AI Content Licensing: How Publishers Get Paid, and How Much

AI content licensing is a publisher granting an AI company the right to use what it publishes, for money, and it comes in two kinds that pay very differently. A training licence lets the buyer copy an archive into its models, and the large ones went to the biggest names as fixed yearly fees: Meta's deal with News Corp was reported at up to $50 million a year. A grounding licence lets an AI product read, quote and link current articles in its answers, and the routes open to everyone else pay by use: TollBit's chief executive told Digiday that nearly 20% of nearly 7,000 publisher sites on its bot paywall had earned from it, at hundreds of dollars to tens of thousands a month.

This guide is for a company whose business is content it publishes and that is not one of the headline names: a news or trade publisher, a magazine, a B2B media company, a reference or how-to site, a forum or community, a newsletter with an archive, a content studio. It explains what AI companies license content for and why the two uses pay differently, what publishers of each size have been paid, the routes open to a mid-sized publisher and what each platform publishes about its rates, what blocking AI crawlers does to your position, the rights to check before you license anything, and a worked example. The news deals in detail are in our guide to AI licensing deals with news publishers, books in selling books to AI companies, and the listing sites' fees in our comparison of AI training data platforms; this guide links them rather than repeating them.

AIData.Broker also acts as the agent of holders who license or sell their data, so we are a party in this market. Every figure below links to its source for you to check. This guide is research from public sources, not legal, tax or investment advice.

Contents

1. Training or grounding: what AI companies license content for
2. What publishers have been paid, by size
3. The routes open to a publisher that is not a headline name
4. What the platforms publish about their rates
5. Blocking AI crawlers, and what it does to your hand
6. The rights to check before you license anything
7. A worked example: a mid-sized trade publisher
8. Questions people ask
9. What this means for you

Training or grounding: what AI companies license content for

An AI company can want your content for one of two jobs, and almost everything about a licence follows from which one it is. The first is training: the buyer copies your archive and uses it to train or evaluate a model. The model learns from the text once and does not need the article again, so the buyer takes a copy, usually in one delivery, sometimes refreshed. The second is grounding, also called display or retrieval: an AI product fetches your current articles when a user asks a question and quotes, summarises or links them in its answer. Engineers call this retrieval-augmented generation, or RAG: the model looks the content up rather than learning it.

The two jobs value different things. A training buyer pays for what it cannot get elsewhere, and says so: OpenAI's data partnership asked for datasets "not already easily accessible online to the public today", and Google's Content Offer Pilot pays partners "for the delivery of their offline content". Depth, editing and clean rights matter; freshness matters little. A grounding buyer pays for the opposite: today's journalism, a trusted name to cite, and the right to show it. Microsoft describes its publisher programme in those terms: publishers "will be paid on delivered value" while AI builders license content "for specific grounding scenarios".

Why the two pay differently

Because the value of a training licence is the archive itself, it is priced as a sum for a term, paid up front or by year. Reddit's January 2024 licences, whose value Reddit stated in its prospectus, carried $203.0 million of contract value over two to three years. Thomson Reuters' chief executive said Reuters' AI licences cover the archive text only, not the live news file. A grounding licence earns one answer at a time, so it is priced per use, per page fetched, or as a share of the AI product's revenue, and it pays only while the product is used.

The law adds a second reason. In the US, whether training on published text without a licence is fair use (the defence that allows some unlicensed copying, judged case by case) is not settled: The New York Times' case against Microsoft and OpenAI has reached cross-motions for summary judgment with no ruling, and the US Justice Department has urged the court in a non-binding statement to hold that such training is fair use. The Copyright Office's view is that where licensing markets exist "unlicensed uses will be disfavored", but Judge Chhabria, like Judge Alsup, refused to count a lost training licence fee as harm, writing that such a market is not one the plaintiffs are "legally entitled to monopolize". Showing your article inside an answer is a different act, closer to republishing it, and that is where buyers have been signing. Our guide to fair use and AI training sets out every ruling.

The ledger shows the result. Of the 130 disclosed licensing deals we track, 61 include training and 47 license content only to be shown in answers; the news guide charts how the share of training deals fell from 2024 onwards.

