Transform your brand with conversational design

Learn how to design, build and launch conversational chat bots from HubSpot conversational marketing expert Connor Cirillo.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

What do I know about conversational design?

Conversational design is one of the hottest new disciplines, and it’s already having a big impact on the world. On a broad scale, it’s democratizing access to specialties and knowledge that were previously restricted to the privileged. The Intuit assistant, for example, is like having a 24/7 CPA. PlannedParenthood helps teens get sexual health questions answered without feeling uncomfortable.

And for individual marketers, conversational design will be one of the most in-demand disciplines in the next decade. Consumers and buyers are on more channels than ever before. And because it would be too easy otherwise, their expectations are higher than ever, too. Savvy marketers like you — yes, you — who understand how conversational design and experiences meet these needs will be more valuable to employers and grow your career, faster.

I wrote this guide to be your blueprint for adding conversational design to your skillset.

Connor Cirillo headshot
Connor Cirillo

Hi, I’m Connor and I help growing companies create great conversational experiences. At HubSpot, I lead Conversational Marketing, where we make marketing, sales, and service experiences more human. I also help companies who are looking to build these experiences for their audiences. You may have seen me on the HubSpot blog, TwilioSIGNAL, or ManyChat’s Conversations.

Chapter 1

Conversational design: what it is, why it matters, and where it’s useful

As humans, we just know when a conversation works. Take the opening scene of The Social Network

Mark: Did you know there are more people with genius IQs living in China than there are people of any kind living in the United States? 

Erica: That can’t possibly be true. 

Mark: It is. 

Erica: What would account for that? 

Mark: Well first, an awful lot of people live in China. But here’s my question. How do you distinguish yourself in a population of people who all got 1600 on their SATs?

Erica: I didn’t know they take SATs in China. 

This feels authentic — it works and we can tell. On the flip side, we also know when a conversation doesn’t work, like this dialogue between a human and an older piece of software

U: Your nose is a question mark, I think thats interesting. 

S: Of course, whats your idea? 

U: I think it fits you perfectly. 

S: Let me see, his nose is still broke. 

U: Do you want me to make your nose broken? 

S: Ah, yes, why did you ask me for this? 

U: Because I thought you might mind.  

We intuitively know this is broken. It lurches, sticks, and struggles. Neither party is on the same page...or heading anywhere meaningful. The whole dialogue is full of grating friction where the first exchange is full of flowing force. 

At the macro-level, that’s what we feel when we read this. But at the micro-level, it’s difficult to put your finger on exactly what’s wrong. 

We know when something is off, but we can’t explain why

What I’m getting at is this: we’re great at conversational design in everyday life, but we’re bad at identifying the components that make it possible. Conversation is so natural and second-nature to us, we rarely stop and think about it. It’s like riding a bike or the lyrics to Say My Name by Destiny’s Child — once you know it, it stays with you (🎶when no one is around you🎶).

Few of us notice: 
  • Conversational parts: Pieces of discourse such as openings (“hey Connor”) and acknowledgment (“okay, I’m with you”) mark a conversation’s progression. 
  • Conversational principles: Certain guiding truths, called Grice’s Maxims, explain how good conversations work (we’ll geek out on this in Section 2). 
  • Conversational expectations: We want to feel seen, heard, and valued in every conversation. It doesn’t matter who we’re with or where we are. Our lizard brains give us a little rush when we feel worth someone’s (or something’s) attention. 

Here’s a quick example of how this works together. Imagine I’m deep in my morning ritual of ordering a coffee at Dunkin’. I might ask, “Hey, I’d like a large iced hazelnut coffee with three sugars and almond milk.” That’s the opening to the conversation. The barista would then say (in an incredibly Boston accent), “Got it, one large iced hazelnut coffee with three sugars and almond milk. Anything else?” That “got it” was subtle acknowledgment that let me know she heard me, and it facilitates back and forth. By asking “...anything else?”, the barista puts the ball back in my court to move the conversation forward. 

If you leave any of those pieces out, the conversation will feel weird or off. Imagine if, instead, the barista said, “Okay, one hazelnut coffee.” That’s all. Now, the conversation has friction because it’s not clear what I need to do next. Pay? Ask about her day? Clarify I want a large? It’s unclear. 

For a conversation to be good, all parties must cooperate and work together. If one party isn’t putting a matching effort into the dialogue, the conversation will grind with friction. It’ll slow down and putter out. You’ve experienced this if you’ve ever asked someone, “What’s wrong?” and received “nothing” as a response. One party uncooperative and it’s clear you and the respondent aren’t on the same page. All of this is friction, and it kills the conversation. 

A useful definition: Conversational design is understanding how to create great interactions between two or more parties. It’s figuring out, agnostic of tools or channels, what makes any given conversation good. Then, it’s creating more goodness by saying the right thing, in the right place, at the right time, every time. 

What conversational design isn’t 

Conversational design isn’t strictly about chatting with chatbots or swearing at Siri. Great conversations aren’t specific to any channel, tool, or industry. They’re not even tied to marketing, sales, or service. Conversations transcend all that.

Conversational design diagram

It’s anywhere and anytime you have a two-way exchange — over text or phone, with a chatbot or human, using a keyboard or voice.

And while studying conversations isn’t a new idea by any stretch (we’ve spoken for thousands of years, after all), it’s more important than ever. The way we all do business is changing, and the marketers who want to take advantage of this shift need to understand the fundamentals of great conversation.

Why conversational design matters: the times, they are a-changin’ 

Historically, we’ve met businesses on their turf. If you wanted to cash a check, you drove to your bank, filled out a slip of paper, and chatted with the candy-doling teller. If you wanted to rent a movie, Mom took you Blockbuster, and you’d pace the aisles hoping they had Shrek a normal movie in stock. If they did, you’d rent it, watch it, and return it rewound — because Mom raised you right.

Blockbuster video gif

Having to go to a business to interact with them sucked, but we’ve come a long way in ten years. That’s because there’s a bit of a revolution going on. Businesses can’t get away with self-centered navel-gazing — at least, not like they used to.

Now, we’re in what Forrester has dubbed, “the age of the customer.” In this age, the rules are different. To succeed, businesses need to meet customers on their turf and terms, or brands will miss the boat.

Consumers are in more places than ever

As consumers, we live in more places than ever and spend more time online than ever before. We’re always connected, and our relationships reflect this. Think of your closest friends. You might use a gif over text to react to a meme they posted on Instagram about an experience you shared. Whoa! 

To reach consumers like us, businesses need to meet us where we are and on our terms. They need to build meaningful relationships — emails with “Hello {{name}}” just won’t cut it. 

Business centric vs customer centric
Via Drift

Put another way, success these days isn’t what you sell — it’s how you sell it.

Few companies understand this as well as Dominos. Case and point: You can order a pizza from nearly anywhere using nearly any device. And not just any pizza — your favorite pizza. 

Sitting on your couch? Order with Google Home. Chasing kids around the block? Use your smartwatch. On the socials? Tweet a pizza emoji at Dominos, and they’ll toss your favorite in the oven. Ordering a pizza isn’t just technically possible, it’s as easy as possible through tech.

Domino's chat options

It used to be that customers walked into a brick and mortar Domino’s or found the nearest store menu online. Said another way, I had to go to Domino’s to get Domino’s. Now, we can ask for a pizza anywhere. That’s ordering on consumer’s terms, and that’s the new way of business 🍕🤝😍.

In return, customers feel less like they’re interacting with Big Pizza™️ and more like they’re chatting with a friend. It’s fun, it’s easy, and it works. It’s also done wonders for the pizza chain once known for its cardboard crust and ketchup-like sauce. 

Dominos stock price graph
Via Vox

This is the power of conversational design: it builds relationships and intimacy with customers, wherever they are. In return, they get affinity for your brand. 

Where conversational design matters: the entire flywheel  

Conversation is the common denominator of every brand interaction, from marketing to sales to service. This makes sense when you’re a customer-centric organization.

A flywheel primer 

At HubSpot, we think about the customer experience in the flywheel: 

HubSpot experience flywheel

You can also visualize the flywheel as Attract, Engage, and Delight. Where marketing attracts and earns customer attention instead of forcing it; sales engages customers in relationships instead of treating them like deal numbers; service delights existing customers by focusing on their success

HubSpot experience flywheel

As the wheel spins (yes, it turns!), it transforms strangers to brand promoters. The faster the flywheel spins, the more a company grows. 

But the flywheel doesn’t naturally spin as fast as a Tour de France cyclist tearing down the Alps. In every brand, friction — anything that makes customers stumble — slows the flywheel down. Friction can be things like confusing onboarding processes or making customers call crappy 1-800 numbers for help. You don’t have to dive deep into your memory banks to know most brands have a lot of friction.

The good news is that great conversational design will actually do the opposite — add force. Put simple, force is what every delightful interaction does to the customer experience.Here are three real examples of conversational experiences removing friction in each part of the flywheel. 

Marketing: how Casper attracted insomniacs

One Wednesday in 2016, the mattress brand Casper launched an unusual chatbot. 

For starters, it only worked between 11 pm and 5 am. If you tried to interact with it before then, it’d brush you off with, “At work can’t talk now. How’s tonight?” 

Later, when you did chat with the bot, it wouldn’t sell you a mattress. It wouldn’t even link you to one. That’s because Casper’s chatbot, Insomnobot-3000, wasn’t there to make sales. It was there to keep you company in the wee hours. While the rest of your friends were in bed, Insomnobot-3000 was up and ready to chat about...well, whatever you wanted to chat about.

