I’m the official mad scientist at Segment. And what that means is that my role is to come up with ideas that others don’t.
^ That’s our pal Guillaume Cabane, a.k.a. “G.”
Technically, G is the VP of Growth at Segment. But I think we can all agree that “Mad Scientist” is a much better title. (And after you learn about what G has been up to, you’ll understand why that title is a perfect fit for him.)
We recently had G into the Drift office to give a presentation — to 100+ marketing and sales folks — on how he manages growth at Segment.
The big problem he addressed:
How do you stop your marketing and sales funnel from leaking?
The point is that 19 out of 20 visitors to your website that you paid for, probably some good money, are just leaving. And so what should you do? Should you just pour more on top of that leaky bucket? Should you just spend more and more of your VC money? I guess it depends on your VCs. Mine would say, “No, that’s crazy.”
Check out the video below to see G’s full presentation. (Or keep scrolling to read a summary.)
Introduction: Reducing Cognitive Load
A key point that G made during his presentation was that people don’t just drop by B2B websites (like Segment’s site) to pass the time and enjoy themselves. As he explained:
No one does that. Not on a B2B website. If they come, it means there’s a reason behind it. They clicked on an ad, or they came from a good piece of content. It’s means that they’re attracted to something, to part of the message, or they have a pain point that they need solved. And if they leave, it means that we haven’t solved that pain.
The underlying mechanism at work here, G told us, is cognitive load.
What I mean by cognitive load is when they come to your website, and they read your message, they need to interpret the message. So they come with a preconceived idea. They think, “Hey, here’s my exact pain point, here’s what I want to solve.” Then they read your message and say, “Huh? I don’t get it. I’m out of here.” And that’s cognitive load: It’s the effort you’re asking of your visitors. And the problem is, unless it’s a perfect match, which happens 1 out of 20 times, they will leave.
So, how do you increase that 1 out of 20 ratio? How do you drive more qualified leads into your funnel and prevent folks from bouncing?
Back to G:
To fix that problem and increase your conversion rate, what you need to do is lower the cognitive load. And the only way to do that is to make the message more in-tune, more relevant, and more personal to your visitors.
Of course, personalization means different things to different people. And for Guillaume, “It doesn’t just mean a better message for everyone.”
Because if you optimize for the average, you’re actively ignoring a heck of a lot of potential customers.
G gave the example of a (hypothetical) shoe company, which decides to “personalize” its message for men’s shoes based on the distribution of shoes sizes for men in the U.S.
Turns out, size 11 is the most common, so the company decides to put a message on their website that says, “We have the best size-11 shoes for everyone!”
But as G explained, “That, of course, is stupid. It’s never going to work.”
And here’s why it’s never going to work:
As marketers, especially in B2B, we’re always saying, “Hey, we have to find a better message through A/B testing, a message that speaks to everyone. Well, the problem is that everyone is different. We all have different attributes. And if you think about it, this is our message on Segment’s website — it’s “Stream data to every product integration your team needs.” Well, that’s good if you’re a product person. But the problem is, if my visitor is a marketer, I would need to say, “Stream to data to every marketing integration your team needs.” But I can’t do that. So I’m kind of stuck between settling for a generic message that speaks to no one, or a specific message that speaks to my biggest audience but ignores all the others.
In order to solve this problem, G set out to “personalize the entire funnel, from collecting the data, to sending email, to sending live chats.”
The framework he came up with for doing this is called “Collect, Analyze, Act.”
Let’s walk through it together, step-by-step.
According to G, there’s no way you can personalize an experience if you don’t know who you’re talking to. So the first step is to learn more about each individual coming to your site.
And to do that, you need to gather more data.
G came up with a simple equation that helps illustrate why this is so important:
More data = more leads, and more leads = more money.
One way Segment goes about collecting this data is through email-based enrichment on their sign-up forms. Once someone enters an email address, the rest of the form gets auto-filled with their information (which is scraped from the web).
