Content monetization in the AI era

How a web resource earns when the click no longer arrives. Every revenue model the open web was built on — display ads, affiliate links, lead capture — is paid per human visit, and the visit is what the ai-search-shift removes: AI summaries halve link clicks (pew-ai-overviews-clicks), human traffic to the most-crawled categories fell up to 40% in a year, and AI crawlers now take thousands of pages per referral (crawl-to-refer-ratio). The content is still consumed. The meter stopped.

Five answers exist. They are not variants of one strategy — they differ in what event gets billed, and each one works at a different scale. The first four bill the AI company; the fifth (per-copy-licensing, added 2026-08-29) bills the reader.

1. Per visit — ads, affiliate, leads

The incumbent model, and the one being measured as it declines. Still the only model that pays a small site anything today, because it needs no counterparty agreement: seo-affiliate-alignment and b2b-ppc-metrics are this spoke's coverage of it. Its problem is structural rather than cyclical — the referral is gone, not merely cheaper — which is the argument great-content-no-longer-works makes for exiting the channel and seo-no-longer-drives-growth softens into "keep the non-commoditizable work."

2. Per fetch — the crawl toll

Charge the bot for the page. pay-per-crawl via HTTP 402, floor price $0.001 per crawl, enforced by a CDN that will block non-paying crawlers by default. Revenue scales with how hard you are already crawled, which pays large publishers and rounds to zero for everyone else — and it prices the fetch, not the thousand answers generated from it.

3. Per use — inference, training, and the licensing economy

Bill the use rather than the fetch: RSL can declare pay-per-inference and pay-per-training terms in robots.txt, Cloudflare is moving from pay-per-crawl toward pay per use, and x402 with the Monetization Gateway supplies the rails for sub-cent per-call billing. The economics are right and the enforcement is unbuilt: declaring terms is unilateral, collecting on them is not. The entire licensing market so far is about 50 bilateral deals since 2023 (cloudflare-agentic-internet-report), each negotiated by hand between large parties. The collective exists because one small site has no standing to negotiate; whether collective bargaining reaches independent publishers is the open question.

4. Per access — sell the feed, not the page

Give the content away and charge for serviced delivery: bulk, real-time, structured, with an SLA. wikimedia-enterprise is the working example — $8.3M in FY2024–25, up 148%, 13 paying customers, first profitable year after four. It is the only model here with audited revenue behind it, and the one that most clearly sells a different artifact than the website: a machine-readable feed rather than pages.

5. Per copy — the reader's purchase is the licence

Bill nobody new. Google's Expert Intelligence (2026-08-27) lets a Play Books purchase authorize Gemini Notebook to ground answers in that book, with a collaborator on a shared notebook prompted to buy their own copy or lose access. 100,000+ books at launch from Bloomsbury, De Gruyter Brill, Johns Hopkins University Press, Macmillan, O'Reilly Media and Penguin Random House. The publisher is paid by the copy, as always; the AI access rides on the sale. See per-copy-licensing for the shape and its limits — chiefly that it works only for content already sold as discrete copies, so it says nothing about a web page, and that the storefront setting eligibility is the AI company itself.

What this means at small scale

The corpus supports one conclusion firmly and one only tentatively.

Firmly: the four models pay in proportion to leverage, and leverage means being crawled a lot or being hard to replace. The floor price of a crawl ($1 per thousand fetches), the size of the licensing market (~50 deals), and Wikimedia's customer count (13) all say the same thing — these mechanisms were built by and for large publishers, and an independent site adopting them collects approximately nothing in year one.

Tentatively, and marked as inference rather than sourced fact: the transferable part of Wikimedia's result is the product shape, not the price. What a machine buyer pays for is a clean structured corpus with guarantees, which is a thing a small well-maintained wiki can produce and a sprawling content site cannot. The three cheap moves that follow — publish RSL terms because declaring costs nothing and creates the claim, keep llms-txt and a feed so the content is usable when someone does want it, and use pay-per-crawl as a cost control against re-fetch waste rather than as revenue — are consistent with the evidence but not demonstrated by it. No source in this corpus reports a small publisher's earnings from any of them.

That absence is itself the finding. Nobody has published what an independent site makes from AI licensing, and every number available comes from a vendor selling the mechanism or a publisher large enough to negotiate.

The fifth model does not fix that, and is worth reading carefully on the point. Per-copy licensing needs no bilateral negotiation, which is the barrier that keeps the other four away from small publishers — but Google's post states no publisher terms at all beyond the ordinary book sale. So the mechanism removes the negotiation and leaves the earnings question exactly where it was.

pay-per-crawl · x402 · rsl-standard · wikimedia-enterprise · crawl-to-refer-ratio · cloudflare-agentic-internet-report · cloudflare-monetization-gateway · ai-search-shift · pew-ai-overviews-clicks · great-content-no-longer-works · seo-no-longer-drives-growth · llm-wiki · per-copy-licensing · gemini-notebook-expert-intelligence