Casper insomno bot

Casper’s bot was intentionally conversational, not informational. The bot could navigate common insomniac topics like TV and work stress in a way that felt human. Lindsay Kaplan, Head of Communications back then, said, “We wanted to make a bot that made 3 a.m. a little less lonely.”

So, Casper used SMS to give insomniacs access to the bot through something they already had in hand — their phone. All they had to do was text a number, like how they’d reach a friend.

This fit well with Casper’s brand goal to be more about the culture of sleep than mattresses. While Insomonbot-3000 did earn write-ups in over 30 publications and air on national morning shows, it primarily brought restless sleepers closer to the brand through a conversation.

Sales: how HubSpot generated 182% more qualified leads
The old world of lead qualification, or as I like to call it, “death by a thousand form fields,” sucks. Prospects don’t have patience for it anymore. Nowadays, 84% of customers say if you want to win their business, you need to treat them like a person — not another number in your CRM.

The good news is, conversational design facilitates humanizing interactions. At HubSpot, where I do this work every day, we saw this firsthand.

Back in the distant past of 2018, we started feeling the pain of live chat. More prospects wanted to chat — yay. But when Sales started struggling to keep up with the volume of interactions, it broke down. Worse, many chats were about product support. This ate up the sales team’s time and prevented them from chatting with prospects who needed nuanced help.

To solve this, we built a chatbot. We looked through live chat transcripts, spoke with sales and support agents, and did a lot of work to figure out how the chatbot could help. We came up with these outcomes: it should engage visitors, triage them, and get them to the right place — fast.

The results were pretty incredible. Compared to live chat, 75% more people engaged with the chatbot. Even better, the new process generated 182% more qualified leads for the Sales team. 

Service: how Lemonade delighted thousands of customers 
If I say the words “insurance claim,” what’s your reaction? Cold sweat? Nervous anxiety? Deep groan? 

Any of those are fair because most interactions with insurance is, well, bad and not good. Lemonade gets that, and they’re out to change it. Their homepage tagline reads, “forget everything you know about insurance” and they mean it. Their super-friendly bot, AI Maya, can get you insured in 90 seconds flat — 24/7, 365 days a year. 

But that’s not all Maya can do. 

Back in 2017, Lemonade started to make some serious improvements to Maya. They equipped Maya to make changes to customer policy and handle complex questions, in addition to signing people up. They called this CX.AI — “infinitely scalable customers care experience.”

Customers loved it, saying things like, “I actually enjoyed this usually tedious process! Maya, you the woman!” The support team loved Maya’s new capabilities, too. In Lemonade’s 2018 year in review, they revealed the number of tickets humans handled plummeted when they upgraded Maya. 

Customer support ticket volume graph

This automation didn’t come at the expense of great interactions, which is a common — and reasonable — pushback. By the end of 2019, Lemonade’s NPS score (a way to track customer satisfaction) was an impressive 76. For comparison, that’s about where Apple sits. Meaning, Lemonade customers love their insurance as much as they love their Apple devices! 

NPS score graph

In the last 3 years, Maya has sold over 1.2 million policies all on its own. Compared to traditional insurance agencies that are powered by human reps and the souls of policy-holders, Lemonade saves thousands of man-powered hours and millions of dollars — all through smart, delightful conversation.

These examples show that understanding conversational design 10x-s the value you, as a marketer, can create. And it 10x-s the experience customers receive wherever they touch your brand, too. Win-win.

The next section is everything you need to know to start doing it well. 

Section Recap

  • Conversational design is: understanding what makes interaction between parties good and then deliberately creating great experiences for your brand.
  • Why it matters: The way we do business is changing. Customers expect human and personalized experiences in more places than ever; conversational design delivers this. 
  • Where you can use it: Any channel, any industry, any part of the business.

Chapter 2

Learning how to approach conversational design

To create great conversations, you need to be able to approach them from both a 30,000-foot view and a ground-level view. What does that even mean? Well...

Let’s say you’re visiting Boston and you want to get from a Red Sox game at Fenway Park to one of Boston’s weirder attractions called The Skinny House. (It’s a thing.) 

Tiny townhouse
By Rhododendrites - Own work, CC BY-SA 4.0, 

You know where you want to start, where you want to end, and that there are turns in between. This is the 30,000-foot satellite view.   

Skinny House to Fenway Route

The ground-level view zooms in. It’s the turn-by-turn directions: get onto Charlesgate, keep right at the fork, and follow signs for Downtown Boston. In conversations, there are the nitty-gritty details — what you say, how you say it, when you say it, and why.

Map and driving view

To build a good conversation, you have to hold both of these views in tension and flip between them regularly. The better you can move between the macro and the micro, the satellite and the street view, the better you’ll get at “seeing the matrix” of conversational design. 

To do that kind of mental toggling, you’ll need context, empathy, and a good understanding of Grice’s maxims.

Build a 360-degree picture with context

Every interaction you have with friends and family is the sum of all the other interactions you’ve had with them. 

If your sister was dumped by a boyfriend at Six Flags, you wouldn’t suggest it for your family trip. If your Dad has a tough, Boomer-y shell, but loves football, you’d kick-off conversation with the latest free-agent signing. In both scenarios, your past interactions inform your next one. Context helps you reduce friction, keep conversations flowing, and build strong relationships.

The same is true for conversations at work. The best ones account for context. Even if it’s subtle, your interactions with your bosses, coworkers, and customers look different. Not only are they all different people, but the goals and subjects of those conversations are different, too. Context matters.

What’s true for your personal and work conversations is also true for the conversations you design. To build a conversation that goes well, you need a 360-degree view of the consumer’s situation. This includes: 

  • Who are they? 
  • Where are they?
  • How did they get there?
  • What do they want?
  • Why do they want to do that?

A CRM and identifiers like page URL will help you with the first three. Empathy will help you with the last two.

Lean on your CRM for a holistic view

As humans, our brains and our memories are our CRMs. This system works fine for our personal relationships but, as businesses, we need something more structured — HubSpot, Salesforce, Zoho, etc. — to help with brand relationships.

Just like your brain-CRM helps you create good conversations in-person, your business-CRM helps you create good conversations for your company. How? By providing context. 

If you’re using a CRM (read: you should be), you’re storing information like: 

  • Name
  • Company size
  • Industry 
  • Lifecycle stage
  • Location 
  • Tickets they’ve filed in the past
  • Last contact date

All of this context can come together to create personalized conversations. 

One helpful way to think about using this information is the personalization-effort spectrum. 

Personalization effort scale

On the left, at the lowest-effort end, you fire a generic message into space without regard for the other side. This is about as personalized as a door-to-door salesman pitching you an outdoor playset when you live in an 800 sq ft NY apartment with no children.

One increment of effort to the right is using the customer’s name. Consider the difference between saying, “Hey there,” and “Hey Felicia.” People love hearing their name (it makes them feel seen, heard, and valued), so this immediately personalizes a conversation and builds a better connection. This is a step in the right direction, but it’s not good enough for the consumer of 2020 and beyond.

We all get plenty of emails and LinkedIn messages that get our name correct...and nothing else. You know, the ones that start with, “Hey Connor!” only to ask if I want a demo of a wildly-irrelevant product. Those messages are frustrating not only because they clog up my feed from engaging with absurd influencer content, but because we all expect more personalization than our name. 

Now, slide over to the right side of the spectrum, and imagine what you can do with a ton of context about someone. Imagine you know Felicia manages a six-person e-commerce team, she’s in the “lead” lifecycle stage, she downloaded an ebook last week, and she’s pinging you at 7 am PST for a quick answer before work. Finding and using this information is hard work, but it pays off. The sooner you’re able to tie in details about who Felicia is, the faster you can personalize the conversation, and the better it’s going to go. 

Remember, all of us expect to be treated in context.
Using context gets tricky because, as consumers, we live on more channels than ever before. When we talk with friends, we may start a conversation about wrinkly english bulldogs in-person, continue it over text, tag each other on @english.bulldog, and tweet about the wrinkliest ones we find. We easily move conversations across channels and keep a single thread going. 

Cute scrunchy face bulldog

It’s this one. It has to be.

This omnichannel contextualization is what we expect from businesses, too. As a business, if you don’t have a way to see interactions across as many channels as possible, you design conversations in a silo. Silos lack context —  avoid this.

If someone’s reaching out in a channel you don’t already have integrated, the first thing you should do is try to connect the dots around who they are. You want to get a value — a piece of information you can look up in your CRM as soon as popsicle possible. For example, if Felicia DMs you on Twitter, you may not have her Twitter handle stored. But you probably do have her email address. So, one of the first values you want to ask for is her email. 

At first, I get why this can feel like it adds friction. Soliciting a little info in the short-term pays off when meaningful force gets added as a result. 

Figure out where they are: note the page 

In addition to what your CRM tells you, pay attention to which URL a customer lands on. (Note: this is useful only for onsite or in-app visitors, but if you’re new to this, that’s a good place to spend your time.) A new visitor to the homepage and a customer coming back to the pricing page need different things and want to have different conversations. 

Here are a few assumptions we could make around page type alone, for a B2C SaaS product:  

  • Home: wants to take a common action
  • Blog/Content: hopes to educate themselves
  • Product: interested in what you offer 
  • Pricing/Demo/Contact: has high-intent to purchase

Often these pages map to one of the five stages of awareness. Stages of awareness is a framework that outlines a visitor’s interest level. It’s popular among copywriters and marketers because once you know what stage someone is in, you have a better idea of what they want and how to meet them. 

