What is activation for an AI product?

Activation for an AI product is the first output a user keeps and uses. Not the signup. Not the first prompt.

Most AI teams count the first prompt as activation because it is easy to track and the number looks great. Then they route that number into onboarding emails, sales alerts, and board slides. None of it holds up, because typing into a box is not value. It is curiosity.

The first prompt is curiosity. The first kept output is activation.

Why is the first prompt a bad activation event?

Because it is cheap to do and cheap to give away.

Model costs keep falling. TechCrunch reported on Sep 22 that OpenAI priced its new GPT-6 API tier at roughly half of GPT-5.6 API pricing. On Sep 28, Anthropic pitched Sonnet 5.5 as cheaper and lighter on token burn. Every AI product can afford a more generous free trial now.

That means more people try your product for free, including far more people who were never going to buy. Ehrenberg-Bass and the LinkedIn B2B Institute put about 95% of B2B buyers out of market at any time. A generous trial pulls plenty of that 95% into your signup count. If activation is "sent one prompt," you just called them all activated.

I have lived the vanity version of this. 28,000 organic views, 4 website visits, zero sales. Big top of funnel, nothing underneath. A first-prompt activation metric is that same trap inside the product.

What counts as a kept output?

An output the user did something with after the model produced it.

Pick the one that matches the job your product does:

1. Copied or exported. They took the draft, summary, or table out of your product and into their work.
2. Sent or published. The email went out. The page shipped. The report went to their boss.
3. Accepted without a rewrite. They approved it as-is, or with light edits, instead of regenerating five times.
4. Run again within a week. They came back and did the same job again. One run is a test. The second run is a habit starting.

You do not need all four. You need one, defined in writing, that your team agrees means "this person got value."

How do I pick the right event for my product?

Ask what the user would pay you for, then find the moment that thing first happened.

If you sell research, activation is the first researched account someone saved to their CRM. If you sell writing, it is the first draft that got sent. If you sell call summaries, it is the first summary shared with someone who was not on the call.

Then check it against retention. Pull the users who hit your candidate event in week one. Compare them to users who only sent a prompt. If the kept-output group is not coming back more, you picked the wrong event. Try the next one.

Any product analytics tool can run this. PostHog is one example: track the event, build a cohort off it, and compare retention against a prompt-only cohort. The tool matters less than the definition.

Why is activation a Signals problem, not just a product metric?

Because every downstream signal inherits the definition.

In how product marketers use PostHog for GTM signals, activation is one of four event families I watch. In product-led sales for AI products, usage earns the human touch. Both assume activation means something. If it means "sent a prompt," your sales alerts fire on tourists and your team learns to ignore them.

Get the definition right and it works the other way. A fit account that hit the kept-output event this week is a real buying window candidate. A fit account that typed one prompt and left is not.

What should onboarding do with this?

Point every onboarding step at the kept output, not at more prompts.

Pre-fill a real example from their use case. Show the export or send button next to the first result. Cut steps between "output appears" and "output leaves the product." Measure time to the first kept output, not time to the first prompt.

Then rewrite the dashboard. Signups, prompts, and kept outputs as three separate numbers. The gap between the second and third is your onboarding problem, and it is usually bigger than anyone wants to say out loud.

Count what they keep. Ignore what they type.

Start Signals, Convert, Grow

FAQ

What is activation for an AI product?

The first output a user keeps and uses, such as copying, exporting, sending, or approving it, or running the same job again within a week.

Why not count the first prompt as activation?

A prompt is cheap and mostly curiosity. As model costs fall and free trials grow, prompt-based activation counts more tourists and hides the users who got value.

How do I validate my activation event?

Compare week-one retention for users who hit the event against users who only sent a prompt. If the event group does not come back more, pick a different event.

Can I track this in PostHog?

Yes, PostHog is one example. Track the kept-output event, build a cohort from it, and compare retention. Any product analytics tool that supports events and cohorts can do this.

How does activation connect to sales?

Product usage alerts and product-led sales triggers inherit your activation definition. A kept-output event from a fit account is a real signal. A single prompt is not.