The growth engineering team at Fyxer might be the most impressive I’ve met over the past decade.

Fyxer, which brings a Cursor-like experience to email, exploded from $1 to $30 million in annual recurring revenue (ARR) in 2025 and they’re forecasting $100M ARR by the end of 2026. Their secret: running 514 (!) experiments, more than two per work day.

The growth engineering team alone launched 360 out of the 514. Their core team, led by Kameron Tanseli, is four engineers strong. This means they’re doing 90 experiments per engineer per year. That’s just insane (I think British folks like Kameron might call it bonkers).

This is an AI-native growth team where growth engineers have full ownership over an end outcome, with a big assist from AI.

Up today: a behind-the-scenes look at Fyxer’s growth engineering team and the real-life experiments that powered Fyxer’s 30x growth.

January 2025: Targeting work emails

When Kameron joined Fyxer, the business was around $1 million in ARR. Many of the initial users were friendlies or early adopters and there wasn’t a clear sense of which people were most likely to pay and expand.

Kameron enriched the signup emails with Apollo’s API and then analyzed the expected lifetime value (LTV) by different types of signups. When Fyxer got adopted as a work tool, users were far more likely to invite their colleagues and Fyxer could start marketing to other employees at the same company.

Fyxer didn’t shut off personal email signups despite the lower LTV. Instead growth engineering partnered with the marketing team to revise their prosumer motion to target, nudge, and optimize for work emails. They added a route in the product for personal email signups to add their work email, too.

February 2025: A credit card gated free trial

In February 2025 Fyxer had a classic 7-day free trial model. Free-to-paid conversion was about 5%, in line with the latest conversion benchmarks.

Kameron ran an experiment to ask new signups to add a credit card upfront. The conversion rate jumped from 5% to 35%.

I asked what happened to signups. Kameron admitted that they dipped; however, the overall number of paying customers doubled after the experiment. Fyxer concluded it was a winner after only 8 days. The paywall was actually optional during this experiment so this was essentially free money on the traffic of new users.

“It was around the time when a lot of AI apps required a credit card upfront,” Kameron told me. “There were changing perceptions around the willingness to do this. We also had a high intent rate since people were connecting to their email.”

February to March 2025: Shifting more users to annual plans

Around this time Fyxer saw most new signups opt for month-to-month plans. Kameron was looking to shift the mix toward annual to help with both retention and cashflow (collecting cash upfront would allow for more to be spent on ads). He tested different annual discounts plus breaking the annual pricing down to the effective price per month.

The control UI was the screen from above, which defaulted to the monthly plan and where yearly plans would have one month free (8% off). The winner: defaulting to the yearly plan, offering a 25% yearly discount, and communicating the effective price-per-month.

This test 2.3x’ed the share of new trials being annual compared to the control. Fyxer now sees 50% of paying customers sign up for annual plans.

March 2025: Raising prices from $30/user to $50/user by adding a Pro tier

By March, Kameron shifted gears to price testing. At the time Fyxer had a single package, which was priced at $30 per user per month. He did a test where he introduced a new, larger feature package (the Pro tier) at $50 per user per month – nearly double the old price.

There were three different variations of price testing:

This ran from March 3rd to March 10th. The winner was Option 3 and the revenue impact was huge:

March 2025: Adjusting trial lengths to 3, 7, 14, and 28 days

Next Kameron moved to trial lengths.

He first tried to gamify the trial experience where users could get a longer trial if they invited team members. Kameron tried multiple versions of this over a few weeks. Nothing worked.

He tested all kinds of trial lengths spanning from 3 days to 28 days. The short 3-day trials didn’t work at all. “People would cancel immediately and we had terrible trial start rates,” Kameron recalled.

He found that a 7-day trial worked best for overall conversion rates, although personal email signups converted better with 14-days (it took these users longer to get to an aha moment). The winning test: segmenting the trial lengths with personal users getting 14 days and everyone else getting 7 days.

The increased trial length for personal users increased their trial start rate from 13.4% (control) to 22.1% (treatment), or an increase of 65%.

April 2025: Team invites

Gamifying the trial length didn’t help improve team invites and so Kameron changed course. He tried moving team invitations to after the paywall, realizing that people would be more likely to invite their team after they took that extra step.

Kameron tested pre-populating the invitations with a user’s closest team members and making a slightly harder to opt out of the invites. The user could easily push “Continue” or un-check the recommended invites one at a time. This had a big increase in the number of invites being sent, and Fyxer sees that one-third of the invites get accepted.