Here are the stages: 

  • Unaware: visitor isn’t aware of their pain, solutions to that pain, or your product
  • Pain aware: visitor knows they have a nagging pain but doesn’t know there are ways to get rid of it
  • Solution aware: visitor knows there are solutions to their pain, but  haven’t chosen a solution, and they don’t know your product exists
  • Product aware: visitor knows your product does something about their pain, but they’re not 100% sure what or why you’re better than competitors
  • Most aware: visitor is almost ready to buy and just needs a nudge

Here’s how site pages loosely map to stages of awareness:*

Site map awareness stages
*This is a rough guide. Layer quantitative and qualitative data points to create a more accurate picture for your site. 

You’ll notice the Homepage isn’t in the image above, so let’s hone in on that one for a second. Homepages are typically a catch-all kind of page. However, for most businesses, customers who land there take a handful of paths. 

For example, when someone lands on the HubSpot homepage, they’re usually trying to do 1 of 3 things. They’re trying to use their product (log in), read content (go to HubSpot Academy), or learn about pricing and features (talk to sales). So, the bot on the homepage acts as a concierge; it helps customers take one of those paths. 

Here’s how Demio takes a similar, concierge-esque approach: 

Demio conversation example

Keep in mind this is only for onsite or in-app visitors, and it’s only one part of the story. Good conversational design pulls in all the context, clues, and channels it can. You want to tie in who they are (based on the CRM, above) as well as how they got there (below). Remember to layer your data points — everyone likes parfaits. 

Trace how they got here: mine the data 

When you’re answering, How did they get here? there are two things you want to look at:

  • Have they interacted with you before? 
  • How did they arrive at this particular moment?

Have they interacted with you before?

To answer the first question, a CRM is again invaluable. Dive into their contact record and how they’ve interacted with you in the past. It’s great if you know Felicia downloaded an ebook a week ago; it’s even better if you know she signed up for a free product trial two days later and has been using it every day since. 

How did they arrive at this particular moment? 

To answer the second question, use identifiers such as URL strings, cookies, and IP addresses. These can all help you figure out how someone came to your website at this moment. Common options include:

  • Return traffic: They’ve visited your site before, but they haven’t interacted with anyone. 
  • Organic traffic: They’re searching for something your product does. They’re likely aware of the problem your product addresses. 
  • Direct traffic: They typed your URL into their browser. They know who you are.
  • Anonymous traffic: It’s anyone’s best guess, but they’re likely new traffic. 
  • Social campaign: They came from a promo you’re running. 
  • Email: They’re on your mailing list and followed a link.  

Again, the goal is to build a 360-degree, unified picture of this person’s entire relationship with your company. Among other things, this helps you empathize with where the person wants to go. 

Combine science and art to build customer empathy   

So far, we’ve looked at these pieces of the 360-degree view: 

  • Who are they? 
  • Where are they?
  • How did they get there?

Now, we want to hone in on these pieces: 

  • What do they want?
  • Why do they want to do that? 

Figuring this out isn’t like looking for a single missing puzzle piece that makes everything click.  It’s more like looking for several missing puzzle pieces scattered around the room. 

It’s also one part science and one part art. The science involves looking at various types of quantitative data, such as Google Analytics and heat maps. The art is making smart inferences with qualitative help. 

The science of empathizing

When you’re looking for ideas about what a visitor might want to do, behavioral mapping tools and heatmaps are very helpful. 

Behavioral mapping 

Google Analytics is the most ubiquitous tool that can help with this — specifically, the Behavior Flow feature. It shows you the common ways visitors move through your website. It lets you drill down to specific segments, traffic sources, and the like.

Sample Behavior Flow from the LearnWhy site.
Sample Behavior Flow from the LearnWhy site.

As you poke around, you’ll notice people try to do a handful of things when they land on any given page. Once you know what pages people want to get to (the science), you can begin inferring why (the art, see below). 

What else this is good for:
Analytics also help you analyze the types of people who come to your page. For example, maybe 90% of the people who land on your product page are first-time visitors. This has a big impact on the conversations you craft for that page; these people don’t know a ton about your brand. Contrast that to pricing, where 80% of the traffic might be known leads. These folks need less, “Hi! Here’s who we are...” 101-stuff and a much faster path to a specific outcome.


Heatmaps, alongside Google Analytics, are also great for data points. For this, I like tools like Hotjar. Heatmaps help answer:

  • Clicks: What they’re interacting with 
  • Attention: How much time they’re investing 
  • Scroll: How much they’ve engaging or looking

Clicks & Attention 

Clicks and attention are kind of like Behavior Flow, except on a page level. They help you understand what people are trying to do. If you uncover a common set of actions, you can create a well-designed conversation that (a) helps answer those questions and (b) gets people to where they want to be, sooner. You can shepherd people toward common tasks.


Scroll depth can help you determine when to engage someone. Imagine your site as a brick and mortar Brooks Brothers store. If a store associate says, “Hi there! Are you finding everything alright today?” the moment I walk in, it would be intrusive and annoying. Even if I know what I want — Regent-fit sport shirts, obviously — I want to try and find it for myself.

But if I was #meandering around the shirts and seemed stuck between the seersucker or the gingham (sorry, I hate me, too), it would be a more appropriate time for the associate to offer some help. The associate could say, “Hey, we’re actually running a deal on this shirt. Can I help you find a few more?”

Someone asking if you need help when you’ve just walked in (the real-life version of 0% scroll depth) doesn’t tee up a good conversation. Waiting until after they’ve engaged in some way (clicks, scroll, attention) gives you more to go on. Context helps you design a conversation where both sides get to work together to help each other. 

The art of empathizing 

So far, all this quantitative data is telling you what users do, but not why

To figure the why out, you’ll need to read between the lines. Science is knowing 30% of your new traffic goes from the homepage to the pricing page.
The art is inferring why? What’s the problem they have, and what do they need help with? 

Perhaps it’s to learn about pricing and features (a common goal). So, the bot on the homepage could prompt, “Hey, do you want to get info and about pricing and features?” Now, the customer doesn’t even have to leave the homepage or find their way to the pricing page. You anticipated their needs based on the information available. 

Four “arts” can help you develop empathy and inference:

  • Rich personas
  • Customer interviews
  • On-site surveys 
  • Chat transcripts

Rich personas

When I reference personas, I’m not referencing the made-up kind. I’m referencing the kind backed by quantitative and qualitative data. If you have access to that kind of persona, use it as a lens to help you make sense of quantitative data. What problems does this persona face? What outcomes are they trying to achieve?

Buyer persona example

Remember, a persona acts as a marketing guide. Ultimately, it should help you build conversations with a person, not a persona. At HubSpot, we say, “It’s fine to market to a persona, but you build relationships with a person.” 

Customer interviews

If you’re not sure what customers are trying to do, there’s an easy way to find out: ask them. Customers will tell you, in their own words, what problems they’re trying to solve. 

If you have the time and resources, there’s no replacement for talking 1:1 with customers. LearnWhy covers how to do this in an extensive guide elsewhere, so I’m not going deep on it here.

On-site surveys 

You could also “interview” customers on-site with surveys. Surveys are great if you have a hunch and want to confirm your direction. They can also be great for gathering responses to specific open or closed questions. 

For example:

  • Open: “How do you feel about Acme Co. products?” (no predefined answer choices; visitor fills in their response) 
  • Closed: “Do you price to be the most important factor when evaluating Acme Co. products?” (2+ predefined answer choices; visitor selects one) 

Note: open is great for voice of customer mining because the customer supplies their own wording. 

Live chat

Live chat can help you quickly assess what visitors are trying to do. This has been a particularly beneficial experience for my team and I at HubSpot. When we’re not sure what conversations to craft for a page, we start with live chat. We read and analyze transcripts, and those help us quickly figure out what folks are looking for. From there, we can layer qualitative input on top of quantitative data, and get a clear picture of why people were on a page, what problem(s) they faced, and where they’d like to go. 

Two reasons why all this context matters

On the surface, gathering context and building empathy looks like a lot of work — and it can be. Which begs the question, why do it at all? Why not just guess at conversational flows and see how customers react? Two reasons: 

1. Context helps you sidestep bad assumptions (which leads to bad business) 

If you don’t build a 360-degree view of the customer, you’re going to miss parts of the story. You’ll be like one of those horses #clopping around with blinders on — seeing part of the street, but very little else.

This means you’ll make some big assumptions. You’ll misunderstand what problems people are facing or what they’re trying to do. This leads to bad conversations, which creates more friction in the flywheel (you’re making it harder for customers to interact with you), which means less business at the end of the day. 

Worse, you may not even know you’re having bad conversations, so you won’t do much to fix it. Think about it. How many times have you told a brand about a mediocre or poor conversation? Probably not often. It’s more likely you got frustrated, figured it out yourself, or switched to another brand.

Unchecked, it’s no exaggeration to say these bad conversations have the cumulative power to kill your business. Especially in a world where how you sell is why you win. The short term damage is acute; the long term damage can gum up the flywheel and be catastrophic. 

2. Context helps you sidestep rookie mistakes 

One of the most common mistakes I see is treating one conversation as the end-all, be-all. Marketers new to conversational design expect customers to love their brand, sign up for their newsletter, and make a one conversation.

That’s not only unrealistic, it’s far from how any customer purchases anything. Just like you don’t ask someone to marry you on the first date — and expect it to go well, anyways — you don’t take a customer 0 to 100 real quick.

