An AI product looks like a SaaS product from the outside and behaves like a metered utility from the inside. Revenue arrives as subscriptions and credit packs, but cost arrives with every request, and that cost does not care which affiliate sent the customer. A program copied from a $49/mo project management tool will pay a percentage of revenue while ignoring the fact that some customers consume 60 percent of their payment in inference costs.
The tracking mechanics are the same as any other SaaS affiliate program: a link, a cookie or a saved referral, a webhook from Polar, Stripe, or Paddle, and a dashboard that reconciles. What changes is the definition of the number you share, the speed at which customers churn, and the number of ways a free tier invites abuse. Here is what to decide before you announce a rate.
TL;DR
- Commission on revenue is not commission on profit. On an AI product, heavy users can erase the margin a percentage is drawn from, so set the rate against gross margin, not against competitor benchmarks.
- Decide whether credits earn commission when they are purchased or when they are consumed, then write it down. The two answers produce very different cash timing.
- A free tier converts beautifully and invites self-referral. Block duplicates on identity, not on intent, and say in your terms what happens when one person runs three accounts.
- AI buyers increasingly arrive through agents, API docs, and listicles rather than a browser session, which makes saved referral records and server-side postbacks more important than a 90 day cookie.
- Churn and price cuts arrive faster here. A clawback window with a published definition prevents the argument you would otherwise have in month four.
Commission on a product whose costs move with usage
Start with the number the affiliate gets paid on, because three different numbers live in every AI invoice.
The first is what the customer paid. This is the only number a webhook can hand you with certainty, and it is the number every affiliate platform uses by default. Polar reports a net amount on a paid order, Paddle reports a subtotal in minor units, and Stripe reports an invoice total. All three are pre-tax or can be read that way, and none of them know anything about your inference bill.
The second is what you kept after variable costs. A customer on a $40 plan who burns $34 in model calls contributed six dollars of gross margin. Pay a 30 percent commission on the $40 and you have paid twelve dollars for the privilege of losing six. That is not a fraud problem, it is an arithmetic problem, and it shows up on your heaviest, most enthusiastic users first.
The third is what you recognize over time, which matters when credits are prepaid. Most teams land on a revenue-based commission low enough to survive a bad month of usage, plus an explicit cap on very large accounts: the arithmetic stays honest without asking affiliates to trust a margin figure they cannot verify.
| Commission base | What it means | Where the number comes from | Main risk |
|---|---|---|---|
| Amount collected | Percentage of the invoice the customer paid | Provider webhook, pre-tax field | Ignores heavy usage, can exceed margin |
| Margin share | Percentage of revenue minus variable cost | Your own cost ledger plus the webhook | Affiliate cannot verify it, disputes get personal |
| Per activation | Flat amount on first paid invoice | Webhook plus your product event | Caps upside, cheap for you on big accounts |
| Tiered by plan | Different rate per price tier | Plan id on the webhook | Plan id drift after a pricing change |
| Capped percentage | Percentage up to a dollar ceiling | Webhook plus a rule in your terms | Ceiling must be visible before signup |
Whichever row you pick, keep the commission on pre-tax money. Tax was never yours, and a merchant-of-record invoice makes that especially easy to get wrong. The same reasoning runs through Affiliate Commissions on Usage-Based Pricing, which covers the invoice as the unit of pay in more detail.
Credits, prepaid packs, and what counts as revenue
AI products sell credits for a reason: they front-load cash and protect you from a customer who disappears mid month. They also create the single most confusing commission question in the category.
If commission lands on the purchase, an affiliate earns once, immediately, on money the customer may not consume for three months. If it lands on consumption, the affiliate earns in small slices, the payout schedule gets noisy, and you need a ledger mapping credit draws back to the referral. There is no universally correct answer, but there is a correct process: pick one, state it in the terms, and make the dashboard match it.
Two edge cases deserve a clause. Expiring credits are not revenue for a service delivered, and some teams exclude them from commission. Refunded credit packs are worse, because a refund can arrive after a commission was paid, and the reversal has to pull back money that left the building. When the Refund Beats the Sale covers that mechanic, and Affiliate Commissions on Trials and Freemium Plans covers the payout shape.
A working default: pay on the credit purchase, hold the commission in a pending state until your normal refund window closes, and record usage-based adjustments in the ledger rather than editing past commission amounts. Numbers that change retroactively destroy trust faster than numbers that are simply low.