I’ll mention that Fyxer’s referral program is rather generous with both sides getting $50 for an accepted referral. This was initially open to all types of users; switching to only users with a work email reduced referral program abuse by a large amount.

May to June 2025: Onboarding flows

When someone adopts Fyxer, they mostly experience the product inside of their email inbox. The product actually changes the inbox itself with things like smart labels and drafted email responses. The experience outside of the inbox has to be frontloaded before someone gets there.

Most of Kameron’s onboarding experiments focused on getting people to connect their email. After all, this is what people were most afraid of (giving AI full access to your email). He tested a variation to the “Connect your email” screen, shown below, to add more social proof and explain exactly what Fyxer’s AI assistant would do.

The treatment screen increased the percentage of users who connected their email from 57.1% to 59.8%. This was a modest uplift (+4.7%), but was statistically significant (p=97.5%).

The experiments mentioned above are only a handful of the 514 experiments Fyxer ran in 2025. The tech stack, workflows, and AI prompts are what made this possible.

1. Ideate

PostHog provides dashboards to understand the bigger picture about the funnel and user behavior. It’s connected to other tools like Customer.io (customer engagement), BigQuery (data platform), and GrowthBook (feature flags and experimentation).

2. Analyze

Claude Code is Fyxer’s AI data scientist. It’s connected to product data (read only access to BigQuery) and has access to written markdown files that describe each schema and its relationships, which provide context around the data. Claude Code then writes and executes SQL before returning results to the growth engineering team.

Claude Code generates SQL queries, answers drill-down questions about the data, creates charts, and can even execute Python code.

3. Spec and wireframe

Growth experiment specs are prompted within Claude Code and sent to Linear via Model Context Protocol (MCP).

The Fyxer team gets an insight from Claude Code (ex: “run a segmentation analysis on personal versus work email users”) and then within two minutes could generate an experiment ticket in Linear. Claude Code even provides the wireframe, which goes in the ticket itself.

“A lot of the experiments we do are very surgical, for example an extra step in the onboarding or an upsell somewhere in the email,” Kameron mentioned. “The AI models are good enough at UI and good enough at coding to just do the growth engineer’s tasks. Growth engineers can focus on data exploration and iterating on results.”

Here’s the exact experiment creator workflow and prompt:

They use Lovable internally for prototyping. This models the email inbox and how interactive emails work inside of them.

4. Code

Claude Opus 4.6, Codex, and Tembo help one-shot smaller experiments such as upsells or configuration changes in the backend.

5. Test

They use GrowthBook for feature flags and experimentation.

6. Share

Growth engineers also have a Manus prompt for turning visual A/B results into Slack launch posts to update the company.

Kameron describes Manus as “essentially Claude Code but available on the go and triggerable via email.” It allows them to create, install, and use skills.

Other tools include NotebookLM to generate weekly workshop slides and upskill the growth engineering team, Firebase, and NodeJS.

Fyxer’s growth engineering team provides a lens into what’s possible with AI.

The team currently has four growth engineers (although they are hiring). Each owns one of Fyxer’s growth loops (a viral loop within a product, onboarding, monetization, the referral journey). Growth engineers own both the strategy (the team has no product managers) and the execution of the strategy.

Fyxer’s hiring process starts with a technical screen that’s the same for product and growth. It’s a take-home test that assesses problem solving, mainly related to time zones. This is “open book” – candidates can use AI for the coding test – but the results get compared against an Opus 4.6 baseline. Candidates who pass the technical screen move to behavior interviews and then an in-person whiteboard session around a real-life scenario the team has personally dealt with.

Backgrounds vary quite a bit, although everyone is an engineer with experience outside of engineering whether as a product manager (most common), marketer, ex-founder, or early-stage hire at a fast-growing startup. Data analyst experience is a huge plus given the quantitative side of the role.

Generalists (with unlimited AI token spend) ideate, wireframe, code, test, and then iterate on results. They’re spending less time coding and more time on the idea generation to build stronger hypotheses and customer understanding.

Ultimately, this is a vision for AI-native work that combines velocity with human judgment. And it’s quite impressive. Kameron isn’t resting, by the way; he plans to double the number of growth experiments in 2026: 1,000 tests and 30% of them one-shotted by Tembo.

Related reads: