Key Takeaways
- A launch brings sign-ups, not users. Launch week brought 182 sign-ups, about 4 times a normal week. The number of people using Kai on 3 days a week went from 17 in early September to 18.
- Ads can't tell you who your product is for. 4 positions on Google cost us €2,772 and brought 36 sign-ups and 0 active users. Jim stopped them on 26 October.
- 5 calls per group, selling nothing. Jim now reaches out to real estate agents, fractional executives and agency owners to learn what they struggle with. 5 conversations are enough to know if there's something there.
- Define value before you count it. Our new North Star counts value days: days Kai did something for you that you'd notice. Kai sorting the inbox of someone who never opens the app doesn't count.
- 3 value days in a week is the bar. 93% of people who reach it come back for value the next week, against 43% with 1 value day.
Introduction
This is How We Grow, where Jim and I write up what we're doing to grow Kai. We recorded #019 at Jim's place on Friday 30 October, with a new setup: a camera on each of us and a GoPro above the table. It was my turn to lead, and I wanted Jim on the record about the 3 weeks since our Product Hunt launch, because they changed how the whole team works.
Where things stand: launch week brought 182 sign-ups, and most of those people didn't stay. Jim describes the mood as "lost and confused, but also excited". We're back in founders' shoes, looking for the people Kai is really for.
Prefer to watch or listen? Here's the full episode
The 3 weeks after the launch.
The launch brought sign-ups, not users
182 sign-ups in a week, 18 people using Kai every week
We finished 5th on Product Hunt on 13 October. Around it we ran everything we could: the launch itself, emails to Morgen's users, posts on our social accounts, and a lot of work on onboarding and on the paid plans. Launch week brought 182 sign-ups, about 4 times a normal week.
The number we care about didn't follow. People who used Kai on 3 or more days of the week were 17 on 2 September and 18 on the day we recorded.
Sign-ups per week, and people using Kai 3 days a week.
No single job people came for
Jim had been talking to users for weeks before the launch, and the launch made the pattern hard to miss. Everyone used Kai for something different. Kai can do a lot, and that's the problem: there was no one job we could take to the next 1,000 people and say "do this with Kai". No use case, so no persona, so no message that works.
Why we're back to discovery.
So the team is starting again from the market: who exactly Kai is for, what those people are trying to do, where their current tools fall short, and why Kai would be the obvious fix. Jim expects the product to change a lot once we know.
Why Jim stopped our ads
4 positions, €2,772, 0 active users
Over the summer we tested positions with ads on Google and Meta. Each one was a guess about who Kai is for: people with ADHD, AI-native builders, people who need an assistant and can't justify hiring one. On Google alone, from August to October, 4 positions cost €2,772 and brought 36 sign-ups. 3 of those people got value from Kai on their first day. None became active.
Our Google ads, August to October.
Jim stopped all 4 on 26 October. He gave 2 reasons on the episode.
The guesses were too broad. "You have ADHD" isn't a use case. "When you don't know what to do next, Kai tells you right away because it's connected to everything" is one, and we never got that precise.
Ads are too far from people. Every step of the funnel can fail: the creative, the landing page, the targeting, the message, whether people are ready to buy. Jim's math: test 3 angles at 200 sign-ups each and you have 600 sign-ups, about 60 people who stay and maybe 1 or 2 who pay. Without a huge budget, that teaches you nothing you can trust.
Why an ad test teaches so little.
5 calls per group, before anything else
Who we're calling
The clearest group among the people who stayed after the launch isn't tech founders, the persona we went after first. It's founders of small agencies, or people running 3 or 4 small businesses at once. Students, PhD students and people with ADHD stayed too, but there's not enough volume to read anything into it. As Jim put it, if we had to bet our life on one group, it would be those founders.
Who stayed after the launch.
So he's calling 3 groups: real estate agents, starting in California, fractional executives and agency owners. The calls aren't about Kai. They're about anything these people struggle with, emails and meetings first, but anything goes. If real estate agents turn out to be drowning in legal paperwork, that's worth knowing too.
The plan: 5 calls per group, selling nothing.
How Jim runs the outreach
From a list to a call, in 7 steps.
- Build the list with Claude. Jim uses MoltSets, an API he connected to Claude, and asks in plain words: fractional CFOs in Europe holding 3 jobs at once, find 500. It comes back with names, emails and LinkedIn profiles.
- Reach out in steps. Visit their LinkedIn profile, send an invite, send an email, send a LinkedIn message, send another email.
- Sell nothing. The message is "we're founders building something for people like you, want to hop on a call?"
