The Only 5 Metrics a Pre-Seed AI Product Needs
Most early dashboards track too much. The five metrics that actually change a pre-seed AI product's decisions, and the vanity numbers to stop watching.
Published 2026-08-16 · 4 min read
A pre-seed AI product does not have a measurement problem. It has a focus problem: founders install analytics, build dashboards, and then make the same decisions they would have made anyway because none of the numbers change what to do next. Five metrics do change decisions. The rest are decoration at this stage.
These are the five, why each one, and what to do when it moves.
The five
| # | Metric | The question it answers | Act when |
|---|---|---|---|
| 1 | Activation rate | Do people reach the value? | It is below where your converters started |
| 2 | Week-2 retention | Do they come back? | It is near zero — you have a product problem, not a growth problem |
| 3 | Where traffic comes from | Which channel produces the people who stay? | One channel clearly outperforms on activation, not volume |
| 4 | Cost per active user | Does serving them cost more than they are worth? | It exceeds what a user is worth, especially with per-use AI costs |
| 5 | Conversations per week | Are you still learning? | It drops to zero — you have stopped talking to users |
1. Activation rate
The share of new signups who complete the action that indicates they got value — not just signed up, but did the thing. This is the single most important early metric because it separates a distribution problem from a product problem.
- Define it as one specific action, found by comparing users who stayed against those who left and locating the behaviour that separates them.
- Low activation with good traffic means fix the product, not the marketing. More visitors to a leaky bucket is not growth.
- High activation with low traffic means the product works and the constraint is distribution — which is the good problem to have.
2. Week-2 retention
Do people who activated come back? Retention at week two is early enough to measure quickly and late enough to mean something. A product people do not return to has no business model regardless of how many signups it collects.
- Near-zero week-2 retention is a product signal, not a marketing one. Stop every growth activity and find out why they leave.
- Retention is the multiplier on everything else. Traffic times activation times retention is roughly your durable user base. Improving retention raises every channel's value at once.
- A small cohort that retains beats a large one that does not. Do not chase signup volume until week-2 retention is respectable.
3. Where traffic comes from
Not how much — from where. Segment by source, and because several platforms and AI assistants strip UTM parameters, use referrer as the reliable method.
- Judge sources by activation, not volume. A channel sending twenty people who activate beats one sending two hundred who do not.
- Watch for AI-assistant referrals separately. This traffic is growing and is invisible if you only look at search and social.
- The output is a decision: do more of the channel that produces activated users, and stop the ones that do not.
4. Cost per active user
The metric most pre-seed AI products ignore and most need. If your product calls a model on every use, serving a user costs real money, and a growing free base can be a growing loss you have not looked at.
- Compute it from your actual inference and infrastructure bills, divided by active users — not signups.
- Compare it to what a user is worth. If cost per active user approaches or exceeds the value, the pricing model or the free tier is broken.
- This is the metric that makes the free-versus-paid decision concrete. See the companion articles on when to stop being free and on usage-based pricing.
5. Conversations per week
The least dashboard-like metric and the one that matters most early. How many real conversations did you have with users or potential users this week? At pre-seed, learning speed is the constraint, and learning comes from conversations.
- Zero conversations in a week is a leading indicator of building the wrong thing. Every founder who shipped something nobody wanted can trace it to a period where they stopped talking to users.
- It is a metric you control completely, unlike the other four. If it drops, that is a choice, and it is the easiest one to reverse.
- It feeds the other four. Activation definitions, retention causes and channel fit all come out of conversations, not dashboards.
What to stop watching
| Vanity metric | Why it misleads early |
|---|---|
| Total signups | Counts people who clicked, not people who got value |
| Pageviews and sessions | Rises with traffic regardless of whether the product works |
| Follower count | Peers and bots inflate it; it does not correlate with users |
| Revenue without cost | For a per-use-cost product, revenue alone hides whether you are profitable per user |
| Daily active users, pre-fit | Meaningless if you have not defined what activity indicates value |
| Bounce rate | Coarse and often misattributed; activation tells you more |
The weekly review
Five numbers, looked at once a week, each with one question:
- Activation: did it move, and did any onboarding change cause it?
- Week-2 retention: is the cohort that activated two weeks ago coming back?
- Traffic source: which source produced activated users this week?
- Cost per active user: did the bill move faster than the users?
- Conversations: how many did I have, and what did I learn?
That is the whole review. It takes fifteen minutes, and every number in it points at a specific action rather than a feeling about how the week went.
A metric is worth tracking if it changes what you do. At pre-seed, five do. The dashboard with forty tiles changes nothing — it just makes the week feel measured.
More in Measuring growth
- How to Tell Whether Your AI Launch WorkedA launch feels like a success or a failure based on the spike, but the spike misleads. The four questions that actually tell you whether a launch worked.
- How to Run a Weekly Growth Review as a Solo FounderA 30-minute weekly review that turns scattered metrics into one decision: what to look at, what to ignore, and how to keep it from becoming a ritual.
- How to Reuse a Failed Launch as a Second LaunchA launch that flopped leaves feedback, a fixed product and material. How to turn a failed first launch into a stronger second one, not start over.