Here, I’ll let G explain:
The way that this works, we split the form into three steps. We collect the email address on the homepage, and while the second step is loading, we query an API, called Clearbit, to get all of the data on the user and on the company, and we send that back to the browser to pre-fill the form. And not only that, there are two other tricks happening here. First, we actually get way more data than what we display on the form. We get about 120 different traits. And we use that to personalize the social proof that shows up at the bottom. For example, if a marketer is filling out the form, we show them social proof from another marketer so it’s relevant.
By pre-filling the form, G is able to reduce friction and decrease the effort needed on the part of the user. Combine that with the relevant social proof being displayed, and he’s seen his conversion rates climb 50%.
But it doesn’t end there. G then takes the data he’s collected and uses it to create “extremely personalized paths” for the different types of people coming to the Segment site. And one way he does this is via their welcome email.
Marketers who sign up, for example, get an email from G in which he explains that he too was a Segment customer in charge of marketing once — so he’s “been exactly in your shoes.” And it goes on from there with specific instructions for how marketers should get started.
Developers, on the other hand, get an email from Segment’s head of customer success, Jake, and it’s completely different. It includes access to all of Segment’s technical documentation — something marketers probably wouldn’t be interested in.
As G explained:
If you cross the emails together and you send the technical docs to the marketer, you fail completely, which means that you’re increasing relevance here. But that’s not the only thing you can do when you have all of those traits.
The other thing you can do, according to G, is you can start building some machine learning-based scoring models:
In our case we use a vendor called MadKudu, and they get all of the data through our enrichment partner Clearbit, so the 120 traits, and they create a model based on past conversions. It’s a regression tree model, and it starts at employee count and goes into how much they’ve raised, the industry they’re in, and the technologies they have running on their website. And it gives us a score, which means, “Is this a good lead or a bad lead?”
If you’re in marketing or sales, you’re probably familiar with lead-scoring tools and models, “and you’ve probably had a bad experience,” according to G.
But there’s good news on this front: The technology has gotten much, much better.
We are able to predict 80% of our revenue within 16% of our leads. In other words, it means I’m able to send one-seventh of the leads to my salespeople while maintaining 80% of the revenue. That’s a significant reduction in noise. Very significant.
And that’s very important because all of the companies I’ve been at in the past, we’ve had a huge problem: We’d have tourists coming to our website. So we have people who land there because they have a problem, and we don’t answer their problem — we talked about that. The other thing we have is we have people who are never going to pay. They have the problem, but they’re never going to pay for Segment. They’re freelance developers, they’re in another country where we have no sales reps, they’re just never going to pay. And so the cost of handing them through my sales process is not acceptable. So I need to exclude them from there.
Using lead scoring, G is also able to create a separate onboarding flow for the users with the highest scores.
The email comes from one of the co-founders, and it puts in copy not only the user, but also one of the sale reps, because it syncs with Salesforce. So it pulls the lead owner from Salesforce and it names that sales rep in the body and gives you the email address and tells you what you should do. So that could be automated and it makes sense, because what happens when one of the co-founders puts in copy one of the sales reps? Well the sales reps going to stop whatever they’re doing right now and they’re going to answer immediately. And why do I need to have that done manually? I don’t need do. So we completely automated that and we actually have a copy, here, which actually replies to the original email, adding it as a fake reply here, and says, “Hey, I’m glad you found us.” And with that reply I’m doing my job as a sales rep and replying within five minutes. And now we have a conversation going between two people at Segment and a customer.
From the customer’s perspective it looks like a conversation, it looks like there’s an important person who said, “Hey, you’ve got to talk to this guy,” and the sales rep is like, “Oh yeah, now I’m talking.” It looks completely natural and human, and that’s what I want them to believe. And this performs extremely well. So that’s the value. Of course, I don’t want to do that with anyone. I can only do that with people who are highly valuable because I have a very high rate of answers.
Finally, G also has a separate flow for people from very large companies. In addition to personalizing the social proof, he offers them one-to-one onboarding on the sign-up form. “Would you like help getting started?” Yes or no?
The question is, why do we do what? Well, think about it: If the person who signs up is Nike, what’s the chance they actually sign up for Segment and install Segment with the technical docs in just a few days? It’s zilch. It’s never going to happen. They’re going to fail, and we know that because we’ve look at the data. What we need to do is get them to talk to a sales rep, or to a technical engineer, as soon as possible. So I want to avoid them having a bad experience where they’re going to fail. I want to avoid failure mode in the app. And when they press “yes,” there’s a screen that says, “Thank you, we’re going to reach out to you,” and that’s it.
