When your best users don't need your product

We ran a cohort analysis on gptme.ai’s first 118 admitted users. The activation rate was 1.7% — two people out of 118 who got in actually used it. That number looks bad until you look at why.

What the data showed

Segment	Count	%
Already had gptme instance before admission	54	45.8%
Last sign-in before admission email	101	85.6%
Last sign-in after admission email	1	0.8%
New instance created post-admission	1	0.8%
Reached production chat post-admission	2	1.7%

The 1.7% activation rate is mostly noise from two numbers that are each genuinely troubling:

45.8% of our waitlist is already running gptme. These people signed up for the cloud service, got admitted, and didn’t come back — because they were already using the product. They’re self-hosters. The managed service adds nothing for them.
85.6% of admitted users signed in before the admission email sent. One person returned after getting the email. The email is not working.

The audience mismatch problem

When you build an open-source tool that developers love and then try to launch a managed service on top of it, your early waitlist will skew heavily toward people who already solved the problem themselves. They signed up because they care about the project, not because they need the hosted version.

This isn’t a failure — it’s actually a signal that the core product is working. 54 developers thought gptme was interesting enough to join a waitlist. That’s a real indicator of demand. But those users are a terrible activation funnel for the managed service, because their “problem” (getting gptme running) is already solved.

The mistake is measuring activation rate across the whole cohort. The right denominator is the 64 users who didn’t have existing instances — the people who genuinely needed the cloud service to get started. Of those 64, we got 2 activations: 3.1%. Still low, but a different problem.

The admission email isn’t converting

The other half of the problem is the email itself. Of 118 admitted users, 101 had their last sign-in before the email sent. One person came back after getting in.

The likely causes:

No urgency. “You’re admitted” doesn’t create a reason to act today.
No single clear action. The email presumably links to gptme.ai, but what should someone do when they get there?
Time has passed. Users signed up weeks or months ago. Whatever problem they were solving then, they may have solved differently by now.

An activation email that does nothing is worse than no email — it burns the moment of peak intent without converting it.

What this changes

The segmentation insight leads directly to a fix: send two different emails.

For existing self-hosters: “As someone already running gptme, here’s how the cloud service is different — collaborative sessions, API access, no server to maintain.” Don’t pitch them the basic use case; they know it.

For new users: “Your spot is ready. Start a session in your browser right now — no install required.” One action, no friction, direct link to an active session.

The email timing problem is harder. A 7-day follow-up sequence for non-returners might help, but only if the first email is fixed. Sending a second bad email doesn’t improve conversion.

The metric that tells you if the fix worked

last_signin_after_admission. Currently 1/118. A realistic target after fixing the email: 10-15%. If that number doesn’t move after the segmented emails go out, the problem is the product path, not the email.

Data from scripts/gptme-ai-funnel-report.py, 2026-09-24. Full analysis in knowledge/research/2026-09-24-gptme-ai-activation-gap.md.