The free tier is your demo, and it is also the abuse surface
Free tiers are how AI products get discovered. They are also the cheapest place in the world to fabricate a referral.
Three patterns show up repeatedly. One person signs up twice, with a referral link and without, to test whether the payout is real. An affiliate creates throwaway accounts to inflate signup counts. A coupon site registers accounts to claim credit for customers who were going to buy anyway. All three are visible in data, and none in a click count.
The defenses that hold are identity-based. Block a referral when the affiliate email matches the referred account email. Rate limit referred signups per IP and device. Require a payment method before a referral counts. And keep a saved referral on the account, so a customer who upgrades in month three still credits the right partner.
Say the rules out loud in the terms, including what happens when you find three accounts from the same person: the accounts are merged, the duplicate referrals are voided, and a repeat offense ends the partnership. Affiliates accept enforcement when the rule was written before the incident. The broader catalog of patterns is in 10 Affiliate Fraud Tactics.

The five kinds of people who promote AI tools
Recruiting for an AI product is different from recruiting for a dev tool, because the audience research happens in public and the tool changes every few weeks. Five profiles are worth separating, because they need different assets and different rates.
| Affiliate type | Where they publish | What they actually convert | Asset they need |
|---|---|---|---|
| Tutorial creators | YouTube, short video, course platforms | Viewers who want a workflow, not a tool list | Demo project files, a current UI walkthrough |
| Newsletter writers | Niche newsletters with a real open rate | Readers who already trust one voice | Exclusive angle, a working discount, no generic copy |
| Comparison sites | "Best AI tools for X" pages, directory listings | Buyers one step from a decision | Accurate pricing table, trial length, export format |
| Agency and consultant partners | Client engagements, internal playbooks | Teams who need deployment help | Multi-seat pricing, referral for the agency itself |
| Power users | Communities, prompt libraries, group chats | Peers replicating a workflow they watched work | Share kit with screenshots that are not out of date |
The fifth group is the one AI products underuse. A customer who has shipped something with your tool has credibility no publisher can buy. Just date your share kit and version it: screenshots and pricing go stale in weeks here.
Listicles, deal sites, and the coupon layer
The affiliate layer of the AI category is dominated by list pages. "Best AI writing tools", "best AI meeting notes", "cheapest AI API" and a hundred variants have real search demand, and the people running them are professional affiliates who will test your program in a week.
That is fine, with two rules. Coupon codes leak: once a discount code exists it gets scraped onto coupon sites, and a buyer already on your pricing page can arrive through a coupon page and hand credit to someone who added no value. Brand bidding turns your own name into a paid placement you would rather not subsidize. Both answers are defensible, but they belong in the terms.
There is a third wrinkle: "best tools" pages are often updated by AI and sometimes written without the operator reading the output, so your listing may describe features you do not have. Keep a public facts page with pricing, model options, and limits, and point reviewers at it.
Attribution when the buyer is an agent
AI products are increasingly bought by software. An agent reads your docs, picks an endpoint, and a human approves a payment. Some of that traffic never renders a page, never keeps a cookie, and never sees an affiliate link.
No provider webhook solves this. All three report a payment with whatever metadata you attached at checkout, and none know which page the agent read. What you can do is make the referral travel with the request: a parameter on the signup route, a custom field on the checkout, a saved record on the account. Affiliate Postbacks and Server-Side Tracking is where that plumbing lives.
Be honest about the ceiling. An agent that arrives with no referral parameter is an organic customer, and claiming otherwise would be fiction. What you can promise is that no referred signup is lost to a cookie: if the referral was attached at account creation, later payments inherit it.
Churn and refunds arrive faster here
AI subscriptions churn on novelty and on price. A model release changes what your product can do, a competitor cuts prices, and a customer who loved you in March has cancelled by June. Add a usable free tier and you get a large population of accounts that will never convert, fine for the funnel and bad for a dashboard that counts them as progress.
Two consequences follow. Recurring commissions are worth less per signup than on a sticky vertical tool, so recruit for volume and expect shorter lifetimes. And publish the reversal rule: a common structure is a clawback window aligned to your refund policy, applied to the commission and not to the affiliate's reputation, with anything cancelled after the window keeping its earned commission.
If you allow prepaid credits, add one more rule: a chargeback on a credit pack voids commission on that pack and any commission already paid from credits it funded. That is the sentence that stops a fraud ring from monetizing your program.
