AARRR, Revisited:
Pirate Metrics for AI Products
Five numbers, one funnel, and the order you fix them in. A practical guide to Dave McClure's framework, adapted for products where every active user carries an inference bill.
By Serverless Ventures
Published: October 3, 2026 · Read time: ~9 minutes
The short version
AARRR tracks five stages of user behaviour: Acquisition, Activation, Retention, Referral and Revenue. Its value is not the acronym. It is the discipline of picking one number per stage, finding the stage that leaks the most, and fixing that one before spending on anything else.
For AI products the framework still holds, with three amendments: activation means the first genuinely useful output, retention has to be read after the novelty spike, and revenue has to clear inference cost as well as acquisition cost.
Where pirate metrics came from
The framework comes from Dave McClure, the Silicon Valley investor who went on to found the accelerator 500 Startups. His complaint was that founders drowned in dashboards: page views, follower counts, press mentions, all moving, none of them telling you whether the business was getting healthier. AARRR was his attempt to get startups to narrow their focus to the metrics that directly affect the health of the business. Read the acronym out loud and you get the nickname.
The model is a funnel. People arrive, some of them get value, some of those come back, some of those bring others, and some pay. Each stage is a conversion rate, and because the rates multiply, a weak stage caps everything after it.
The five stages, one at a time
1. Acquisition: how do people find you?
Every channel that brings a person to the product for the first time: search, social, paid campaigns, content, app stores, integrations and partner listings.
- Track: visitors or installs by channel, cost per acquired user, and the share of each channel that goes on to activate.
- Trap: judging channels by volume. A channel that delivers half the traffic but twice the activation rate is the better channel.
2. Activation: do they get value the first time?
The moment a new user does the thing your product exists for. Visiting a second page or signing up for a trial are leading indicators; the real event is the first outcome the user would miss if it disappeared.
- Track: share of sign-ups reaching the activation event, and time to reach it.
- Trap: defining activation as sign-up. That measures the form, not the product.
3. Retention: do they come back?
Repeat use over time: return visits, sessions per week, email opens, continued subscriptions. Retention is the closest thing to a direct measure of product-market fit.
- Track: cohort retention curves (day 1, day 7, day 30) and churn. The question is whether the curve flattens or keeps falling toward zero.
- Trap: blended monthly active users. New sign-ups can hide a cohort that is quietly leaving.
4. Referral: do they bring others?
Users promoting the product through invites, shared outputs, reviews and word of mouth. Of the five stages it is the hardest to measure, because most referral happens in conversations you never see.
- Track: invites sent and accepted, share of new users from referral links, and a simple "how did you hear about us?" on sign-up to catch the dark-social share.
- Trap: paying for referrals before retention is solid. Incentives bring in people who leave as fast as the people who invited them.
5. Revenue: do they pay, and is it enough?
Not just whether a user pays, but whether they pay enough: at minimum, break-even, and ideally well above what it cost to acquire them.
- Track: conversion to paid, average revenue per user, lifetime value (LTV) against customer acquisition cost (CAC), and monthly recurring revenue (MRR).
- Trap: top-line revenue without a cost line. Growth that loses money on every customer only accelerates the loss.
Using it: one number per stage, then the leakiest stage
The standard recipe is four steps: define a conversion metric for each stage, instrument it, run experiments to move it, and shift marketing and product effort toward what works. In practice two habits make the difference.
- Write the definitions down. "Activated" and "retained" must mean one specific event each, agreed before anyone looks at the numbers. Definitions that drift make every trend line a negotiation.
- Work on one stage at a time. Compare each stage's conversion rate against your own history and any honest benchmark you have, pick the weakest, and give it the next quarter.
A worked example (illustrative numbers)
Take 10,000 monthly visitors, a 20% sign-up rate, 40% of sign-ups activated, 35% of activated users retained at day 30, and 25% of retained users paying.
| Scenario | Activated | Retained | Paying | Recurring cost |
|---|---|---|---|---|
| Baseline | 800 | 280 | 70 | None |
| +20% visitors (buy more traffic) | 960 | 336 | 84 | Every month |
| Activation 40% to 48% | 960 | 336 | 84 | Once |
| Retention 35% to 42% | 800 | 336 | 84 | Once |
Because the stages multiply, a 20% relative improvement anywhere produces the same 20% more paying users. The difference is what it costs. More traffic is rented and has to be paid for again next month; a better onboarding flow or a stickier product is paid for once and compounds. That is the real argument for fixing the inner stages first.
The order matters: retention before acquisition
AARRR is listed in the order a user experiences it, not the order you should work on it. Pouring acquisition spend into a product whose day-30 retention curve never flattens is filling a leaking bucket. The RARRA variant, popularised in the mobile growth community, reorders the same five stages to make this explicit: Retention, Activation, Referral, Revenue, then Acquisition. Other teams prepend an Awareness stage (AAARRR) to separate "has heard of us" from "has visited". The variants disagree on naming and agree on substance: get users to stay before paying to bring in more of them.
Where the funnel misleads
- It is linear; growth is often a loop. In products where usage creates content or invites (shared documents, generated media, marketplaces), referral feeds acquisition directly. Model those loops, or the funnel undercounts your best channel.
- It is single-user; B2B buying is not. The person who activates is often not the person who signs the contract. In B2B, track activation at the account level as well as the user level.
- It invites vanity metrics back in. Sign-ups, downloads and page views slide easily into the Acquisition box. Each stage should be measured by a behaviour that predicts the next stage, not by volume.
- It says nothing about why. AARRR tells you where users leave. Interviews, session recordings and qualitative frameworks such as Google's HEART tell you why.
AARRR for AI products: three amendments
Most of the products we build and back at Serverless Ventures are AI-native: agents, copilots, generative media tools. The five stages still apply, but three of them need redefining.
The first output the user would actually use: a merged pull request, a sent draft, an exported clip. Not the first prompt. Measure time to first useful output and the share of first sessions that reach it.
AI products see a novelty spike: people try them because they are new, not because they need them. Judge cohorts from week four onward, and watch whether usage shifts from exploration to a repeated task.
Every active user costs money to serve. Revenue must clear CAC plus inference, so track gross margin per active user, not just conversion to paid.
The revenue amendment changes how you read retention. In traditional SaaS the marginal cost of a highly engaged user is close to zero, so more usage is always good news. In an AI product your heaviest users are also your most expensive. A flat-priced plan can have excellent retention and negative margin in the same cohort. Pricing that tracks usage (credits, tiers, or outcome-based fees) keeps the Revenue stage honest as the Retention stage improves.
Referral, meanwhile, tends to come for free when the output is shareable. A generated video, a report, a design: each one that leaves the product with its origin visible is an acquisition channel. If your AI product produces artifacts people send to others, instrument that path before you build a referral programme.
A one-page checklist
- One written definition per stage, each a user behaviour rather than a volume count.
- Activation defined as the first useful outcome, with time-to-value tracked alongside it.
- Retention read by cohort, and for AI products, from week four onward.
- Referral measured with links plus a "how did you hear about us?" question.
- Revenue reported against CAC and, for AI products, against cost to serve.
- One stage owned per quarter: the leakiest one, with retention fixed before acquisition is scaled.
Background on the framework's origin and the canonical stage definitions from the ProductPlan glossary entry on the AARRR framework. The worked example uses illustrative numbers, not data from a specific product.