Guide
Growth in theAge of AI
In an AI product, signing up and making a first payment are easy. Making growth last, and not losing money on every request, are different jobs. This guide collects the questions that tell them apart.
Author: Alp SuleimanPublished
Introduction
Growth and durability don't show up on the same line
The dashboard is going up. The real question is who that curve belongs to.
In AI products, sign-up, trial and first payment are easy. Something easy to buy can be easy to leave; every use also creates a cost; and copying a feature keeps getting cheaper (as our essay argues). Our framing (interpretation): a single ARR line can't tell these three pressures apart.
a16z's view: the leading AI companies don't necessarily have a retention problem, but they do have a measurement problem (a16z also says these companies may reach even better long-term retention). The rest of this guide is about building that measurement: cost first, then who stays, then price, team and defensibility. It ends with one method proposal and three test questions.
Unit economics
Every request has a bill
Is cost falling or rising? Both can be true.
A request sold inside a flat-price plan is like a taxi with an agreed fare and the meter still running. The passenger pays the same amount; the longer the ride, the more fuel you burn. In classic software the marginal cost of a user was usually very small. Here every use sends you an invoice.
There are two readings of where that invoice is heading, and they seem to contradict each other.
- 01
The same work gets cheaper every year
In a16z's LLMflation analysis, the cost of an LLM of equivalent performance falls 10x every year (the yardstick is an MMLU score; OpenAI, Anthropic and Llama models). The author says the methodology is "far from perfect" and the rate may slow down, but still considers the overall conclusion valid.
- 02
But demand moves to the newest model
Ethan Ding's argument (a blog post, not data): users don't stay with the older, cheaper model; they move to the newest one, so subscription margins don't recover even as the unit price falls.
To us the two don't contradict; they measure different things. One is the price of the same work; the other is the price of the work the customer now asks for. The second is the one that matters to you.
Retention
The tourists left. Who stayed?
The early curve may not be a good guide to long-term loyalty.
a16z calls them "AI tourists": users who sign up, try the product for a couple of months and leave. According to Rodriguez and Immerman, this wave makes early retention curves a less decisive indicator of long-term retention. Their fix is simple: rebase retention and CAC from M0 to M3. In their words, M3 represents your true customer base after the tourists have churned out. The curves they looked at mostly flatten around M3.
Mind the sample limit: the piece looks at a selected set of "dozens" of AI companies above $1M in ARR. Rather than assuming your own curve flattens at M3, check where it flattens in your matrix.
The question from this site's piece on the new battle for value applies here too: a competitor shipped something, and how many users left because of it? Interpretation: a retention push that doesn't know why people left risks being an effort to hold on to tourists.
Price
The cost of selling cheap
Do price and retention move together?
In the same analysis Poyar cuts the data by monthly price band, and the gap is large. The table below is his data; the same sample and the same limits apply.
Poyar's line sums up the table: the downside of being easy to buy is being easy to cancel. His suggestion is to separate experimental spend from serious spend, and he proposes counting serious spend as revenue from customers who have paid at least $250 (an annual plan paid upfront, a deal above $250, or a low spender who sticks around).
| Monthly price | GRR | NRR |
|---|---|---|
| Above $250 | 70% | 85% |
| $50–249 | 45% | 61% |
| Under $50 | 23% | 32% |
Median retention by monthly price band (AI-native companies)
Above $250
- GRR:
- 70%
- NRR:
- 85%
$50–249
- GRR:
- 45%
- NRR:
- 61%
Under $50
- GRR:
- 23%
- NRR:
- 32%
Source: Poyar and ChartMogul, 2025 analysis; AI-native products by monthly price band. The post gives no company count per price band; the AI-native sample is small (about 200 scraped companies), so read it as directional.
To us, price band and customer type are tangled together here. A different customer usually buys the expensive product, for a different purpose. The table doesn't show that cutting price breaks retention; it shows that the two go together.
Pricing
Who gets the cost?
The pricing model decides who is left holding the inference bill.
With flat pricing the bill stays with you; with usage-based pricing it passes to the customer; with a hybrid it is shared; with outcome-based pricing the customer pays for the result and you carry the risk (interpretation). In ICONIQ's July 2026 report (surveys from Q4 2025 and Q2 2026, about 300 executives each, self-reported), the share of consumption-based pricing rose from 35% to 42% in six months and outcome-based from 18% to 23%; companies blend 1.7 pricing models on average.
The Pricing Model Picker helps you decide which model fits, with a few questions; this guide doesn't decide for you. Then test the model on your own numbers in the AI unit economics calculator: under flat pricing, find the request count at which your heaviest user turns unprofitable; under outcome-based pricing, find where margin goes if the success rate falls.
Growth team
Speed alone isn't enough; what you measure decides what counts
What is AI changing in the work of a growth team?
One example from Poyar: Fyxer ran 514 experiments in 2025, more than two per work day. In an experiment that added a credit-card prompt to the trial (the paywall was optional), sign-ups dipped and the number of paying customers doubled.
This is one company, as told by Poyar; it does not generalize. The source stops here; the rest is interpretation: a team that watched sign-ups would have called that experiment a failure. Speed sets the number of experiments, the metric sets what counts as a win.