In all likelihood, your customer would rather go 0 to 10... Maybe they want to find your blog or explore the new product you launched. Maybe they’re just looking for your Insta account. Remember, context will help you figure out the specific thing the customer wants to do. 

It’s better to have many small, delightful conversations that add up to a better relationship than it is to cram all your conversations into one interaction.

Once you know who someone is, how they got to you, and what they want, you’re almost ready to start building conversations. Understanding Grice’s Maxims — don’t worry, we’re about to go through it in-depth — is the last foundational concept you need. 

Grice’s Maxims: 4 things all good conversations have 

Paul Grice was a philosopher from the early 1900s who spent a lot of time studying conversations. He was also a bit of an overachiever who crushed cricket, chess, and piano — note to self: be that good at at least one thing in my next life. Grice’s most relevant contribution to what we’re talking about here is a set of guidelines called the maxims of cooperation. For a conversation to be objectively good, it has to conform to these maxims. 

It’s important to remember these maxims center on things both parties have to do. Think of a conversation as a verbal game of catch. You have to grab the ball out of the air and toss it back. For the conversation to continue, the other side has to do the same. 

Throw and catch a ball

Why are these useful?

Think of the last great conversation you had. 

If we were talking in-person, chances are you could tell me exactly when and where your last great conversation was. It’s likely the last time someone made you feel great (seen, heard, and understood) because that’s what good conversation does. As Maya Angelou said, "I’ve learned that people will forget what you said, people will forget what you did, but people will never forget how you made them feel." 

Now, try and remember exactly why that last great conversation was so good. Chances are this is pretty difficult for you to put your finger on. As we mentioned earlier, most of us aren’t good at explaining this stuff. It seems so amorphous. 

Grice’s maxims help solidify things. His four maxims outline the structure of conversations and explain why the good ones are good. This is important for you because it helps you see and audit conversations with more clarity. It also helps you write conversations with more assurance. 

Grice’s maxims state that good, cooperative conversations must be: 

  1. Informative
  2. Truthful 
  3. Relevant 
  4. Fitting  

1. Informative: say what’s necessary

To keep a conversation flowing, both conversants need to provide a certain level of information. Otherwise, the conversation halts. Take this typical after-school interaction:

Parent: “Hey! What’d you learn at school today?”

Kid: “Nothing.”

Parent: …

Kid: …

Parent: ...

What do you do with “nothing”?! The conversation dies. Compare that to:

Parent: “Hey! What’d you learn about at school today?”

Kid: “Sea animals. Did you know dolphins can beat up sharks?!”

Parent: “No way.”

Kid: “Yes way. And guess what else?”

Parent: “What?”

Both sides help build the conversation. They’re saying enough to keep the conversation going, so it pings and pongs the way enjoyable dialogue does. One side says something, the other one adds, and so forth.

What this looks like 

In conversations with customers, informative maxim looks like asking and responding to the right question. This is easy when you know what you’re looking for. If a person asks, “Who should I talk to about upgrading?” they want a name, a “someone”. If they ask, “When will you resume shipping to Canada?” they’re expecting a date, a place in time. 

In other words, there’s an implicit response we’re looking for in every question we ask. 

Question response table

A good place to observe this maxim in action is with your sales or support teams. Chances are, they’re really good at keeping conversations going, asking the right questions, and responding to what’s asked — it’s kind of their job. There’s a lot you can learn from watching them do it.

2. Quality: say what’s true 

The old way of business was you vs. the customers. You’d negotiate, take things on and off the table, and try to walk away with a “win.” That’s not how the world works anymore. Now, it’s you and the customer vs. the problem they face. Your role isn’t to secure a win over the customer, it’s to secure a win with the customer. It’s not Mario Kart, it’s team play on Super Mario Party. 

Brand plus customer diagram

Your agenda aligns with the customer’s, so you can be transparent and honest in what you say. You should also be transparent about who or what your customer is talking to. If it’s Eric from support, say so. If it’s RateBot, make that clear too. 

What this looks like 

Here’s an example of the quality maxim, via Quickbooks. Right away, the Quickbooks assistant honestly tells me (a) I’m chatting with a bot and (b) what that bot can do. 

Quickbooks conversation example

This immediately improves the quality of the conversation and builds trust as well. The result is a conversation that feels more natural. 

3. Relation: say what’s relevant

This is where you get to use all that juicy context you unearthed earlier. The maxim of relevance is about using context (everything you know about the person you’re talking with) to say something that fits the situation:

  • If they’re on the pricing page, it’s more likely they want to talk with sales than find the blog. 
  • If they came from your newsletter, you don’t need to prompt them to sign up for your mailing list. 
  • If they’re calling support, it’s unlikely they want a chipper, cheery, “Hey, thanks for calling support, how is your day?!” type conversation. They want resolution.

Here’s are two examples of getting relevance wrong:

  • Say you have an older, CEO-type customer typing in what we millennials call “full psychopath mode” — correct grammar, punctuation, and capitalization. If you’re responding to them in emojis, gifs, and half sentences, your responses are going to be jarring and create friction. 
  • Or, imagine a brand has one voice and the agents inside it have another. Humans and bots should feel like extensions of one another, and this includes consistency in the way they talk. A written style guide is immensely helpful with this. If you have a written guide, use it. If you don’t have one, create it. 

To create a smooth, frictionless conversation, you need to take all the context you have (including who the customer and brand are) and use it. 

What this looks like 

Here’s how a bot from Wave illustrates the maxim of relevance well. After I told it I needed help with receipts, it gave me a few relevant issues to choose from. 

Wave conversation example

Note that this brief conversation demonstrates the first maxim and second maxims as well — the bot is honest, and it provides just enough information to keep the conversation going!

4. Fitting: say what’s appropriate 

This last maxim is where I see businesses go wrong most often. 

You want to use the bare minimum of words and phrases to move a conversation forward. This is ideal whether the conversation is over email, with a human, or via a chatbot. Why? We get overwhelmed easily. 

When we’re thrown too much information, our lizard brains shut down. You experience this in your inbox when a business sends you a “hey, here are the 8 next things we need to be doing together.” Or when you’re asked to fill out a long, traditional web form that asks for a ton of things upfront. 

Long tedious from example

All this form is missing is “Horoscope” and “favorite Tarantino movie” — Zodiac and Inglorious Basterds, respectfully.

Tackling one of these feels overwhelming, impersonal, and very one-sided. 

What this looks like

To keep conversations more manageable and enjoyable for your customers, there are a few different things you can do:

  • Use the 3:1 rule. Never say more than 3 things (thoughts, message bubbles, etc.) without asking them to interact in some way. Maybe it’s a question, yes/no answer, or simple acknowledgment. Whatever it is, it keeps the other person an active partner in the conversation.
  • Trim words with the Hemingway App. It’s a great, free tool for refining your prose and making sure your conversations say only what they need to.
Hemmingway writing app

  • Read any conversation you write out loud. Cut unnecessary words you hear, then read it out loud again. Keep repeating this process until there’s nothing left to cut. Our brains smooth over bad prose and typos — reading it out loud mitigates this.
  • Play to the strengths of different channels. Email is good for posterity, so that’s somewhere you can give a bit more information. Other channels let you present clickable options for customers — that can save a lot of words. 

Here’s how the beauty brand Sephora presents just the right amount of information (with relevant Covid-19 links!) for a nearby store request in FB messenger: 

Facebook messenger conversation

Informative. Truthful. Relevant. Fitting. These four principles are what every conversation must have to be good.

1 + 1 = 3, or why you need everything you just read  

Great design is invisible. When you do everything in this section well (context, empathy, maxims), prospects aren’t going to think, “Wow, that was the best conversation I’ve ever had. Must be some stellar conversational designer over there!” Instead, they’ll remember how they felt (heard, understood, and valued) and how they figured something out. 

Great, invisible conversational design rides on using all the disciplines we just covered in tandem. Paradoxically, that’s when you get the most out of each one:

  • You need context to identify a goal. 
  • You need art and science to empathize with that goal. 
  • And you need the foundational qualities of a good conversation to string it all together.

That’s how you know they’re starting there, and trying to get here. Then, it’s a matter of filling in the middle. That’s what I cover next.  

Section Recap

  • Build Context: Use your CRM, page URLs, and other technology to build a 360-degree view of who you’re talking with
  • Empathize with customer goals: Use both art and science to put yourself in the customer’s shoes. On the science side, leverage Google Analytics and heatmaps. On the art side, lean into personas and qualitative inputs from surveys, chats, and interviews. Put it all together to build a complete picture.
  • Avoid mishaps: Use context and empathy to avoid bad assumptions and “all or nothing” conversations that gum up the flywheel. 
Grice’s Maxims: All good, cooperative conversations are (1) informative, (2) honest, (3) relevant, and (4) fitting. Knowing this helps you dissect and build good conversations.

Chapter 3

Building chatbots and experiences that drive results

At this point, I’m assuming you have a 360-degree view of your customer. You know: 

  • Who’s on the other end of the conversation
  • What matters to them
  • When they’re showing up
  • Where they’re showing up
  • Why they’re there

Now, we need to get in the weeds of how you can help this person achieve their goal. You have a good idea where they are and where they want to wind up. What’s missing is getting them there

That in-between comes down to a few things: 

  • Creating a great on-page experience (widget rendering)
  • Facilitating a good conversation
  • Mapping out successful paths

While the previous sections look at conversations as a whole, this section heavily features conversation in the context of chatbots.