What to pay, and what to cap
Rates in this category are wide, because gross margins are wide. A pure API wrapper and a model you host yourself have almost nothing in common economically, even if both sell a $30 subscription.
| Program shape | Typical rate range | Best fit | Watch out for |
|---|---|---|---|
| Recurring percentage | 20 to 30 percent | Stable subscription with healthy margin | Heavy usage accounts, long tails |
| First year only | 30 to 40 percent | Fast churn, or a large one-time setup | Affiliates who plan for lifetime income |
| Flat per conversion | $20 to $80 | Low price point, high volume | No upside for the affiliate on big accounts |
| Credit pack percentage | 10 to 20 percent | Prepaid consumption products | Refund and expiry handling |
| Hybrid | Flat plus small recurring | Enterprise or agency referrals | Two numbers to explain on every dashboard |
Set the rate by working backwards from margin on your worst realistic customer, not your best. If a 30 percent commission leaves you negative on a heavy user, either the rate is too high or that plan is priced wrong, and both are easier to fix before launch. Sled ships a 30 percent default with a per-affiliate override, so an agency partner can carry a different rate than a YouTube creator.
Minimums matter more than rates for small affiliates. A $20 floor covers fixed transfer costs, and six payout methods (PayPal, Wise, Payoneer, Revolut, bank transfer, crypto) mean the affiliate is not paying a fee to collect a fee.
Terms clauses worth having on day one
A short list, taken from programs that survived their first year.
Define the commission base in one sentence, including tax and discounts. Say what counts as a conversion: first paid invoice, not signup. Describe the reversal window. State the self-referral and duplicate-account rules. Name the off-limits channels, usually paid brand search and cookie stuffing. Require affiliate disclosure and a real trial behind any review. Reserve the right to pause a partner for fabricated claims without touching their earned balance.
The AI-specific additions are two. Ask affiliates not to publish generated screenshots or synthetic reviews of your product, because a fabricated UI screenshot is a support ticket you will handle badly. And reserve the right to remove listings that misstate pricing after a change, since your own pricing will move faster than their page will.
Affiliate Program Terms: Free Template has a version of this you can edit.

A launch checklist for an AI affiliate program
Six steps, in order, before you announce anything.
- Compute margin on your heaviest realistic customer, then choose a rate that keeps that customer profitable after commission.
- Write the commission base, credit handling, and reversal window into the terms, and make the dashboard display exactly those rules.
- Connect the provider webhook and verify the paid path in test mode: trial, first invoice, renewal, refund, credit purchase.
- Publish a facts page with current pricing, limits, and model options so affiliates stop guessing.
- Set self-referral and duplicate-account rules, then test them by referring yourself with a second address.
- Build a dated share kit with real screenshots, and schedule a refresh whenever the UI or pricing changes.
If you are on Polar, the wiring is covered in How to Add an Affiliate Program to Polar.sh. Stripe and Paddle have their own guides, and the webhook semantics differ enough that copying a handler between them is how refunds get lost.
FAQ
How much should an AI product pay in affiliate commission? Start from margin, not from the market. A 20 to 30 percent recurring rate works when gross margin after inference costs stays above roughly 70 percent on a typical heavy user; below that, a lower rate, a first-year-only structure, or a flat per-conversion amount keeps the program solvent. Whatever you pick, keep it on pre-tax revenue and publish it on the program page so affiliates are not guessing from a competitor's blog post.
Do free credits or a free trial count as a conversion? Record them, do not pay them. Store the referral against the account when the free period starts, and let the first paid invoice create the conversion. That keeps your conversion rate honest, keeps the affiliate attached if the customer upgrades next quarter, and avoids paying commission on accounts that were never going to spend money.
Can affiliates promote an AI tool using AI-generated content? Usually yes for drafting, no for claims. A generated comparison table that misstates your pricing, or a synthetic screenshot of a UI you never shipped, creates a refund and a support problem. Write the rule as: generated drafts are fine, generated claims and screenshots are not, and every product claim must be checkable against your published facts page.
How do I stop affiliates from self-referring on a free tier? Match the affiliate email against the referred account email, rate limit signups per IP and device, and require a payment method before a referral counts. Require a real conversion, meaning a paid invoice, before any commission is created. Then enforce visibly: void the duplicate, merge the accounts, and state the rule in the terms before it is needed, so the conversation is about a rule and not about trust.
Building an affiliate program for an AI product on Polar, Stripe, or Paddle? Sled tracks clicks, saved referrals, and recurring commissions, blocks self-referrals, and pays out through six methods for a flat $19/mo with no cut of your commissions. Start free trial.