Now we’re a bit more aggressive, so now actually as soon as you enter your email, we have a screen which offers you to do a hands-on onboarding, and which offers you some calendar times. Just to give you an idea, about 3% of the people select “want help,” and the conversion rate once they select is about 70%. What’s really important here is most of the emails for your high-value customers you do as a marketer, if you’re a marketer, is to get engagement, is to get people to respond saying, “Yes I want to talk.” And you’re going to do version after version to get there.
The problem is that you’re asking for something from the customer. Here I’m reversing who’s asking for what. The customer is asking for some of our time, and we’re branding our time as being valuable. I’m selling something: I’m selling valuable onboarding. And, of course, that’s why the conversion is much higher on the booking rate. And the good thing is that most sales reps don’t need to do anything. They just pick up the phone.
Despite all of these improvements, however, G still had a huge problem.
The problem is it’s limited to post-engagement stages. We’re still only improving that 1 out of 20 people, that 5%. And sure, we did +30% on the sign-up rate and sure the boss is happy, but I’m still looking at 19 people who just left the door without saying hello. And that pains me very much everyday. And so when I think about that, I think about what I can do to solve that problem. And it’s not by improving that match rate, it’s not by finding a better average, it’s by finding a similar process that enables me to find another key to the user data. And in our case, because we do B2B, we found that the IP address is a very, very good key for that.
So you’ve probably heard about IP to domain enrichment, I think I started hearing about that in 2010, about seven years ago now (how time flies), and it was horrible. Really, really bad match rates and really, really bad data. But it has improved significantly over time. And I’m going to cover that. But basically the way it works is, we send an IP address to Clearbit, and they return a domain, we get some of the data and we also get the whole 120 different attributes, which means not just the domain, but the whole company object.
On average, the match rate for IP to domain enrichment is 26%. But that figure, according to G, is highly dependent on company size, “which makes sense: The more people there are in a company, the more confidence they have on the fact that that IP belongs to that company. That’s how they measure the signal.”
The average match rate for people who sign up for Segment, meanwhile, is 50%
And that’s very interesting, because those are the people we want, the people who have intent. So our match rate on the people we want is actually higher than 26%, even though the match rate for our website traffic is 26%.
Once you’ve collected and analyzed this type of granular data on your website traffic, it’s time to see how you can use it.
The first thing G used it for?
There’s a reason why we did that. It was a new process for us at the time, collecting all of that data, and we wanted to see if it worked, and we had a pipeline that was great, through Zapier and a lot of other things. So we just tested a few campaigns. And the core principle, if you want to think about it, is this:
You get IP addresses, then you get companies, and because you have the same company object as what we had for sign-up, we can use the machine learning model, and so we can score anonymous visitors. Now we start knowing, of all of those people who leave, those 19 people, are they valuable? Or actually, who is valuable of those 19 people? Is there anyone who’s valuable? And the answer is yes. And the answer is there are a lot of valuable visitors who are just leaving without doing anything. We can know how much money we’re leaving on the table. They’re not just tourists, they’re people who come from valuable companies, we know they’re in our top score, we know they came for a reason, and they are leaving.
G went onto to say that if you were to present that problem to your board of directors, and you didn’t have a plan to fix it, “I hope you’re fired.”
However, one big question G had to answer when it came to sending cold emails was, “Who should they target?”
And the way to do that is we go back to another API, and we ask for specific profiles, specific roles. And in our case, at Segment, we like to talk to the VP of Engineering, VP of Analytics, analysts, data engineers, and what not. So for each company that we find, we say, “Let’s get 3 to 5 of those top roles and let’s send them emails.” And the data is pretty interesting.
We have the anonymous ID, the page they visited, the company name, the lead scoring model, the domain of the company, and the count of how many people came from the same company. Because those are anonymous visitors, they’re unique visitors — cookie ID’d. And in one case, the case of the company Simple.com, which was a good fit, there were 13 people that came from a company that is not a customer. What does that say about intent? For sure it says something. It says that we’re missing something completely and that we need to do something.