How your widget renders has huge implications on success

The opening scene of a movie sets the tone for what you’re about to watch. Tarantino’s Inglourious Basterds, a 2009 heavily-awarded war film, does this particularly well.

Tarantino opens the movie on a dairy farm over a glass of fresh milk (the “fresh cream” scene). The tense opening, between a Nazi officer and a French rural farmer, runs 15 minutes and sets the tone for the main characters and the rest of the movie. Tarantino depicts the officer as calculating and cold-blooded — traits developed throughout the rest of the movie.

Fresh cream movie scene

Much like scene-setting, how your widget appears sets the tone for the entire conversation (though, hopefully, your conversations don’t take Tarantino turns). Here are three areas you want to pay close attention to. 

  1. Choosing an honest icon
  2. Writing your initial message
  3. Building in good UX

1. Choose an honest icon 

Imagine walking into a new brick and mortar Warby Parker store. You’ve passed through the glass doors and just walked over to browse the latest frames. A friendly-looking sales associate approaches while you try on tortoise-shell glasses. He has short-cut hair, a square frame, and his name tag reads, “Terrell.” But — here’s where it gets weird — when Terrell opens his mouth, he says, “Hi! I’m Warbybot” in a mechanical-sounding voice.  

That’s jarring, right? Your eyes would get huge, you’d take a step back, and you’d wonder if there’s something seriously wrong with the tortoise-shell frames on your face. This is more Twilight Zone than real life, but your website visitors are going to have a similar reaction if you set a human face and name as your icon...then serve up a robot interaction. 

Like Lowe’s does. 

When you visit the Lowe’s Home Improvement site, there’s a smiling human face in the bottom-right corner. When I recently got stuck trying to order something in the flooring department, I clicked this icon.  

Then Flooring Bot popped up. 

Lowes conversation example

Now, Flooring Bot immediately announced what it was, and that’s a good move. But the experience began with friction because I initially saw a human face. I expected a human, and I got a bot. 

A human face in the widget launcher sets an expectation there’s a human on the other end.

If you break this expectation, you’re going to introduce friction. Remember, one of Grice’s maxims is, “say what’s true.” You want to honestly show who is on the other end. There are two ways to do that: explicitly or implicitly. Casper opts for the explicit option. On their homepage, their chat widget renders with a human icon and the text, “chat with a human.” It’s very clear what I’m getting into here. 

Casper homepage

Casper chat widget icon

The implicit option is more subtle. A good example is when you have a non-human icon for when there’s a bot on the other end and a human icon for when there’s a human on the other end. 

Smart brands, like Outdoor Voices (OV), switch between the two seamlessly. When you land on OV’s site, the chat icon shows the brand logo. When you click this logo, a bot appears. But when you ask the bot a question and was handed off to a human, the icon changed from OV to K. Also, the label at the top of the chat switched from “Outdoor Voices” to “Krista.” The implication is a human named Krista has taken over. 

Outdoor voices chat example

Most conversational tools, if they’re worth their salt, will support this kind of transition or handle it automatically for you. 

Either way of disclosing (implicitly or explicitly) who’s behind the chat is fine, as long as it sets an honest expectation. Deception will always cause friction. 

So, do I have to use a robot icon for a robot?

No. If you can, steer away from a run-of-the-mill bot icon and opt for something specific to your brand or campaign. You could use your logo, a company mascot, or some other creative option. 

Whatever you choose, keep in mind this is a manifestation of your brand. It’s a way someone can talk to your brand. Start with your logo if you’re totally unsure.

2. Write a compelling hook

There are three common options for loading a chat widget onto a page:

  1. Just the icon 
  2. Icon plus prompt 
  3. Fully loaded chat 
Allbirds website example

Option 1 

Free ebook chat trigger example

Option 2
There’s no inherently right or wrong option here. It all depends on the context. To revisit the brick and mortar example, if you walked into the brand’s store, would you want to be greeted with your options immediately (fully loaded chat)? Or simply know there’s a sales associate available if you need them (just the icon)? 

If you go for a partial or full display, you’re going to need a great hook. A hook is an opening line that grabs a visitor’s attention. 

How to use conversion copywriting tactics in your hook

I find it helpful to think of hooks as subject lines. Especially if the prompt appears before the visitor clicks into chat, this is an accurate parallel. The subject line exists to get someone to open an email; the hook exists to get someone to open chat. 

For example, here’s what ConvertKit says to someone after they’ve clicked the option to create a new email but haven’t typed anything for a few minutes. 

ConvertKit conversation example

This is a great, contextual hook. To write a good one for your setting, you want to use a few copywriting principles:

  • Have a goal. It can be a small goal, such as subscribing to the newsletter. Or it can be a big goal, like a purchase. Either way, have a goal. 
  • Match the hook to the goal. Everything that happens in the conversation, including the message prompt, should drive toward the goal. If you want them to head toward San Francisco, don’t point them toward Canada.
  • Make it relevant. This isn’t a billboard off I-95. This is a targeted message to a specific persona you know a lot about. Use all the context you gathered in section 2 to craft your message. 
  • Make it enticing. “Do you want to increase your traffic by 42%?” (um, YES) is more intriguing than “Learn about getting more site visitors...” (meh, I’ll pass).
  • Ultra-specific. Specifics create intrigue and relevance. “Snag your free copy of The Ecommerce Playbook for reducing churn” is much better than “Get a free bonus gift.” 
  • Short and simple. Say use instead of utilize, get instead of procure, easy instead of uncomplicated. As Winston Churchill put it, “Short words are best, and old words when short are best of all.” (Reminder: the Hemingway app helps with this.) 
  • Deliver on your hook. Just like a subject line leads into the email, the hook should lead into the conversation. Don’t pull the classic spam move and tease a hook that has nothing to do with what you plan to say (“You won’t believe Nick Cannon did this…”). 
Conversational starter example
A good example of several principles (enticing, relevant, specific, goal-oriented) in a good hook. 
This is not the place for open-ended questions

Avoid free-response questions in your hook. For example:

  • “How can I help?” (well, I’d like a jet plane) 
  • “Who would like to speak to?” (uh, Drake?) 
  • “What can I do for you?” (track down my true love, please) 

These questions make people think, which decreases the chances they’ll engage. These questions also increase the chances you’ll get a response your bot can’t handle.  

Instead, you want to set up a Plinko board using close-ended questions. Only, instead of pushing them to a bad outcome (like rigged arcade games), you want to push them toward the jackpot — the outcome they want. 

3. Facilitate great user experience (UX)

Most consumers look for chat in the bottom right corner of their screen. It’s also possible to embed chat in the page design itself (less common) or even serve it over ads, as you’ll see in a screenshot later. 

However and wherever you place your chat, you want to make sure your visitor has a great experience — one that’s frictionless and enjoyable. To create a great user experience (UX), here are a few things to keep in mind. 

Make “what next?” crystal-clear

Whenever someone interacts with your bot, it should be very clear how to move forward. You want to avoid scenarios like the one in the screenshot below, where it’s not obvious what each option is or which one the visitor should choose.  

Messenger chat example

Does “Yes!” do something different than “Take the quiz”? It’s not clear, and that’s poor UX. In conversation writing, every phrase and word carries a lot of weight because one wrong move (an ambiguous question or poor reference) can break the conversation.

Shopify’s bot is a good example of clear paths done right. 

Intercom chat example from Shopify

Experience-wise, there are few things Shopify does well:

  • When the bot pops up, it provides less than five options to choose from
  • Each option is clear and you can imagine where they’d take you
  • Shopify wants you to select one of these options, so they turn off your ability to type something in. “Choose an option above…” appears where “Type a message…” normally would be. This makes your way forward crystal clear. 

Keep in mind, to guide visitors through a clear path you have to have a clear path. We’ll get into the specifics of crafting this path in the “How to map your conversation” section below.  

Be brief 

Brevity isn’t just part of Grice’s maxims (see “say what’s appropriate”), it’s good UX. Short, concise responses are easy to read and follow. 

A good example of brevity is Roo, Planned Parenthood’s sexual health chatbot. Roo gets questions about a lot of messy topics. In response, Roo could be very wordy. Thankfully, it’s not. 

Here’s how Roo answers, “How do I tell someone I like them?” 

Planned parenthood conversational design

Entire movies try to answer this question (Grease, Twilight, When Harry Met Sally), but Roo tackles it in two sentences. That’s great UX. 

When it comes to brevity, there are usually a few ways conversations go off track. Avoid any of the options below:

  • The great wall of text: Ever asked a friend how they’re doing and they send you a novella? I bet you didn’t read it right away. Walls of text are overwhelming and hard for anyone to digest. 
  • Morse code mishap: Other friends break their novella into 20 different short texts. This, too, is obnoxious and hard to read. 
  • Death by “fine”: When you get too brief (“How are you?” “Fine.”) the conversation fails. The other person doesn’t have enough to go on, and you break Grice’s maxim of quantity. 

Pick a tool that makes sense 

Marketers often get bad cases of “shiny object syndrome.” This is where we pick a tool (or tactic) because it’s trendy and has fun features. We dream of what we could do and forget to focus on what we should do. 

We saw this a couple of years ago with the Facebook messenger gold rush. Marketers were convinced this platform was “it” and “the next big thing.” There was bot goldrush, too — every company and their grandma started making bots. 

