Each week I reverse engineer the products of leading tech companies. Get one annotated teardown every Friday.
Growth Dives: Behind Fyxer's 90% month-3 retention
The story behind Fyxer's 90% month-3 retentionRetention isn't won with shiny new features - it's won by obsessing over the core job. Read online here or download as PDF Last week, I shared some of Fyxer’s series B metrics on LinkedIn:
In the comments, someone asked about the retention stat in particular 👀
That same week, Andrew Chen published a post on retention, where he observed:
What does this mean in a nutshell? Well, if you want to win with retention, pick a problem that happens daily. Build a solution for it that’s good enough for customers to switch. Then you will have great retention. At the simplest level, that’s what Fyxer has done.
That’s the product–market fit side of the story. And for early-stage startups, it’s the most important. But what’s next? What’s the post-product-market-fit retention strategy?After working with over 20 products, across B2B and B2C, seeing 1000s of tests, I have never seen a new feature increase retention. I’ve heard of a few cases, but not seen it with my own eyes. It is always improvements to the core product that improve retention in a big way. To win with retention, you have to ask: What did customers come to us for? Then optimise to make that incredible, rather than adding things they might (but realistically don’t) want. For example:
Deepening (rather than expanding) product-market fit should be the focus. I promise you, there’s no amount of sexy new features that will increase retention. For Fyxer, better retention doesn’t come from shiny new features. It comes from: better emails and better email categorisation. So the real retention question for Fyxer becomes: how do we get better at drafting people’s emails? It all starts with people. The competitive edge in AI isn’t data quantity - it’s data qualityMany people think that you win by training on as much data as possible. That’s not actually the case. What matters is what you feed the model, and how close that data is to your target customers. One third of ChatGPT’s retention improvements have come from refining the model with examples from customer use cases. If people are trying to write prose or code, the ChatGPT team improves the model on the use case people care about. It’s the same for Fyxer. One of our co-founders, Rich, was on a podcast last week called Founded & Funded run by Karan Mehandru at Madrona. On the episode, Rich shared: We have over-indexed on quality of data - it’s the biggest differential we have to the rest of the market. Before starting Fyxer in 2023, Rich and his brother Archie had run an executive assistant agency since 2016. That meant when GPT-3 launched, they already had:
As Rich put it: “We had instant product–market fit because we built the data before the product.” That’s when they teamed up with Matt, now our CTO. Matt explains it simply: To get the best performance out of AI, you need to train it. That involves building massive datasets that tell it how to behave, in a really specific way: How should I phrase my writing to most closely emulate the user’s tone? What time should I propose for a meeting? General AI models (like GPT-3) are trained on huge, messy public datasets. That gives them breadth, but not precision. The problem is that public data doesn’t contain consistent, labelled examples of what “good” looks like for those cases. To get a model to behave like a skilled executive assistant, you need it to make very specific decisions. So, to build those datasets at scale, you need a human data team. This creates a consistent, high-quality training data set, which gives the model a much clearer view of how to behave. As Matt says, “The more aligned the human data team is, the stronger the model becomes.” And that data is created in-house by human labellers never from customer inboxes or live emails. So, to bring it back to retention. Retention for Fyxer comes from deepening our core value proposition: writing incredible emails. To do this, we train general models on high-quality, very specific datasets that we create with a team of data labellers. Quality in = quality out. This is not only how we will improve retention, but also what drives a moat for the business. As Archie explains:
In conclusion, retention for Fyxer is about human data labelling. But also focus.I’ve covered how Fyxer won with early retention (how the concept and creation of the business set them off on the right foot) and how the current product strategy (focusing on draft quality) is how we win. But what I left out is how hard it is to focus a team. In product we’re used to thinking big, blue sky thinking, jumping on trends. Especially in AI, there are so many shiny things we could be building in. So many cool features. SO many use cases. So many potential customers. With all that, it’s actually incredibly hard to focus on one thing. Ruthless prioritisation involves saying ‘no’ 100s of times a week. And that’s the unsexy bit. The actual challenge with retention is not 'what to build' it's more 'what not to build'. This issue is sponsored by Attio ✨ Attio is the CRM for the AI era. Connect your email, and Attio instantly builds your CRM - with every company, every contact, and every interaction you’ve ever had, enriched and organized. With Attio, AI isn’t just a feature - it’s the foundation. You can do things like:
Join industry leaders like Flatfile, Replicate and more. Did you like this inside look? Reply and let me know! See you next week, Rosie 🕺 |
Growth Dives
Each week I reverse engineer the products of leading tech companies. Get one annotated teardown every Friday.