Just to give you an idea of scale, we found at the time that we launched 1,500 new companies per week, of which 300 matched our top score. Per week. So there’s no way your sales team could do anything about that. It’s way too big. And that resulted in 5,000 new contacts per week. So I can’t just drop that on my 10-person sales team and say, “Hey, go ahead and contact them.” It’s not going to work. Also, they did not express any intent, so we can’t expect our sales reps to handle them at the same conversion rate. So we need to automate that. When we go back to the API and we pull specific roles, we can see the roles here from the same companies. We’re pulling the name, the title, the role, and all of the rest. You can see that we’re pulling product people, engineering people, senior director of business, director of product…perfect people. So, we did emails.
So, what’s the open rate like for G’s data-enriched cold emails?
And we’re talking about an open rate of a cold email. It speaks to two things. 1) The copy is relevant. Because we extract data on which scripts they’re running on their website, we match those on those we support as a platform. We inject it into the email. We use the industry to look at which customers we have in that industry, and we name the customers in the email. Just like what we did on the sign-up form. So, the copy is good, but of course, it just speaks to the intent and the quality of the data. Those companies have intent, so they’re opening the emails. It happens that we’re emailing them at the right time. Surprise. And you say, “OK, well you’ve got a good open rate, G, but does it work? Do you generate business?”
So here’s an answer, and this is just one example. In this case, we emailed a person named Henry, and Henry is the director of product management. And Henry forwarded the email to Joy, and Joy forwarded the email to Andrew. Andrew said, “Hey, we’ve got to talk.” So we were able to penetrate through different profiles at the company with one email, which speaks to the intent and to the power of the email, which is amazing.
But once again, despite this incredible result, G still had a problem with how his cold emails were performing:
We’re creating a huge friction compared to engaging them directly on the website. We are forcing them to come to us, to leave, to wait for an email, hope we email the right people and then to answer to us. Sure, it works, and it’s much better than most systems, but the content on the website still didn’t solve the problem, and we still offered them a very poor experience on the website, so we can’t really settle it there.
Build a pipeline that moves faster using live chat.
There were some technical challenges that we needed to face. We needed to detect the company, we needed to score the company, and we needed to trigger the live chat in a matter of seconds. Because if the pipeline takes like 5 minutes, well your visitor is no longer there. And when I look at the distribution, time spent on site for the people who don’t convert, most of the people are below 15 seconds. So if you have 30 seconds, you’ve lost more than half. So time is of the essence here. And so we did pretty much the same pipeline: We get the IP address, we get the domain, we score the company, and then we fire live chat if the score is high enough.
So, why did G decide to use lead-scoring with live chat? Why didn’t he just show live chat to everyone?
There’s a reason for that, and the reason is: I’ve tried live chat for five consecutive years at past jobs, and I failed for four consecutive years, which is a long time to fail. Luckily, I was in a different job so the boss didn’t know I failed in past jobs when I offered the same idea again. “Oh, is this going to work? Oh, I trust you G.” And then I’d fail again. But the question is, why did I fail?
I failed because salespeople are expensive, and asking them to handle low-value chats is not worth it. They were just saying, “G, you’re funny with your stuff, but I can’t just spend all of my day answering to random people on the homepage.” And as much as I wanted to do that, that’s not happening. So the point is, we created a test pipeline, and we said, well, what if we could power live chat just for those valuable people, when we know they’re valuable? What would it do on the conversion? What would it do for our sales reps? The interesting story is we lost almost no leads, because the other people who were low-score eventually would not convert, and the sales people started giving me bottles of wine saying, “This is fantastic! I’m getting money and I’m not doing anything.” So that’s how I measure success.
One reason why G’s live chat pipeline has been so successful is that he’s able include personalization, just like on his forms and emails.
We inject the company name in the chat, and then load Zapier to get the pipeline of data. And the pipeline takes the data from Segment, it sends the IP address to Clearbit, we get the company name, and then we say “We have advice for [company name].