- Stop at 5 calls. 5 conversations per group are enough to know whether there's something there. If not, he moves to the next group.
- Build for whoever leans in. When someone says "wait, you can solve that?", the team builds it and goes back to them within days.
Calls with our own users run in parallel, but Jim noticed a bias: those people signed up for a reason, and they're a narrow slice of the market. The cold groups are there to correct for it. When we recorded, the first internal calls were done and the cold campaigns had gone out less than 24 hours before. From experience, Jim expects a week before the first call and 2 or 3 calls per 100 people reached.
I asked whether Kai should stay a tool for one person or become one for a team, since an accountant with the problem likely has colleagues with the same one. Jim's answer: it depends on the use case, and the calls will tell us.
Our new North Star: the value day
Same users, different answers
After the launch, several of us analysed who was getting value from Kai and came back with different answers. David counted 15 people getting value, Jim got a different number, because each of us defined value our own way.
The old North Star dated from before the product was finished. It counted agentic actions: any time Kai used one of its tools. The problem showed up fast. A lot of people connected their inbox and never came back, while Kai kept sorting their email in the background. Those runs counted as value, and we paid for them.
What our North Star counted, then and now.
One unit for everything
Jim rebuilt our metrics around one unit, the value day: a day Kai did something for you that you'd notice. A task, an event or an email Kai made because you asked, a meeting you recorded, an action on something Kai flagged or drafted. Kai working in the background doesn't count, and neither does chatting or reading a brief.
Everything else is built on it:
- Active: 3 or more value days in the last 7 days. Our North Star is the number of people who are active this way.
- Activated: 3 value days in your first 7 days.
- Retained: a value day in your second week.
What counts as a value day, and what's built on it.
The bar is 3 because people who reach it keep coming back.
Back for value the next week, by value days this week.
What the new number showed us
Since August, 4.4% of sign-ups reached 3 value days in their first week. Split by where people came from, the difference is stark: 16% of people who used Morgen before trying Kai, against 2.5% of everyone else.
Active in their first week.
Paid traffic showed the gap best. Those people finished onboarding and connected their inbox, and never had a single value day. Jim's point: if Kai were built for the people we were reaching, activation would be 50% or 60%, not around 5%.
Where activation is, and where Jim wants it.
The honest caveats
- The numbers are small. At 18 people, 1 person is 6% of the North Star, so we read it over 4 weeks, never day to day.
- September reads low. Acting on an email Kai flagged went unlogged for part of September and October, so value days from that stretch undercount.
- Calls are slow. A week to the first call, 2 or 3 calls per 100 people, and 5 calls might show nothing.
- We're giving up volume. Product-led growth needs a lot of sign-ups to conclude anything. Moving to conversations means fewer, deeper signals, and slower sign-up growth while we look.
What's next
Jim runs the discovery calls, group by group, and the team builds for the first group that leans in.
On my side, our content moves from selling Kai's features to selling a workflow: one reel a day built around a single job, with the workflow sent to whoever comments.
Our content plan.
I also brought Jim an idea from another founder, a €0.99 trial instead of a free one. Fewer people tried the product, and the share who stayed and paid doubled. Jim likes it as a filter for the merely curious, but we haven't decided anything yet.
The trial idea, in that founder's numbers.
Our next hackathon is in about 3 weeks, and the next episode may be a special one with David.
If you want to follow along, the whole series lives at How We Grow, and the changelog shows what we ship week to week.
Find your activation rule with Claude
We picked 3 value days within a user's first 7 days because 93% of the people who reached 3 value days in a week got value again the next week, against 43% with 1. Your product's number may not be 3. We've written a check like it as a skill you can run on your own users: open it, copy it into ~/.claude/skills/activation-rule/SKILL.md, then give Claude a CSV of your users' events and signup dates and ask for it by name.
Activation ruleTurns an export of your product's events into an activation rule: how many value days in the first week predict who stays.
---
name: activation-rule
description: Find your product's activation rule from an export of user events. Defines a "value day" (a day a user got value from something they did themselves), counts value days per user in their first 7 days, tests thresholds of 1 to 4 against who is still active in week 2 and week 4, picks the rule that separates best, and reports the activation rate with small-sample caveats. Use when you say "define our activation metric", "what's our activation rule", or hand over a CSV of product events.
---
# Activation rule
You hand over a CSV of your product's user events. This skill defines a value day with you, counts how many value days each user had in their first 7 days, and finds the number that best separates the users who stay from the users who don't. The result is a rule of the form "activated means at least N value days within 7 days of signup", plus the share of signups who meet it.