Which route pays for which

Training money comes from direct archive deals with AI companies, from the buyers that take text archives (section 3), from Google's Content Offer Pilot for offline content, from collective training licences, and from a broker that takes an archive to several buyers. Grounding money comes from per-crawl and per-use paywalls such as TollBit and Cloudflare, from answer engines that share revenue such as ProRata and Perplexity, from Microsoft's and Google's publisher programmes, and from direct display deals like Meta's and Amazon's. The headline deals of 2023 and 2024 bundled both, which is one reason their prices are so hard to compare.

What publishers have been paid, by size

Very few prices are public. Of the 130 disclosed deals, only 23 carry a dollar amount, and almost every amount for a publisher was reported by the press, citing people familiar or reading a parent company's accounts, rather than stated by the parties. A reported figure may be a ceiling, a first-year sum or a number that bundles cash with technology credits.

Size here means more than circulation. A buyer of news pays for some mix of a live feed, the right to show content under a known name, training rights to an archive, and relief from the risk of being sued, as the news guide explains; the larger the publisher, the more of these it can sell at once. A mid-sized publisher usually has the first two in modest measure and the third only if its archive is deep or offline, so the lower rows of the table, not the top ones, are its comparables.

What the table shows

The amounts fall by orders of magnitude as the publisher gets smaller, and faster than size alone explains. The eight-figure yearly sums all went to publishers that combine a large archive, a live feed, a brand an AI product wants to cite, and the means to sue: News Corp and The New York Times license to some AI companies while suing others. Specialist publishers licensing retrieval to companies report six figures, a listed specialist media group called its deal immaterial, and local publishers were paid in kind. For most publishers the realistic first income is per use, and it starts small.

Two things are missing from the record. No AI company has published a price per article, and the per-page rates on the paywalls are set by each publisher and not disclosed. Our rates page computes a median for text licences of $16M per contract year, but it rests on the large news, Reddit and academic deals, so read it as the top of the market, not a price for your archive. Academic and book publishers sit apart again: John Wiley & Sons reports $49.1 million of AI licence revenue in its year to April 2026, covered in our books guide.

Asking prices are not sale prices

A few publishers list archives at a price on data catalogues. The language-resources catalogue ELRA lists three years of the Spanish daily El Mundo for commercial use at €8,850 for 45,658 articles, and a listing for curated Indian magazine text asks $150,000 for a corpus of 60M+ words. These are asking prices; no sale at them is on record. Publishers themselves expect little: in the Reuters Institute's 2026 survey, only 20% of publishers expected licensing revenue from AI companies to be substantial.

The routes open to a publisher that is not a headline name

A publisher outside the largest few has seven routes, and most license non-exclusively, so the same archive or feed can earn through several at once. The record's summary of 2026 is that the programmes open to a small or mid-size publisher pay per page fetched, as a share of ad revenue, as a share of a retrieval licence, or by a value test, while the lump sums for archives go to the largest news groups.

The exception is a deep, specialist or offline archive, which can still be licensed for training to the buyers that take text. The table sets the routes side by side; the sections below take each in turn, and section 4 lists what each platform publishes about its rate.

Direct deals

Most direct deals went to publishers a buyer sought out, and the mid-sized and local ones often took payment in kind: OpenAI funds Axios's expansion into new local markets, gave Village Media funding, API credits and technical support, and gave Grupo Folha and Grupo UOL ChatGPT Enterprise, API and Codex access. Le Monde's deal with Perplexity pays for the use of articles plus a share of possible advertising revenue. Communities have signed too: Stack Overflow licensed its questions and answers to Google on a non-exclusive basis and to OpenAI through its OverflowAPI, with no terms disclosed.

Buyers are staffing up to do more of these deals. Mistral AI advertised for a negotiator to secure content "for AI training, RAG, and distribution", asking for a track record of closing deals of $10M+; Microsoft advertised a role to work with "providers of licensed content". The licence terms and the guide to what a data licensing agreement contains cover what to ask for, and the news guide explains how to value technology credits against cash.