The problem is, starting with a tactic and fitting a goal around it is bass-ackwards thinking. It’s like showing up in a convertible to help a friend move and saying, “oh, I thought you might enjoy riding in this more.” Okay, but that’s not at all what your friend needs. In that context, they don’t need a leather-lined insta-worthy ride. That’d deliver a terrible moving experience. They need a heavy-duty truck that’s pulling a hefty trailer. They need a vehicle that solves their problem. 

Similarly, you always want to start with who you’re interacting with, what you’re both trying to do, and why it matters. Only then can you pick the right tool. 

Don’t pick a conversational channel because it’s hot. Pick it because it adds to your overall customer experience goals. 

Also, keep in mind geographics play a role in messaging tools. Many countries have a preferred platform, and it’s not always the one we use here in the US. Start with the customer and their preferences instead of trying to force them into the tool you chose. 

Note: We’ve only skimmed the top of UX here. There’s much more under the surface, including various accessibility concerns. If you can, work with designers and developers to ensure your implementation is usable for all visitors.  

3 key ways to facilitate a great conversation

Throughout this entire guide, we’ve mentioned a lot of different ways to create good conversations. First and foremost, great dialogue starts with context — understanding why conversations work, what makes them good, who you’re having them with, and what you’re trying to help them accomplish. 

Here are 3 key ways to translate all of that context into the conversation itself. 

1. Set and meet expectations

Expectations are beliefs about what will or should happen. They matter here because, like the Twilight Zone Warby Parker example showed, you create friction when you don’t meet customer expectations. And friction kills conversations. 

Here are a handful of key ways you can set and meet visitor’s expectations for a frictionless conversation: 

Set expectations: 

  • Icon: indicate human or not
  • Hook: indicate the goal of the conversation
  • Capabilities: state how the bot or human can help 
  • Actions: make it clear whether you want a click or a typed response
  • Timing: if it’ll take someone (human or bot) a while to respond, say so 
  • Priming: use open and closed questions to set up responses

Of course, visitors are bringing their expectations into the conversation as well, especially if they’ve interacted with your brand before. It’s important to meet those, too. 

Meet existing expectations: 
  • Brand: follow voice and style guides; be “on-brand” 
  • Setting: adjust your tone for context; chipper is less appropriate when you’re interacting with a frustrated customer (e.g. support)
  • Personalization: consider every interaction the customer/visitor has already had
  • Conversation: deliver on your hook or your opening promise

Here’s a good example of how the hip DTC mattress brand, Purple, set and met expectations in a recent interaction I had.

First, some context. Purple served me this experience in an ad on Outside Online. 

Purple conversation example

I had recently visited the Purple site and browsed their mattresses. (I was familiar with their playful brand.) This conversational ad re-targeted my initial interest.  I assume the goal of this conversation was to draw me back to Purple’s website and reconsider a purchase, which it did. 

Here’s how the conversation went.

Purple mattress conversations
Purple mattress conversations

Here are all the different ways Purple set and met expectations in this conversation:

  • Icon: I’m chatting with a bot and the icon shows the brand logo (NOT a human).  
  • Hook: Purplebot introduces the goal of the conversation: find the perfect mattress.
  • Capabilities: Purplebot immediately discloses its only job is to help me find that mattress: “Hi! PurpleBot here to help you find your mattress in 3 simple steps.” 
  • Actions: Purplebot indicates what I should do in the first step (and future steps) when it says, “Choose yours below.” 
  • Priming: The bot drives me toward my quiz results with a series of close-ended questions.
  • Brand: The friendly tone, emojis, and exclamation points all match Purple’s funky, friendly style. 
  • Personalization: Because mattress shopping royally sucks, and I’ve already experienced that pain, this done-for-you approach is incredibly relevant...and enticing. 
  • Conversation: The conversation follows through on everything the hook promised: my perfect mattress in 3 simple steps. 

Purple also did a great job with call and response in that conversation. Let’s look at this component below. 

2. Use call and response

One reason a standard web form sucks is that it’s all one-way. Each of us wants to feel seen and heard, and a form just doesn’t do that. Businesses routinely ask us for 8 fields of information and give us no validation or value until we’re all done. This feels unnatural because there’s no conversation where a human would ask for 8 things in a row without acknowledging you. 

Compare that to ordering a pizza with Dom, the Domino’s pizza assistant. 

Dominos conversation design
Dominos conversation design

Conversations like the one I had with Domino’s provide a better long as you facilitate a back and forth, as Dom did. 

When someone answers your bot or gives you info, acknowledge them. This could be as simple as saying “Got it” or “Okay” as Dom does above. Acknowledgments keep the experience two-way and collaborative. It creates what I call the “ping and pong” of natural conversation. 

Here are some other ways to ping and pong:

  • 3:1 rule: Don’t send more than 3 messages without asking for reciprocation 
  • Right question: Common question words — who, what, when, where, why, how — map to implicit responses. (See Grice’s first maxim in section 2.) 
  • Right answer: When a visitor uses a common question word, respond in the way they expect. 
  • Make turns clear: When you need a visitor to respond, make it obvious. 

3. Reduce cognitive load

In any conversation you craft, you want to take out the hard work for the customer and your team. Removing the hard work for the customer means making their responses easy. Removing the hard work for your team means fewer data points to parse on the backend and higher conversion rates. The best conversations will check this win-win box.

Here’s how to reduce cognitive load:

  • Focus on ONE outcome: don’t go from 0-100 in one conversation; aim for the 0-10 approach. Remember, many great interactions > one poor one. Help the customer find a gingham shirt vs. turning them into a Brooks Brothers brand advocate. 
  • Limit choices: The more steps and decisions you can take away, the easier it’ll be for the visitor to engage. E.g. “Would you like me to connect you with the team? [Yes please] [No thanks]”
  • Group options: Limit choices by putting like items/paths in the same buckets. See how Gear Collection does this below. 
Gearbunch conversation design

  • Use rich components: Play to the strengths of whatever channel you choose; use buttons, images, carousel cards, and anything else to your advantage.
  • Set quick replies: Instead of asking “how can I help?”, give visitors predefined paths they can take to common end goals or destinations. 
HubSpot conversation design
How HubBot uses quick replies.

  • Watch your vocabulary: Use simple words wherever possible. Take extra caution with international conversations; watch out for phrases or colloquialisms that are easily misinterpreted (like “two shakes of a lamb’s tail” below — it’s cute and personable, but it isn’t very clear).
Allbirds conversation design

If you don’t have natural language processing, machine learning, or a tool like LearnWhy at your disposal, it’s best to keep the conversation simple. Again, the less guesswork there is (for the customer and you), the better the conversation will flow and the more satisfied the visitor will be. 

Word of caution: you can take this too far 

It’s easy to take the idea of reducing cognitive load, or any other tip I shared above, and make it your end-all-be-all. When you do that, and forget why the customer is there in the first place, you wind up with Eventbrite’s mistake. 

During the Pandemic, the only option Eventbrite’s help center bot gave was, “I have questions about the Coronavirus.” 

Eventbrite conversation design

While this looks like a reduced cognitive load, it actually frustrated users. Folks visited the Help Center for plenty of reasons outside of the virus. As one user commented on Twitter, “...the refund functionality isn’t working and your chatbot only wants to talk to me about the coronavirus. I just want to issue a few refunds…” Whoops. 

Reduce cognitive load, but never forget why people come to the page in the first place and what they want to do while they’re there. 

How to map your conversation from beginning to end 

Conversations can take many turns and be complicated. That’s the bad news. The good news is, as I wrote at HubSpot, “If you can read a ‘pick-your-path’ adventure book, you can build a bot.” I stand by it. 

Here’s what I mean. Creating a conversation is a lot like building a choose-your-own-adventure map. If they say this, they go here. If they click this, that happens. If they make this mistake, I say that. 

Of course, all these paths are a lot to keep in your brain at one time. To organize things, you want to build a flowchart before you build anything else. 

The flowchart approach 

Flowcharts are my favorite way to map conversations, and it’s the way I work with anyone new on my team. They’re logic-driven and easy to put together. If you don’t already use a flowchart tool, I’m a big fan of Lucidchart

Mapping a conversation in Lucid Chart
Conversation template in Lucidchart

Other mapping options 

If you don’t have or want to use a flowchart tool, you can go low-fi and use sticky-notes, a whiteboard, or pen and paper. I’ve even seen nested bullets points in a Google doc and spreadsheet. But if this is your first time mapping conversations, try the flowchart approach. From what I’ve seen, it’s the most helpful way for new conversationalists to map their first interactions. 

Note: If you’re designing voice interactions, more robust tools will be helpful. But most people don’t start with voice interactions. 

Start with the happy path

The happy path is the perfect scenario of back and forth. It’s where the visitor quickly reaches their goal and nothing goes wrong. 

Let’s create the happy path for the Purplebot example I showed earlier in this section. 

Here’s the context again: 

  • Goal: Get the prospect to check out their ideal Purple mattress
  • Who: Eager shopper who needs a new mattress, is considering Purple, and has briefly checked out the inventory 
  • Where they’re starting: Separate web page on the internet
  • Where they want to end up: Looking at their perfect mattress 

Now, here are the first two blocks you want to chart: where they are and where they want to end up.

Purplebot example mapping

Add in the initial message (don’t worry, you can revise it later). Now, what response(s) could you get to that initial message? 

Purplebot example mapping

From there, keep asking, “What happens next?”; “How do I get them to the end goal?”; and “What information do I need to collect along the way?” And keep filling in logical steps until your visitor reaches the outcome they want. In this case, the perfect mattress. 