A team that ran this on its own data landed on 3 value days within 7 days. Yours may land elsewhere. The data decides.
Talk to the person in their language, write every file in English.
## What you need
1. **The events**, one row per event: user id, timestamp, event name. If there's a column saying whether the user did it or the product did it (a source, a trigger, "automatic"), include it.
2. **The signup date** of each user: one row per user, user id and signup timestamp. If the export has no signup table, use each user's first event and say that you did.
3. The export should run at least 4 weeks past the oldest signups you want to read. Note the date of the latest event in the file.
Ask for what's missing in one message, then start. Work in a scratch folder with Python and pandas (check with `python3 -c "import pandas"`, offer `pip install pandas`, install only on a yes). Never edit the original export, and never print an email or a name, only user ids.
## 1. Look at the data
1. Count the rows, the users, the first and last timestamp, and list every event name with its count (the top 40).
2. Convert every timestamp to UTC. Remove exact duplicates (same user, time and event) and say how many.
3. Drop events dated before the user's signup and say how many. Keep users with no events at all: they have 0 value days and they stay in every count.
4. Tell the person what you found in a few lines before going on.
## 2. Define the value day
A value day is a calendar day on which the user got value from an action they took themselves.
Go through the event names with the person, one at a time, and ask: "Did the user do this on purpose, or did the product do it for them?" Use this test: if the user were offline all day, could this event still happen? If yes, it doesn't count.
- **Counts:** finishing the thing the product exists for (sending, exporting, publishing, running a report, recording a meeting), and acting on something the product suggested, because the user decided.
- **Doesn't count:** logins, page views, onboarding steps, settings, reading a notification, and anything the product runs in the background or on a schedule. Background work is the trap. One team's first metric counted its product sorting inboxes in the background, and people who never came back looked active.
Ask about every event whose name doesn't settle it. Write the final list of counted events, show it, and wait for a yes before computing anything.
## 3. Count value days in the first 7 days
The first 7 days are day 0 (the signup day) to day 6, in UTC. A user's value days are the distinct calendar days in that window with at least one counted event.
Keep only users whose window is over. Then show how many users had 0, 1, 2, 3, 4 and 5 or more value days.
## 4. Define "still active"
- **Active in week 2:** at least one value day on days 7 to 13.
- **Active in week 4:** at least one value day on days 21 to 27.
Use the same value-day list as in step 2, so a login never makes anyone active. Only users old enough to have lived the whole week count: 14 days since signup for week 2, 28 days for week 4. Say how many users are in each group.
## 5. Test the thresholds
For each threshold N from 1 to 4, split the users into those with at least N value days in their first 7 days and those with fewer. Report this table for week 2 and again for week 4:
| N | Users who meet it | Still active (met) | Still active (not met) | Gap in points |
|---|---|---|---|---|
Under each table, give the share of all signups who meet N.
## 6. Pick the rule
1. Throw out any threshold that fewer than 20 users meet. A share measured on 10 people moves 10 points when one person changes.
2. Among the rest, take the largest gap in points at week 4, and check that week 2 ranks them the same way.
3. If two thresholds are within 5 points of each other, take the lower one. More users reach it, so it's steadier and easier to act on.
4. If no threshold opens a gap of at least 20 points, say there's no clear activation rule in this data. Possible reasons: the value-day list is wrong, the product doesn't have a habit yet, or the sample is too small. Don't force a rule.
State the rule in one sentence: "Activated means at least N value days within 7 days of signup."
## 7. Report
Give the person:
- The value-day definition: the list of counted events and the ones left out, with the reason.
- The two tables.
- The rule, and the activation rate: the share of signups who meet it, over all signups whose 7-day window is over. Add the same rate by signup week for the last 8 weeks, each week with its number of signups.
- The caveats, every time:
- Every share comes with its n. A share on fewer than 30 users is a direction, not a number (3 of 8 reads 38% and could be a lot of things).
- A week with fewer than 20 signups isn't a trend. Read the rate over 4 weeks.
- If the file has a source column (paid, referral, organic), split the rate by it. Different groups can hide different rules.
- The rule predicts who stays. It doesn't prove that pushing someone to hit it makes them stay, because people who would stay anyway also do more early.
- A counted event that wasn't logged for some weeks makes those weeks undercount. Check each counted event's weekly count for holes.
- What to do next: aim onboarding at the rule, and run this again after any big product change.
## Never
- Count something the product did without the user (background, scheduled, automatic).
- Count a login or a page view as value.
- Pick a threshold that fewer than 20 users meet.
- Report a rate without its n.
- Edit the original export, or print a user's email or name.