Per-crawl and per-use paywalls: TollBit and Cloudflare

These products sit between your site and the bots that fetch it, identify each AI crawler, and charge it for access. TollBit is on more than 7,000 publisher websites. On its own page, TollBit says the publisher sets rates by bot, page, directory and time window, starting from standard licences for summarisation and full display, and that TollBit charges its own transaction fee to the AI customer rather than taking a share. The income is uneven: only nearly 20% of the sites had earned anything. A paywall also earns nothing until it is switched on: Digital Trends received 4.1 million bot scrapes in one week while it had not activated TollBit's charging.

Cloudflare built the same thing into the network it runs for its customers' sites. Its pay per crawl lets a site set a price per domain; an AI crawler either presents payment or receives a "Payment Required" answer with the price, and Cloudflare acts as the merchant of record, according to its documentation, which still calls it a closed beta. In September 2026 it added Pay Per Use, in which a buyer makes an offer that the publisher accepts, declines or counters, and payment follows reported use. Cloudflare bought the UK licensing platform Human Native in January 2026 and tied it to its pay per crawl work. No pricing has been disclosed for either product ( the record).

Answer engines that share revenue: ProRata, Perplexity, Taboola, Bria

A second group licenses your content to power its own AI product and pays you a share of what the product earns. ProRata runs Gist, an answer engine built only on licensed content, and pays each publisher by how much its content contributes to each answer; it has signed more than 500 licensed publications, among them Fast Company, Frommer's and Future's specialist sites, and earlier Industry Dive, Skift and Snopes. Perplexity shares advertising revenue through its Publishers' Program and set aside a pool for publishers in its Comet Plus subscription. Taboola extended its AI answer engine into an ad network for AI apps that passes most of the ad spend to publishers, and Bria licenses publishers' text for retrieval in companies' AI products and splits the revenue.

In each case the share is of the platform's own revenue, which depends on how many people use the product, so a generous percentage of a small product is a small sum. An agent's cut, by contrast, is a share of a negotiated licence price, and the two cannot be compared until you name what each is a share of.

The large platforms' publisher programmes: Microsoft, Google, Amazon, Meta

Microsoft opened its Publisher Content Marketplace (PCM) in February 2026, co-designed with "leading U.S. publishers" including the Associated Press, Condé Nast and Vox Media, and says PCM "will support publishers of all sizes". Publishers set their own pricing and usage terms and choose which AI companies they deal with; Copilot is the first buyer, and Microsoft keeps a share of each transaction that it has not disclosed. People Inc.'s chief executive described his company's part as pay-per-use, and Nine Entertainment's agreement lets Copilot reference the text of its mastheads to ground answers. PCM licenses content for grounding answers, not for training.

Google runs three separate programmes. Its News AI pilot pays news publications "for the delivery and use of enhanced content to develop new AI features", with more than 400 publications taking part by October 2026. Its AI contribution pilot, run through Search Console, pays a publisher when Google judges its content meaningfully contributed to an AI answer; Digiday quoted publishers calling the early numbers low. Its Content Offer Pilot is the one that fits an archive: it pays for offline content, the fee is agreed with each partner after an invitation, the deal is non-exclusive, and you keep your rights, and Google says it has tested it with thousands of partners.

Amazon and Meta sign directly. Amazon licensed editorial content from Condé Nast and Hearst for its shopping assistant, and says over 200 outlets power Alexa+, with no payment terms public. Meta's deals with USA Today, People Inc., CNN and others cover real-time news and links in the Meta AI chatbot, on terms not disclosed.

Collective licences

A collective licence lets many small publishers license through one body, which negotiates once and divides the money by each member's contribution. The US News/Media Alliance, whose 2,200 publisher members can opt in, has a framework with ProRata and a retrieval licence with Bria for companies' AI tools ( the record). Denmark's press collective DPCMO, which represents 99% of the Danish news industry, signed with ProRata while suing OpenAI. In Norway the government licensed national press material through the collecting society Kopinor for training and publicly releasing Norwegian language models.

Collective training licences are newer and thinner. In the UK over 250 publishers had opted in to the Publishers' Licensing Services scheme by June 2026, and in the US the Copyright Clearance Center announced a voluntary AI training licence; neither publishes a fee or a payout. Joining a collective does not hand over everything: under UK law a rights holder gives consent specifically for each right a collecting society manages and may withdraw rights on notice, so read what the membership terms cover.