Purplebot example mapping

When we reach this point, we have the skeleton of our conversation. We also have the“happy path.” 

Of course, not every visitor will have a happy journey. That’s why it’s important to plan unhappy paths, as well. 

Next, tackle the unhappy path

At every juncture in your map, you want to ask yourself: 

  • What am I asking this person to do? 
  • How could they get it wrong? 

For your map, list out the “got it wrong” options at each juncture. If you can clarify your prompt and eliminate some or all of those wrong paths, go ahead and do that. But for many prompts (especially open-ended ones), there will be at least a few ways your conversant can screw it up. 

For those, you need a graceful way to fail, fallback, and reframe. Let’s say your bot asks for an email address and the response it gets is “ham sandwich.” That’s not at all what the bot is looking for, and it has no clue what to do with this. Your job is to figure out how to get things back on track.

If you do that job poorly, you’ll create a conversation like this: 

Friendly the Wolf conversation example

An infinite loop of doomed unhelpfulness where the visitor leaves in a fit of rage (“Well I’M sorry YOU suck!” *keyboard slam*). 

To avoid that unpleasant scenario, and manage unhappy paths well, here’s what you need to remember: acknowledge, rephrase, prompt. These are the three components you need to both tell the conversant what went wrong and show them how to make it right. 

  1. Acknowledge: State you’re stuck and be specific about what went wrong (“I’ve never seen an email like that.”)
  2. Rephrase: Say what you need differently, and give an example (“Something like would be great.”)
  3. Prompt: Ask for what you need directly and politely (“Enter your email again below.”) 

Here’s a remix of how I showed this in Chatbots Magazine

HubSpot conversational design example

If I only acknowledge, I’m rude and unhelpful. If I only acknowledge and rephrase, I leave the customer hanging. But when I acknowledge, rephrase, and prompt, I keep the conversation going and the other person engaged. 

Here are a few other tips: 

  • Keep it “you + user vs. problem”: You’ll notice I didn’t blame or belittle the visitor in my example. The goal isn’t to make the other person feel like an imbecile. The goal is to work with the person on the other end to help them reach their goal.
  • Fall back to humans: Always fall back to a human when the conversation isn’t working. As Friendly demonstrated above, there are few things more frustrating than getting stuck in an infinite loop. Remember, it’s not bots VS. humans, it’s bots WITH humans. Bots should strive to make your humans more efficient — not replace them altogether. 

In summary, these are your basics of mapping: 

  1. Start with where the customer is and where they want to be
  2. Create the happy path between those two points
  3. Plan for unhappy paths along the way 

This sounds plenty simple, and it is in theory. But keep in mind no map is perfect the first time around.

How to improve your map draft 

In her book, Bird by Bird, writer Anne Lamott explains the concept of shitty first drafts. Shitty first drafts are the initial thing you get down, so you can work with what you’ve written later. Lamott confesses, “All good writers write them. This is how they end up with good second drafts and terrific third drafts.” 

Conversational marketers do this, too. The first conversation you map won’t be brilliant, and that’s normal. You’ll need to work with it, revisit it, reshape it. You’ll want to red-pen it until it flows the way it would between two people. 

Here are some pointers on the re-shaping process.  

Picture sitting in a theater

Imagine you’re in a front-row seat watching two actors on stage (or screen) play out a similar conversation. 

For example, if you were trying to get someone to register for an event, imagine the parallel scenario of a dad signing his kid up for a lacrosse tournament. If you were watching this in real life, how would it go? 

If the dad walked up to the registration booth and the person said, “What’s your name, email, phone number, position, and shirt size?” that’d be pretty unnatural. A real-life conversation would look more like the booth person saying, “Hey, what’s your name?” “Connor,” the Dad would reply. “Oh, great Connor,” the registration assistant would say, “Glad you’re here. What’s your phone number? Just so we can ring you if anything happens to Connor Jr.” 

Picking an analogous scenario (shopping, troubleshooting, ordering a pizza) to your personal life, and thinking how that conversation would go, is a great way to ground the conversation you build. 

Read it out loud early and often 

By far the easiest way to edit your conversation is to read it out loud. When you read out loud, you’ll trip over areas that are passive tense, jargon-y, too long, or overly formal. This will feel like briars catching on your sleeve. The pieces that snag are the ones that don’t match how we speak with our friends over coffee or in person.

So, take what you’ve written and role-play it with yourself, start-to-finish. If you come across a snagging spot, mark it with a pen, but keep going until you’ve run through. Edit the conversation, then do this again. 

When I write a conversation, I do this hundreds of times. And I always find stuff that looked good when I wrote it but sound unnatural when I say it. When you slow down and read things out loud, you’ll figure out these areas for your scripts, and start unscrambling them. 

Even if your bot is meant to be read, write your conversation as if it’ll be spoken. I find most people don’t spend enough time having the conversation in their head or out loud. 

Have someone role-play the scenario with you 

After you’ve role-played the conversation out loud, grab someone else to act it out with you. Is this a little awkward? Yes, but it’s incredibly helpful. 

Find a teammate, hand them a coffee, and tell them they’re the bot or the prospect in your role-play. Explain the goal of the conversation and then run through your dialogue verbatim. This will help you find any more briars — any areas that feel weird or unnatural for human interaction. 

It will also help you think through:

  • How you’ll get back on track if the prospect says something wrong
  • How different questions or phrases can be misinterpreted
  • Where you could be more clear, specific, or concise 
  • Where the potholes are and how to get around them 

A word on filler phrases, grammar nazis, and emojis  

When you’re role-playing a conversation, you and your co-workers will both use filler words like “Um,” “uh,” and, “hmmm.” In moderation, these are fine to use when they fit your brand and style guide. Remember, you want the conversations you have in a bot (or through any other automated means) to match how your brand and humans talk. The same logic applies to grammar, emojis, and gifs. Purple used emojis and exclamation points in their conversation earlier, and that fits their style. But if a 200-year old insurance agency sends me dab gifs, I’m going to feel weird about it. 

Use shapes to organize your flowchart

As you get more comfortable with flowcharting, you can start creating a flowchart syntax. For example, I use different shapes in my flowcharts to represent different values. So, a rectangle may indicate the bot makes a statement, a rhombus indicates the bot asks a question, and so on. 

You may not start with this, but it is a handy way to organize your charts over time. As you create your charting language, it’ll get easier for you to see how a conversation develops and edit it.

One final word on mapping: if you’re frustrated trying to do this, you’re probably doing it right. It means you’re thinking, “I just want it to be good, and it doesn’t feel good yet.” That’s a great position to be in — keep going! 

Once your map has been through the wringer, it’s time to get it live and monitor what happens. 

Section Recap

  • Be widget-smart: How your widget renders on the page has big implications. Be thoughtful about your icon, initial message, and widget appearance. Consider the user experience. 
  • Set and meet expectations: Conversations don’t happen in a vacuum. Use all your context from section 2 to both set and meet the visitor’s expectations. 
  • Use call and response: Don’t act like a web-form. Keep the conversation flowing with questions, acknowledgments, and other devices. 
  • Reduce cognitive load: Use a variety of tactics to make the conversation as easy as possible. 
  • Map your conversation: Before you build anything, map your conversation. Start with where the visitor is and where they need to go. Add in the happy path. Then, tackle the unhappy path.  
Edit, edit, edit: Relentlessly improve your conversation by reading it out loud, acting it out, and imagining it in a theater.

Chapter 4

Launching your bot and measuring your results

Conversational marketing is a very iterative process. Your work doesn’t stop once you’ve created a conversation. You still need to test, launch, and monitor the thing to answer, “Is this better than what we had in place?” 

From pre-launch to troubleshooting, here’s a primer on doing this well. 

Pre-launch testing: try to break it  

After you’ve run through your conversation with yourself and another person, you want to quality assurance (QA) test it. Which is a fancy way of saying, you want to try to break what you built. 

Run through your conversation and see if it works as you expect: 

  • Does it look like it’s supposed to? 
  • Does it show up when and where it’s supposed to?  
  • Can it handle everything you need it to handle? (typing, buttons, other inputs)
  • Can you complete the happy path without any glitches? 
  • Can you break it? (e.g. by typing when the visitor needs to press a button, entering the wrong input) 

Do this until you make sure every piece stitches together the way it should. Once it does, you’re ready to go live.  

Going live: how to run a good conversational experiment   

Conversational marketing may be a new discipline, but that doesn’t mean you need a new way to test it. Approach launching conversations the same way you approach other marketing activities (campaigns, emails, landing pages) — as an experiment. 

Experimenting means you use a scientific approach:

  • Define the problem you’re solving
  • Write down your hypothesis 
  • Figure out how you’ll measure that hypothesis
  • Collect data for measurement
  • Analyze the data to see if it supports your hypothesis

Why are you doing this? Define the problem and a hypothesis  

When you launch your conversation you should know, very clearly, what problem you’re trying to solve. For example:

  • Too few qualified leads are reaching the sales team 
  • # of customers is increasing and support productivity is decreasing. Support reps are answering too many basic questions that customers could self-serve instead. 
  • Our NPS score has decreased because customers are having to wait over 30 minutes for a response to their complaints 

You should also know how a conversation might help your problem. Be explicit about this: We have x problem, and we think doing y thing will make z impact. Here’s an example: Sales is getting too caught up in support conversations. We think adding a chatbot to high-intent pages will improve sales productivity by 10%.

A good hypothesis will be: clear, specific, and testable. It will also be meaningful and worth testing (your success or failure matters to the business). 

