AI payment models are finally getting a clear map. On November 2, 2026, Prebid's LLM Taskforce published a field guide that follows the money through the main ways publishers are currently paid when AI systems use their pages, from Cloudflare charges that start at $0.001 per page to direct deals worth millions a year. If you run an ad supported site, the guide is a useful way to compare what each model actually pays, who sits in the middle and where the catch is.
This article summarizes the guide, written by Anton Nashkerskyi of Sigma Software and Krish Arvapally of Dappier, and checks key details against Cloudflare and TollBit documentation. Deal values are the figures the guide reports from published coverage.
Table of Contents
What Are AI Payment Models?
AI payment models are the arrangements that decide who pays a publisher when an AI company crawls, summarizes, displays or trains on its content, and on what basis. The Prebid field guide groups them into four models: pay per access, pay per use, revenue share and direct deals. It also covers a fifth route outside those four, on-site answer engines funded by advertisers.
Our earlier guide to AI content licensing explains why licensing matters. This piece focuses on the mechanics: who bills whom, what a publisher keeps and what limits each model has today.
Pay Per Access
Under pay per access, the AI company pays for each page its crawler fetches. The guide's main example is Cloudflare's Pay Per Crawl. According to Cloudflare's documentation, the feature is in closed beta and Cloudflare is the merchant of record. The guide says the publisher sets a price of at least $0.001 and can vary it by content, while Cloudflare bills the AI company, pays the publisher and does not publish its share. AWS offers a similar per request charge in CloudFront, settled in stablecoins.
The appeal is simplicity. For a site already on Cloudflare, the guide notes it is a setting rather than an integration, and each charge matches a request the publisher can see in its own logs. If you are deciding which bots to allow in the first place, our explainer on AI crawlers and robots.txt covers the controls.
The weakness is that the price is per fetch, whatever the AI company then does with the page. As the guide puts it, "A crawler that caches a page pays once, however many answers it later serves from that copy." Cloudflare has also not named the crawlers that pay through it.
Pay Per Use
Pay per use charges for a permitted use, such as summarizing or displaying a page, rather than for the fetch. A marketplace verifies the buyer, bills it and pays the publisher. On TollBit, the publisher sets a rate per 1,000 pages for each license type. TollBit's rate documentation suggests publishers consider their current RPM as a benchmark for the summarization rate. The guide says TollBit's examples run $1 to $5 per 1,000 pages, and that TollBit takes no share of the rate and charges the buyer a fee on top.
Other options are emerging. Cloudflare's Pay Per Use has been in beta since September 30, with the buyer proposing a price for each use and reporting its use through an API. Microsoft's Publisher Content Marketplace is a pilot where publishers set licensing and usage terms for grounding uses. Dappier's RAG Marketplace lets publishers set a price per query, and FORA is an open protocol for the same kind of transaction.
The advantage is that the price can follow the use, so a full display of an article can cost more than a summary. The limit, according to the guide, is that "the demand side of the marketplace is not quite built out," largely because scraping tools offer alternative access. Early buyers named include Ceramic.ai and You.com for Cloudflare, Yahoo for Microsoft and the news app Particle for TollBit in 2024.
Revenue Share
In a revenue share, an AI product shares part of its own revenue with publishers and allocates it by its own measure of each publisher's contribution. The guide gives three examples. ProRata shares 50% of its revenue with content partners. Perplexity's Comet Plus, a $5 monthly subscription, pays publishers according to human visits, search citations and agent actions, and Axios reported that publishers receive 80% while Perplexity keeps 20% for computing costs, with an initial pool of $42.5 million. Google runs an invite-only pilot that pays sites whose content it judges to have contributed significantly to an AI answer.
The catch is transparency. Perplexity has not said how it weights its three factors, Google does not publish its calculation, and pooled payouts rise and fall with the operator's own revenue. The guide also points to a lack of uniformity in business models, which leaves publishers managing fragmented relationships.
Direct Deals
A direct deal is a contract between the AI company and the publisher with no intermediary, so the publisher keeps the full fee and negotiates covered uses and payment terms itself. The guide cites several reported examples. News Corp's deal with OpenAI is reportedly worth more than $250 million over five years. Dotdash Meredith, now People Inc., receives at least $16 million a year in fixed payments from OpenAI, plus a variable component. Meta will reportedly pay News Corp up to $50 million a year for three years, and Amazon reportedly pays The New York Times at least $20 million a year.
Those numbers are not a benchmark for most sites. In the guide's words, "deals take months to negotiate and have so far gone mainly to the largest publishers, leaving most of the long tail outside them." For smaller publishers, the marketplace and on-site models may be more practical places to start.
On-Site AI Answer Engines
Some publishers also earn from AI on their own pages. An on-site answer engine lets readers ask follow-up questions about an article, and advertisers pay rather than AI companies. The publisher shares that revenue with the provider or sells the space itself. The guide lists Taboola's DeeperDive, ProRata's Gist Answers, Direqt, Inline AI and Dappier's AskAI as examples.
Dappier shared how its version works. Ads in the conversation include programmatic ads, product listings and sponsored conversations, and Dappier typically shares revenue 50/50 with the publisher, with payouts based on earned advertising revenue, similar to a traditional ad network. Dappier covers LLM, token and hosting costs. Because an engine that draws on other publishers' articles has to license them, these products also add buyers for pay per use. For a related edge network option, see our look at the Cloudflare Monetization Gateway.
How the 4 Models Compare
| Model | Who pays and on what basis | Example | Main limit |
|---|---|---|---|
| Pay per access | AI company pays per page fetched | Cloudflare Pay Per Crawl, from $0.001 per page | A cached page pays once |
| Pay per use | Buyer pays per permitted use via a marketplace | TollBit, rate per 1,000 pages | Demand side not built out |
| Revenue share | AI product shares its own revenue | ProRata 50%, Comet Plus | Opaque allocation |
| Direct deal | AI company pays under a contract | People Inc. and OpenAI | Mostly the largest publishers |
5 Smart First Steps
The guide closes with five practical steps.
1. Know your volume
Count a month of declared AI crawler fetches in your logs so you can test any quoted rate against your own traffic.
2. Assess your audience opportunity
For an on-site answer engine, weigh reader engagement, ad revenue and any licensing cost together.
3. Ask who is buying
For any network or marketplace, ask which AI companies pay through it today.
4. Compare what you keep
Check whether a fee comes out of your rate or is added for the buyer, and how a revenue share calculates your part.
5. Check the terms
Confirm which uses are covered, for how long and whether it is exclusive, and agree how payments are booked, the minimum payout and any tax withholding.
No single AI payment model wins for every publisher yet. Pay per access is easy to switch on, pay per use prices the actual use, revenue shares depend on someone else's formula and direct deals mostly reach the biggest names. Start by measuring your crawler traffic, benchmark any offer against your RPM, and keep testing as the marketplaces mature.