Licensing an archive for training to the buyers that take text

An archive that is deep, specialist or offline can be licensed for training even by a small publisher, because the buyer pays for the content rather than the brand. The buyer board lists the buyers that publish terms; those that take text archives are below. Their fees and payout terms are compared in the platforms guide linked at the top of this one.

Two more routes carry an archive without a deal of your own making. Google's Content Offer Pilot pays for offline content, which includes back issues that never went online; our guide to selling paper archives covers digitising them. And a larger publisher that already licenses to AI companies can carry a smaller one's content for a royalty: Wiley's research business sells AI licences that include content Wiley licensed from other publishers. The text and archives page, what labs buy and our page for publishers show what each buyer asks for.

A broker or agent that puts the archive before several buyers

The last route is an agent that acts for the holder alone and puts the archive before several buyers at once, so that their offers, rather than one buyer's first number, set the price. Credtent, a licensing agent for text, audio and video catalogues, routes 85% of gross revenue from every licensing deal to the rights holder and keeps 15%, with the holder approving each deal. An agent's fee is a share of a negotiated price, which makes it comparable to the price itself; a platform's revenue share is not. The page on going direct or through us sets out what each route gains and gives up.

What the platforms publish about their rates

No platform publishes a price a publisher can expect, and the record is explicit that neither Microsoft nor Cloudflare states one. What they publish is the shape of the deal: who sets the rate, and what share of what the publisher receives.

The platforms fall into three groups by who sets the price: the publisher (TollBit, Cloudflare, PCM), the platform by a formula of attribution (ProRata, Perplexity, Taboola, Bria, Google's contribution pilot), or a negotiation with each partner (Google's News AI and Content Offer pilots, an agent such as Credtent). The group tells you what you control: the rate itself, nothing but whether to join, or the terms of one deal.

Where you set the rate, the risk is that no bot pays it, and there is no public benchmark: begin from what the bots already fetch on your site, which the paywalls measure before they charge, and adjust. Where the platform shares its revenue, the percentage is published but the base it applies to is not, so ask for the usage reports a platform sends its publishers before you judge it. A guaranteed floor, the structure Dotdash Meredith reportedly won from OpenAI ( a minimum plus a variable part), is the one protection none of the per-use platforms offers; our licence term on price structure explains how to ask for it in a direct deal.

Blocking AI crawlers, and what it does to your hand

Every AI company that trains on the web or grounds answers in it fetches pages with a named crawler, and a publisher can ask each one to stay away in its robots.txt file, the plain text file at the root of a site that tells bots what they may read. Our census reads the robots.txt of the most visited 300,000 sites in the regions we cover. In the current pass, 15.8% of the 22,561 sites that answered shut at least one AI crawler out of the whole site, and 12.6% shut out two or more, one of them a training crawler. The AI crawler census shows the daily count.

The census measures what sites ask of crawlers, not what crawlers do, and it counts a site as blocking a crawler only when the file shuts that crawler out of the whole site. The sites it reads are the most visited ones, so a smaller publisher's peers may choose differently. The table gives the share that shut out each of the main AI crawlers.

Sites shut out the training crawlers far more often than Perplexity's answer crawler, and for a publisher that pattern makes sense, because the two kinds of crawler do different things to your position.

Blocking training crawlers makes the archive scarcer

A training buyer pays for what it cannot crawl, so an archive every lab can copy for free is worth little in a training licence. Blocking the training crawlers stops future copying, which is the first step to having something to sell. It does not recall what was already taken: robots.txt restrictions grew quickly in 2023 and 2024, and a study of the C4 training set found them cutting off 28%+ of its most actively maintained sources, but a restriction applies to future crawls, not to copies already made. Text behind a paywall, in print or in a members' area was not open to crawlers in the first place, which is why it is often the most valuable part of an archive.

