Reward sharing without losing sight of usage costs.

An AI referral incentive can create demand and a compute bill at the same time. Connect the reward to paid activation, price the promised usage, and decide how your application will deliver it.

Discuss your setup

RefRef · Program design notes · Updated October 3, 2026

01Cost of the offerRefRef

A credit is a unit you define.

A credit might buy one image, a document summary, or part of an agent run. Its displayed value and your cost to serve it can be very different. Estimate model usage, retries, tools, storage, and support before choosing the number on the invitation.

Model pricing can separate input, output, caching, and tool charges. See Anthropic’s pricing documentation for one provider’s structure. Use your own observed workload mix; the example below does not assume any provider’s current rate.

Bounded product usage

Offer a defined amount of work, such as 100 standard generations. Your application must specify eligible models, enforce consumption limits, and explain any expiry to the customer.

Monetary account credit

Offer a fixed monetary amount against future charges. Budget the lost revenue as well as usage costs. A dollar of invoice credit is not the same as a dollar of internal compute budget.

02Example programRefRef

Give both sides 100 standard generations.

Illustrative design: an existing user refers a new user. After the new user’s first eligible $30 subscription purchase, each receives 100 additional standard generations. The bonus is delivered once, with its eligible model and consumption limit enforced by your application.

For RefRef, use a purchase-based referral Program to connect the two participants and record custom Rewards of 100 generations for each participant. Define a custom unit for generations, then use reward webhooks or the reward event feed to connect earning to your application. Your application grants the generations once per Reward and enforces the eligible model, usage balance, and expiry. RefRef records fulfillment; it does not meter or consume the generations.

A first successful generation is useful activation data. Your app can track it, but do not assume a custom usage Event establishes a referral: only the built-in signup and purchase Events bind or inherit one. Keep purchase eligibility distinct from your product’s usage limits.

03Reward economicsRefRef

Price all 200 generations, not just the new user’s half.

Illustrative USD model for one paid referral in its first month. Full bonus consumption is assumed; no customer performance claim.
Input or resultIllustrative amount
Collected subscription revenue$30
Normal included usage and variable service costs$12
Contribution before the referral bonus$30 − $12 = $18
Bonus usage across both participants100 + 100 = 200 generations
Base incremental cost per generation$0.02
Base bonus cost200 × $0.02 = $4
Contribution after bonus, base case$18 − $4 = $14
Stress cost per generation$0.08
Contribution after bonus, stress case$18 − $16 = $2

If you want at least $10 of first-month contribution from this example, the bonus budget is $8. At $0.08 per generation, that buys 100 generations in total, or 50 for each side. At $0.02, it buys 400 total. The affordable offer depends on the cost you can actually control.

These calculations assume bonus usage is incremental and does not replace usage customers would otherwise buy. Subtract displaced revenue, additional acquisition costs, and any omitted payment or support costs in your own model. Fixed overhead and tax are outside this example.

04Setup responsibilitiesRefRef

Your product owns the usage ledger.

  1. Establish the referral. Enroll the referrer, share their link, capture the signed handoff, and send trusted signup and purchase facts from your backend.
  2. Evaluate the reward. Configure eligibility and benefit limits. Reconcile reward records with refunds and adjustments before delivery.
  3. Grant the entitlement. Handle the reward webhook or event feed in your backend, grant each bonus once per Reward ID, and confirm fulfillment. For a smaller pilot, an operator can grant the bonus and record fulfillment in the Console.
  4. Meter actual work. Enforce model access, usage allowance, and any expiry in your application. Compare actual incremental cost with the budget before widening the cohort.

RefRef’s accounting record is not proof that the bonus reached the user. Its sharing Widget does not display a usage balance or Reward history. Your app provides those views and decides how a later refund affects remaining bonus usage.

05Run a small pilotRefRef

Use the reward budget calculator to stress-test contribution after usage costs. Then use the double-sided program template to record what each participant receives.

Keep the first experiment measurable.

For a bootstrapped AI product, begin with a small cohort and a fixed bonus for one bounded workload. For a developer API, a raw request count may hide large token and tool-cost differences; define a metered unit you can enforce. For a subscription, specify whether future purchases earn anything before promising recurring benefits.

Measure paid activation, retained subscribers, bonus consumption, and contribution after incentives. A rise in free accounts alone does not establish that the program acquired valuable customers. Account creation controls and abuse review belong in your application and operating process; referral attribution is not a complete fraud system.

Selling to teams? Combine this cost model with the B2B SaaS account referral design so a person joining a team does not become a second acquired account.

Bring your program brief.

Tell us who refers, what qualifies, and how you will deliver the benefit. RefRef’s external onboarding is not yet generally available; discuss your integration and rollout with the team.

Discuss your setup