As you can see, we actually inject the company name in the message. Now that’s interesting, and begs the question of, “Why are we doing that?” We’re doing that for very much the same reason that putting the first name or the company name in the email subject works better, it actually worked way better a long time. Almost no one is doing that because it’s pretty hard to do, which means I have an unfair advantage versus all of the other people who are using live chat. I’m surprising my customers with relevance. But not only that, just like with the email, we actually inject stuff like the tools that they are using at the company which we know. I inject stuff like the relevanT customers, or sometimes we’re just honest and say, “We use a machine learning algorithm to see if you’re a good fit, and you are.”
So, does this approach actually work?
I’ll let G tell you:
We actually increased the click through rate by 5x between injecting company name or not injecting the company name via relevant data for the IP address, and that is just crazy.
However, there’s still room for improvement.
It’s still only a fraction of all visitors, to be honest. And it’s good but the question is, “How does it interact with the rest of the funnel?” And that problem we’ve been facing more and more, as we power live chat to more and more of our traffic. Live chat doesn’t sit alone. Live chat is in the middle of the funnel. People who come to the website and they see the chat, but then, chat is not the end. Then they sign up and then they use the product. And the problem is that my sign-up and my email does not have the context. If you think about a possible customer journey, this is a possible customer journey: People visit the website, they’re scored, we decide to display the live chat, they have a conversation, they go on with their life, they spend two days doing something else, they forget about us, and then they remember, and they come back and they say, “I want to sign up.” And they sign up.
So, what does G do for cases like those? Instead of sending those folks a generic email for a random person at your company, e.g. “Just so I can help you, what challenges are you trying to face?”, he’s able to tie live chat data back to Segment and create a specific path for people who come in via live chat.
The email that they get is now, “So glad you signed up after our live chat conversation yesterday, should we hop on a call to continue our conversation?” And that comes from the same person that they had the chat with that past day. And it’s a text-based email. And there’s now two options: Either they believe it’s a personal email, which they could very well believe, because it’s a text-based email coming from the same person they had the chat with (and at that time it was a human). Or they don’t believe it, but it’s still super relevant. I’m offering better value.
However, G admits that this is going pretty far with personalization, and he’s asked himself before, “Where’s the limit?”
There isn’t one.
And to prove it, I’ll leave you with the latest example of what G has been up to:
Imagine I want to surprise my most valuable leads, those 16%, those are worth tens of thousands of dollars for me. So I’m happy to spend a bit of money if I can increase my conversion rate. So what if I could send them a small gift, not two weeks after they pay, not two weeks after they sign up, but while they’re having a live chat conversation? So, this is what I’m working on. The idea is that they visit the website, we display live chat, and instead of telling them, “Do you want to talk to us?”, I’m asking them: “Would you like a coffee or a tea?”
So I ask, “Tea or coffee?”, and they say “tea.” And I respond, “Tea, that’s great. What do you want in there? Sugar, milk?” The visitor says no sugar but a cloud of milk. And we say, “Just to confirm, is this the right address?” Because we have the address of the headquarters from the enrichment. And this is using Drift’s latest release of LeadBot, which handles multiple questions and has buttons to capture structured data. What happens afterwards, is we ping a brokering platform, they have access to Postmates, and I can fire Postmate’s webhook to buy a tea for that customer. It’s going to take 15 minutes, which is slightly longer than what we’d wish, but it surprises the customer.
That’s the whole point as a marketer: I’m here to delight my customer with something that is delightful and funny. And someone was telling me at a previous conference, “Oh, Guillaume, all of what you do is great, but it’s still some spam.” But at least you can drink it! There’s actually a study on hot drinks that you might want to look up that proves that having a hot drink makes you believe that the person offering you a hot drink is agreeable and helpful, that they’re trying to help you. So there’s a psychological effect. It’s just fun, it’s nice, and it’s five dollars out of tens of thousands. So I don’t care.
Final Thought: You Don’t Know Your Customer
You‘ve got to accept the fact that you don’t know your customer. Don’t tell me “No, I know my customer.” Unless you have one customer, you don’t know your customers. I don’t know all the details of all the people in my family, so if you know your customers and you have a few hundred, then there’s something weird. But I think it’s OK, because I think the AI and the enrichment will take it from here.