Cloudflare Pay Per Use: What Is Operational and What Is Still an Experiment
Cloudflare is testing usage-linked publisher payments with partners. Separate available controls, research, partner experiments, and future plans before forecasting revenue.
Published August 10, 2026: Cloudflare is experimenting with payment models that connect compensation to the use of publisher content in AI search. It calls the direction “Pay Per Use,” but also says plainly that this is an experiment. Publishers should not treat it as a generally available revenue product or forecast income from it.
The useful distinction is between shipped controls, a research project, partner experiments, and planned availability. Cloudflare’s announcement supports an evaluation framework today; it does not yet support a universal implementation guide, payout estimate, or claim that usage-based compensation works at Internet scale.
Four different product states
| Initiative | Evidence state | Safe conclusion |
|---|---|---|
| AI crawler controls | Available Cloudflare controls, with plan and product differences | Publishers can make access decisions and inspect traffic |
| Smarter search signaling | Research program with broader availability planned | Cloudflare is testing whether signals can reduce repeat crawling and improve retrieval |
| Pay Per Crawl | Previously launched publisher-control model | Payment is associated with crawl access, not proven downstream use |
| Pay Per Use | Partner experiments with Ceramic.ai and You.com | Usage-linked payment models are being tested; the commercial outcome is unresolved |
Cloudflare says its own data suggests that more than half of good-bot crawl traffic refetches pages that have not changed. That is a vendor-owned network observation, not a universal measurement of every publisher or crawler. The announcement does not provide a public sample definition that would let an outside analyst reproduce the percentage.
What the partner experiments test
Cloudflare describes Ceramic.ai’s model as pay per query, with payment designed to occur when opted-in publisher content appears in search results. It describes You.com as testing on-demand payment for a specific piece of premium content. These are different economic events and should not be combined into one “AI licensing rate.”
A publisher evaluating participation needs a ledger that identifies the answer engine, query or request event, content URL, excerpt or asset used, counting rule, invalid-traffic policy, price, currency, minimum payout, reporting delay, dispute path, and whether the user later visited or converted. Without those fields, a usage total cannot be reconciled.
Cloudflare also describes possible reporting such as queries, pages, snippets, and search-result positions for participants in the Ceramic program. Treat these as announced partner-program fields until the publisher can inspect the actual report and its definitions.
Questions before a publisher opts in
- What counts as use? A fetch, result appearance, excerpt, citation, answer inclusion, tool call, or completed user outcome are not equivalent.
- Who records the event? Require the operator, timestamp, request identifier, and a reconciliation method.
- What content is licensed? Define URL patterns, media, excerpts, premium material, territories, and update behavior.
- What remains blocked? Search retrieval, agent action, and foundation-model training require separate decisions.
- How is value calculated? Preserve price, denominator, refund or fraud rules, currency, and payment threshold.
- What can be audited? A publisher should be able to connect a report to server access, content version, and contract terms without exposing user data.
- How can participation end? Record revocation, cache or retention behavior, outstanding payment, and the expected delay before access changes.
Use the publisher-control outcome matrix to keep access, training, search retrieval, citation, and referral separate. Use the AI visibility measurement crosswalk before translating a new report into an editorial or commercial claim.
A minimum pilot scorecard
- Content scope and rights were reviewed before opt-in.
- Search use and model training are contractually and technically separated.
- Every paid event has an inspectable definition and denominator.
- Reports can be reconciled with dated publisher evidence.
- Revenue is reported after fees, disputes, reversals, and payout thresholds.
- Crawl load, discoverability, referral traffic, and content freshness are measured separately.
- The publisher has a revocation and rollback procedure.
The decision rule should be declared before the pilot. For example: continue only when auditable net revenue exceeds the operational and rights-management cost without increasing unwanted access. A promising direction is not yet a durable business model.
Ask a question or join the discussion