Once you craft it, this statement is your anchor. Everything you measure, ignore, and interpret is tethered to the problem and your hypothesis. 

What should you measure? Identify key results 

Once you have your hypothesis, you need a way to measure it. You need key results. 

Keep in mind, the metric you track is the behavior you incentivize. So be careful when you select a key result, and make sure the behavior you encourage lines up with what you’re trying to do.

Quantitative metrics guide

Here are examples of key results you could track, organized by the area of business you’re problem-solving. 

  • Marketing: lead generation
  • Sales: qualified conversations, qualified leads (QLs), or productivity gains 
  • Services: time to resolution or productivity gains
  • Experience: net promoter score (NPS), customer satisfaction score (CSAT), or response times

Productivity gains refer to making your team more efficient. For example, the hypothesis above (we think adding a chatbot to high-intent pages will improve sales productivity by 10%) targets a productivity gain for the sales team. Ways to measure this include # of sales tasks the team can complete each week or % of hours go toward support issues.

Remember, the specific key result(s) you track should line up with the specific problem you’re trying to solve. Everything is tethered to that anchor. 

Qualitative measures to add in

In addition to collecting quantitative metrics, you want to spend time reading through conversation transcripts. Look at the emotions of visitors in the conversation — do they sound happy or pissed? Are they satisfied or frustrated? This practice will help you assess the quality of conversations. 

It’s also a good gut-check. Say you want more visitors to self-serve before they engage with the support team. You’ve been measuring deflections, and those look pretty great (a lot fewer people reached support). But when you look at the transcripts, you realize something is off. Yes, fewer people reached support, but it’s not because they got the answers they needed. It’s because your bot sent them unhelpful links and they left frustrated. 

In other words, the quantitative side looked like a success but the qualitative side indicated a failure. You need both types of analysis to holistically answer, “Is this better than before?” 

What about benchmarks?  

Be wary of using industry-wide benchmarks as your comparison point. There’s going to be so much variance between contexts (B2C, DTC, B2B), that it makes more sense to use your data. Ground your metrics in what you’ve been doing and what you’ve already seen. For example, if your forms are converting at 25%, set your key result at a 30% conversion rate or higher. 

Watch out for vanity metrics 

The first time you run a conversational experiment, you may see metrics such as 1,000 conversations started. That’s cool, but if your goal was to generate qualified leads (QL) for sales, and you only got 1 QL out of 1,000 conversations...well, you can almost hear the hot air balloon deflating. 

No metric is inherently bad, but be wary of putting too much weight on two metrics in particular: 

  • conversations started 
  • deflections

These two rarely give you a good picture of whether your hypothesis is true or false. Conversations started shows you engaged people, but it doesn’t prove you helped them do anything. Likewise for deflections. A deflection shows you re-directed a visitor, but it doesn’t indicate they wound up somewhere meaningful. They could have simply stormed off in frustration, which is not at all a win. For a more holistic picture of how your bot is doing, choose specific metrics that relate to the specific problem you’re solving. 

How long do we run the experiment? Give it time to work

I generally keep conversations in the wild for two weeks, or until I get a hundred conversations. Only then can I get a good sense of what’s going on. 

The two-week mark typically provides two things:

  • Statistical significance
  • A chance for visitors to work through all paths

If you look at a smaller batch of interactions, say 25 conversations, or one day, there’s too much variance. It’s easy to over-index a few people who are frustrated and then over-correct. So, unless your bot is saying or doing some absurd things, give it some time to work. Then you can ask, “are people getting to where we want them to go?” 

Note: Need a quick refresher on statistical significance? Here’s a good resource

Is it working? Analyzing chat and program-level metrics

Remember how, when you’re crafting conversations, you need to toggle between ground level and 30,000 feet? You’ll need to toggle with metrics, too. When you’re forming a conclusion about whether a bot is a success, there are two buckets of data you want to look at: chat-level and program-level. 

Analyzing chat-level metrics 

When I initially run experiments, I pay close attention to chat-level metrics. I’m specifically looking to see if conversations are performing better than whatever we had in place for our key results. 

Once I know a conversation is better than what we had in place, I don’t go deep into chat-level data often (just now and then to answer a specific question). Instead, I monitor program-level metrics. 

Analyzing program-level metrics 

While chat-level metrics fade into the background after the experimentation phase, I still look at program-level metrics every day. 

These are metrics such as :

  • How many total bot engagements do we have?
  • How many of those engagements got to humans?
  • How many became qualified leads after talking with a human?
  • How many qualified leads went on to buy?

If, for example, one of the overarching goals of all our bots is to generate leads, this helps me assess whether or not we’re doing that. Especially as your programs get more mature, this is a good practice to start.  

Use both buckets of metrics to figure out, “is conversation working for our brand?” 

Getting back on track: how to troubleshoot 

Sometimes, your experiments will be a raving success. You’ll gather quantitative and qualitative data that paints a compelling picture you’re solving the problem you set out to solve. In those cases, keep going!  Other times, your experiments will fall flat or spiral. That happens to every marketer. 

Here’s what to do when things don’t look so great:

  • Focus on what you can control
  • Analyze the mini-funnels
  • Identify your biggest drop-offs
  • Make improvements 

Focus on what’s in your control 

One of my big life philosophies is, “control what you can.” When you’re troubleshooting, there are some things inside your control and some things outside your control.

Factors outside your control include:

  • Who’s coming to the page
  • Why they’re coming to the page
  • Sudden changes in page traffic (e.g. an SEO term tanks)
  • Page load time and other technical pieces

Factors inside your control are: 

  • what happens in the conversation
  • the experience visitors have when they engage

What this looks like for me: I don’t get too worried if chat volume goes up or down by a few hundred each week. That’s due to site traffic, which is outside my control. I do get worried if traffic stays flat, but conversations drop a meaningful amount. This indicates something inside my control is wonky. 

When that happens, I look at the mini-funnels inside the conversation. 

What’s happening in the mini-funnels? 

Mini-funnels depict how visitors move through the bot. Picture something like Google path analytics on a bot level. 

Say you’re giving visitors a coupon code to use for their first subscription purchase. 100 people click the chat icon to receive the coupon. 80 people give their names, 10 people give their emails, and 2 use the coupon.

This is a simple funnel, and there’s a clear cliff where folks fall off. That’s important to notice. 

Where are your biggest drop-offs?  

Drop-offs are where you’re losing a substantial percentage of people during the conversation. For example, the dropoff between the name and email step above. When you spot these cliffs, you want to ask questions like, “Why are people getting stuck here?” and, “what can we do to optimize this path?” 

For example, at HubSpot, we once chatted with visitors interested in a content management system (CMS). We asked these visitors for a website URL, but many visitors fell off at this point. They didn’t provide a URL. We realized this is because a big percentage of these folks don’t have a website URL, which is why they’re interested in a CMS in the first place. So, we looked at how to rephrase the question, explain why we wanted the information, and also asked whether we even needed it at all. 

Use a similar approach to address your drop-offs: 

  • Ask: why are people getting stuck here?
  • Identify: how to optimize the step (change the question, provide clarification, or remove it altogether)
  • Implement: an improved solution 
  • Monitor: whether that solution improves drop-offs and improves your key results

While this all looks neat and tidy in bullet points, it’s going to feel pretty hard at times. 

Hang in there: it’ll take a while to get all this right 

While we’ve been having conversations for thousands of years, the discipline of conversational design is pretty new. It’s a blip on the anthropologic timeline of having conversations. 

Conversational design timeline

So, don’t beat yourself up. The first conversation you design and test probably won’t be great. But you know what? The first email you sent or the first landing page you put together probably wasn’t that great either. You’ll get better at conversations, just like you got better at those. 

On your more difficult days as a conversational marketer, keep these things in mind: 

  • You’ll learn through failure. You’ll build a bot you think will do one thing and realize it did something else entirely. The more mistakes you make, and the more you learn about what went wrong and why, the better you’ll get. I’m still learning and making mistakes, and I’ve been doing this full-time for years. So, keep your chin up. 
  • It’s an iterative process. Conversations aren’t like brand campaigns where you spend months and months of building and then ship it. This is a more iterative marketing discipline because you get such a large amount of qualitative and qualitative data in real-time. 
  • Conversations are a hard thing. Conversations are two-way engagements with your brand and a prospect. It’s people, and it’s relationships. It’s an order of magnitude more complicated than an email or blog post; it’s an order of magnitude harder to do well, too. 

Hopefully, I’ve saved you some headaches and hardships in this guide. But especially if this is your first time building a conversation, don’t expect to nail it on round one. It takes a while to get good at this, and that’s normal. But don’t worry — you’ll get there soon enough. Keep your head up.  

Section Recap

  • Use old approaches: Conversational marketing is new, but testing isn’t. Use the same scientific approach you use for your other marketing experiments.
  • Nail the basics: Write a strong hypothesis, identify meaningful key results, and run the experiment long enough to get statistically significant results. 
  • Troubleshoot: If the conversation isn’t performing well, focus on what you can control, analyze the min-funnels, identify drop-offs, and make improvements. 
  • Hang in there: Conversational marketing is hard, and it’s going to take you some time to get it right. Keep your chin up and look for opportunities to learn from your mistakes.  

Questions about this guide? Want to learn more about conversational experiences? Keep up with me on Twitter or Linkedin

The fastest-growing companies are built on deep customer insights
Subscribe and stay up to date with all things customer insights.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
No spam, guaranteed. 😁