Growth.Talent reports Elena Verna, head of growth at Lovable, saying "Right now I'm spending 95% innovating on growth and only 5% on optimization"; in the same page, Kieran Flanagan of HubSpot argues that AI erodes the boundaries that justified separate growth teams.
Defensibility
What is left once it's copied?
What has to sit under the product for growth to last?
Sequoia wrote in 2023 that generative AI's biggest problem is not finding use cases, demand or distribution, but proving value. Bessemer's State of AI 2025 says something similar: retention can be fragile, especially when switching costs are low, and early hypergrowth alone means less now than ever before.
The piece on this site unpacks that line layer by layer: proprietary data and context, embedding in the workflow, distribution, trust and owning the outcome. To us, each of those layers should show up as a flattening in the retention curve. If it doesn't, what you have may be a feature.
Method
Three changes to what you measure
This section is our proposal, not what the sources say. These combinations don't appear in them; they are hypotheses to test on your own data.
Our proposal (hypothesis)
- 01
Start retention after the tourists leave
Find where flattening begins in your cohort matrix and calculate retention and CAC from that point. a16z suggests M3; the right month for you is the one your matrix shows. The number to compute is CAC per customer still here at that point. The "Start month" option in the Cohort Matrix runs a similar calculation, starting retention and CAC from the month you choose; attributing the CAC to remaining customers is an assumption.
- 02
Plot inference cost by cohort, not just revenue
Next to the retention curve, add a curve of inference cost per user for each cohort. Hypothesis: if the ones who stay are the heaviest users, retention looks good while unit economics get worse. The matrix alone won't show it; the two curves together will. If your inference bill is in dollars and your revenue in lira, take the currency effect out as a separate step.
- 03
Split ARR into two buckets and look for real growth only in the second
Adapt Poyar's experimental versus serious spend split to your own price structure: set the threshold yourself. Then read growth only in the serious bucket and strip out the price effect, as in the Real Growth guide. The aim is to stop experimental spend from calling itself growth.
All three are hypotheses. Try the first version on a small cohort; if your own data disproves it, drop the method.
Test questions
Three questions
Answering takes a minute. Not knowing the answer tells you something.
Are you counting the user who started paying, or the one who will stay?
The gap between the two numbers is, roughly, your tourist count.
Cohort Matrix →Is your heaviest customer the one who earns you the most, or the one who costs you the most?
Under flat pricing, they can be the same person.
AI unit economics calculator →What did your last experiment teach you, and what decision did it change?
If the answer is "sign-ups went up," no decision has changed yet.
Real Growth guide →
Frequently asked
Who is an "AI tourist"?
A term from a16z: a user who signs up for an AI product, tries it for a couple of months and then leaves. a16z writes that, because of them, early retention curves are a less decisive indicator of long-term retention, and it suggests measuring retention and CAC from M3 instead of M0.
If inference cost is falling, why are unit economics still a problem?
a16z says the cost of an LLM of equivalent performance falls about 10x a year, while adding that its methodology is "far from perfect" and the rate may slow. Ethan Ding argues that demand moves to the newest model, so subscription margins don't recover.
Does an annual plan fix a retention problem?
It hides the problem in the number; it doesn't fix it. Poyar notes that NRR on annual plans runs 10-20 points above monthly; a16z says high retention on annual contracts can give false confidence and suggests looking at usage and adoption.
Why split ARR into experimental and serious spend?
In Poyar's data, products in the low price bands show markedly weaker retention, and he himself suggests separating experimental spend from serious spend. The method section of this guide proposes adapting that split to your own price structure, as a hypothesis.
Start
Look at who stayed first, then at what they cost you
This whole guide comes down to one habit: while the dashboard goes up, look at which users stayed and what each of them costs you. Start with one cohort.
Where to start
- If your pricing model is unclear: Pricing Model Picker.
- If you want to start retention after the tourists have left: Cohort Matrix, with the start month option. The tool does not identify who is a tourist; it only calculates from the month you choose.
- If you are unsure about margin: AI unit economics calculator. You can carry the blended monthly revenue and margin from there to the Cohort Matrix; you paste the matrix yourself.
Let's look at your numbers together
If you'd like to go through your own cohorts and unit economics together, we can start with a short exchange.
Talk to Alp →Sources
- 1.a16z — Retention Is All You Need (Rodriguez, Immerman, September 10, 2025)
- 2.Growth Unhinged — The AI churn wave? (Poyar, ChartMogul, December 10, 2025)
- 3.Sequoia Capital — Generative AI's Act Two (September 20, 2023)
- 4.ICONIQ — 2026 State of AI Report: The Builder's Economy (July 2026)
- 5.Bessemer Venture Partners — The State of AI 2025
- 6.a16z — Welcome to LLMflation (Appenzeller, November 12, 2024)
- 7.Ethan Ding — Tokens are getting more expensive (July 31, 2025)
- 8.Growth Unhinged — The AI native growth team (Poyar, March 4, 2026)
- 9.Growth.Talent — AI is killing and creating the growth team (April 11, 2026)