Blocking costs little in traffic, because AI crawlers send few readers back: Cloudflare measured that Anthropic's crawlers fetched as much as 500,000 pages for every visitor they referred at the peak in 2025. Google-Extended and Applebot-Extended are separate switches for training, so a publisher can refuse training while staying in Google's and Apple's search.

Blocking answer crawlers is a price, not a refusal

An answer engine's crawler is the route by which grounding income arrives, and blocking it outright gives up the income along with the use. The paywalls in section 3 turn the block into a price: TollBit and Cloudflare stop an identified AI bot at the door and let it in when it pays, while PCM, ProRata and Google's pilots take content through the programme. The pattern that fits most publishers is to block training crawlers, charge or license answer crawlers, and leave search crawlers alone.

What a block is worth in law

A robots.txt line is a request, not a lock. A US federal court held that robots.txt directives have no effect on a bot's ability to access the works if the bot's maker chooses not to heed them, so ignoring one is not circumvention under the Digital Millennium Copyright Act. Getting round a technical barrier is different: in Reddit's case against SerpApi, the court let the circumvention claims proceed where the scrapers allegedly evaded CAPTCHAs and other controls. A gate that checks identity, such as a paywall, a bot wall or TollBit's and Cloudflare's paid gates, therefore carries more weight than a robots.txt line.

Europe gives the request more force. The EU's text and data mining exception lets anyone copy lawfully accessible content for mining unless the rightholder has reserved its rights "in an appropriate manner, such as machine-readable means", and the AI Act requires providers of general-purpose models to identify and comply with that reservation. The EU's data protection board lists respecting robots.txt among the measures that can help justify scraping personal data. In the UK the text and data analysis exception covers only "research for a non-commercial purpose", so a commercial AI company cannot rely on it to copy a UK publisher's pages.

Not every bot obeys

Blocking reduces the crawling; it does not end it. TollBit measured that 30 percent of AI bot scrapes in the last quarter of 2025 bypassed explicit robots.txt permissions, and Digiday reported that the UK publisher Reach had seen well over 100 million successful scrapes of its content; Reach has sent OpenAI a cease-and-desist letter. News Corp is suing the search engine Brave, alleging it disguises its crawlers to evade publisher blocks, and wikiHow is suing OpenAI over content taken "directly or through intermediary datasets". The practical lesson is that text free on a site is already being fetched, so a licence sells rights and delivery, not access.

Check also what you have already allowed. Meta's updated Commercial Terms, in force from 30 October 2026, have businesses with Meta business accounts authorise Meta to crawl the domains and pages associated with their products or services; the clause does not name AI training, and its scope is for your own lawyer to read. Our guide to whether it is legal to scrape data for AI covers the scraping cases in full.

The rights to check before you license anything

A publisher can license only what it holds the rights to, and a publication is rarely all its own. Every route above puts the rights on you: Google's pilot requires partners to have "the necessary rights" to the content, and a buyer's lawyers will ask that third-party content is identified and removed or licensed.

The questions come from the way publications are made: many hands, many contracts, and material brought in from outside. Sort the archive by who created each piece before you sign anything; the rules cited are US law, and other countries ask the same questions with their own answers. The table at the end of the section sums them up.

Staff work

Articles your employees wrote as part of their jobs are works made for hire, and for those the employer "owns all of the rights comprised in the copyright" unless a signed writing says otherwise. Check that no employment contract or newsroom agreement reserves rights to the writer.

Freelancers and outside contributors

This is often where an archive is weakest. Copyright in a contribution to a magazine, newspaper or site vests initially in its author, and without an express transfer the publisher of a collective work is presumed to hold only the privilege of reproducing and distributing the contribution as part of that publication, a revision of it, or a later issue in the series. When newspapers and magazines licensed freelancers' articles into electronic databases, where each appeared on its own, the Supreme Court held in New York Times Co. v. Tasini that the privilege did not cover it and that the publishers had infringed the freelancers' copyrights. Tasini concerned databases, not AI, but it tells you what to check: whether each freelance contract expressly granted rights beyond the publication itself.

