For most of my career the working mental model of growth was a funnel. Pour users in at the top, watch them narrow through acquisition, activation, retention, revenue, referral, and count what survives. It was the right model for its time, and it carried an assumption so basic nobody bothered to state it: the product is the same product for user one and user one million. Growth meant pushing more volume through a fixed pipe.
AI-native products break that assumption. Every session generates signal the product can learn from — the prompt that failed, the suggestion that got accepted, the output the user edited before shipping. A product that retrains on its own usage is not a fixed pipe. It is a loop: usage produces data, data produces a better model, a better model produces more value, more value produces more usage. The output of one cycle is the input to the next.
Data network effects are not new — we wrote about them in Hacking Growth. What is new is the cycle time. The old loop turned quarterly: collect data, retrain, ship. The AI-native loop can turn nightly, or per session. When the loop spins that fast, the compounding asset is no longer the channel or the campaign. It is the model. Your acquisition spend depreciates the day you stop paying; the loop appreciates while you sleep.
Virality changes shape too. The viral factor is still the same arithmetic — invitations sent times conversion rate — but AI moves both terms at once. When the product's output is an artifact people want to share, every artifact is an invitation, so i climbs without anyone designing an invite flow. And because the shared artifact demonstrates the product better than any landing page could, conversion climbs with it. The output is the ad.
The unit of compounding isn’t the campaign anymore. It’s the model.
The discipline that matters most, though, is the one that always mattered: tempo. The core argument of high-tempo testing was that the team that learns fastest wins, because learning compounds the same way interest does. AI raises the ceiling on throughput — agents can draft variants, run the analysis, and write up the result — which means the bottleneck moves from running experiments to choosing them. Judgment about what to test becomes the scarce input. That is a promotion for the growth lead, not a layoff.
So what do you instrument? Not just the funnel stages. Measure the loop itself: how much usable signal each cohort generates, how long it takes that signal to show up as product improvement, and whether time-to-value is falling for the cohorts that arrive after a retrain. If the product is genuinely learning, next month's users should activate faster than this month's — without you touching the onboarding. That curve, not CAC, is the health metric of an AI-native loop.
I don't think anyone has the full playbook yet, including me. The funnel took a decade to formalize; the loop will get formalized faster, because the people building these products are also the people measuring them. This page is me thinking out loud about what goes in that playbook. If you're running one of these loops in production, I want to compare notes.