A commissioned contribution to a collective work counts as a work made for hire only "if the parties expressly agree in a written instrument signed by them", and any transfer of copyright must be in writing and signed. So the test is the paper: contributors whose contracts assign copyright, or grant licensing rights broad enough to cover AI uses, can go into a licence; those with older or narrower contracts need a new grant or should be left out. Where your contracts changed at some point, the date of each contract decides which side of the line a contributor's work falls.

Syndicated and wire copy

Stories you ran from a news agency, a syndication partner or a content exchange were licensed to you for publication under that provider's terms, and are not yours to license onward unless those terms say so. The agencies license their own archives directly, as Reuters does with its archive text, so a buyer has no reason to pay you for the same copy. Strip wire and syndicated pieces from the archive, together with press releases republished in full.

Images, video and audio

Photographs come from staff photographers (the company's), freelance photographers (theirs unless assigned, as with text) and agencies. Getty Images, Shutterstock, Adobe, Alamy and Freepik all forbid customers to use licensed images for machine learning or AI unless separately authorised, and Getty does not warrant that people in its photos consented to AI uses, so a text licence should leave agency images out. Video and podcast archives are licensed separately; see our guides to selling videos and audio and speech data.

User comments and forum posts

Posts on a forum and comments under articles belong to the people who wrote them, and the platform can license them only as far as its terms of service grant it a licence. Reddit's user agreement shows what an explicit grant looks like: its licence from users "includes the right to use Your Content to train AI and machine learning models". If your terms grant only what you need to display posts on your own site, a licence to an AI company may need new terms, and changing terms for content already posted is a question for counsel. User content also carries personal data, and regulators look at its licensing: the FTC opened a non-public inquiry into Reddit's licensing of user posts, and the Dutch data protection authority ordered access to information about it. Our Reddit guide and the guide to GDPR and licensing data for AI training cover both.

Earlier licences and exclusivity

List every licence the content is already in: an aggregator or database deal, a syndication contract, a platform programme, an earlier AI deal. An exclusive grant anywhere can block every other route, while most of the programmes above are non-exclusive. The guide to exclusive and non-exclusive licences explains what exclusivity is worth, and the guide to training data provenance why buyers pay for a record of where each piece came from.

A worked example: a mid-sized trade publisher

The following is an example, not a real company. A B2B trade publisher covers one industry. It runs a news site with a metered paywall, a daily newsletter, an archive of articles going back many years, long-form buyer's guides and market reports sold to subscribers, a members' forum, and an annual print directory whose older editions were never online. Its writing comes from staff and from freelancers, whose contracts changed some years ago from first publication rights to a full assignment. No AI company has approached it, and it wants to know what it can earn.

Step 1: split the holdings by use. The daily news and newsletter are a grounding product: current, citable, valuable for a few days. The long-form guides, market reports and old print directories are a training product: deep, specialist and, being behind a paywall or in print, not on the open web. The forum is user content and is set aside until its rights are settled.

Step 2: sort the rights. Staff pieces go in. Freelance pieces under the newer contracts go in; those under the older contracts stay out unless the writers sign a new grant, which the publisher offers with a share of any licence income. Wire copy and agency photos come out. The forum's terms grant the publisher a licence to display posts on its own site and nothing more, so the forum stays out of any training licence.

Step 3: set the crawler policy. The publisher blocks the training crawlers (GPTBot, ClaudeBot, CCBot, Google-Extended, Applebot-Extended and the rest) to keep its paywalled and long-form content scarce, keeps the search crawlers open, and puts the answer crawlers behind a per-use gate through TollBit or Cloudflare rather than blocking them.

Step 4: turn on grounding income. It applies to Microsoft's PCM, joins ProRata through its trade association's framework if it is a News/Media Alliance member, and takes part in Google's programmes if invited. All are non-exclusive, so it can do all of them. It expects the income to be modest: the record says only that the paywall sites that earn at all earn hundreds of dollars to tens of thousands a month.

Step 5: offer the deep archive for training. The long-form guides, reports and old directories are what a training buyer pays for. They fit what Mercor's listing for publishers asks for, "long-form content you hold the rights to", and the print directories fit Google's Content Offer Pilot for offline content. The short daily news would not: Mercor's listing says "not short articles alone". The publisher describes the cleared archive (titles, word counts, years, subjects, rights status) with the datasheet builder, runs the rights check, and gets a valuation before any buyer names a figure.

Step 6: decide how to sell it. It can approach buyers one by one, list the archive on a listing site at a price it sets, or have an agent put it before every buyer that fits at once. It keeps any licence non-exclusive and limited to training, with a defined term for the copy and limits on reproducing text in outputs, so the grounding income in step 4 is untouched. Whatever offer arrives, the offer check reads it against market terms.

The result is two income lines that do not compete: small, recurring grounding income from the live product, and a one-off or periodic training licence for the deep archive. The rights work in step 2 is what makes the second possible, so it is the step to start with.

Questions people ask

What is AI content licensing?

A publisher granting an AI company the right to use its content, for training models or for showing it in AI answers, in return for payment. Training licences are usually fixed fees for an archive; grounding licences pay per use or as a share of revenue. Of the 130 disclosed deals we track, 47 are for display only.

How much are AI content licensing deals worth?

Almost no publisher deal has a price stated by the parties. Reported figures run from close to $250 million over five years for News Corp with OpenAI, through about $3 million a year for the Indian Express Group, to "six-figure deals" for retrieval licences through Snowflake. Among the few stated in a filing, Reddit's licences carried $203.0 million over two to three years. Every deal is in our deal ledger.

How does licensing content for AI training work?

You license a copy of an archive, usually non-exclusively, for a buyer to train or evaluate its models, for a fixed fee and a defined term for the copy; the models it trains usually survive the term. Buyers pay for depth and for what is not on the open web, and will ask you to show you hold the rights to every piece. The guide to data licensing agreements covers the contract, and licence term length what survives the end.

Should your company license content to AI?

If AI bots already fetch your pages, a per-use gate earns from that traffic at little cost, and most grounding programmes are non-exclusive. A training licence is worth pursuing when your archive is deep, specialist or offline and you can show clean rights. What the record argues against is a low first deal for everything: a low first price for a whole repertoire becomes the price every later buyer cites, which is why Spain's book collecting society refused one.

Which content licensing companies work with AI?

Four kinds: AI companies that sign direct deals (OpenAI, Microsoft, Google, Amazon, Meta, Perplexity, Mistral), per-use paywalls (TollBit, Cloudflare), answer engines that share revenue (ProRata, Perplexity, Taboola, Bria), and buyers and agents that license archives for training ( Mercor, Protege, Defined.ai, Credtent and others). The buyer board and the register of companies list each with its terms.

What AI content licensing solutions are there for publishers?

The tools that need no negotiation are the per-crawl gates (TollBit, Cloudflare's pay per crawl and Pay Per Use), Microsoft's PCM, ProRata's Gist programme and, by invitation, Google's pilots. TollBit's publishers keep 100% of the rate they set, and ProRata's framework with the News/Media Alliance pays publishers 50% of all Gist.AI revenues. None publishes what a publisher can expect to earn.

How big is the AI content licensing market?

No reliable total exists. What can be counted: 130 disclosed deals, of which 23 carry an amount, and a few revenue lines companies report themselves, such as Reddit's other revenue, mostly content licensing, of $140.0 million in 2025. Our guide to the size of the AI training data market explains why the published estimates cannot be checked.

Does blocking AI crawlers stop my content being used for AI?

It stops the crawlers that obey it from copying new pages; it does not recall copies already made, and not every bot obeys. In the US a robots.txt line is a request, not an access control; in the EU a machine-readable reservation takes your content out of the text and data mining exception. Block training crawlers, and charge answer crawlers rather than blocking them.

What this means for you

A mid-sized publisher will not be offered the sums in the headlines, but it has more routes than it did two years ago, and most of them run side by side. The money divides cleanly: the live product earns per use, and the deep archive earns a training licence, if its rights are clean.

The order matters. Rights work and a crawler policy cost little and come first, because every route depends on them; the grounding programmes can then run in the background while you decide whether the archive is worth a training licence, and on what terms.

1. Separate grounding from training. Price the live feed and the archive separately, and never let one buyer bundle them into one number.
2. Set the crawler policy deliberately. Block training crawlers, gate answer crawlers through a per-use paywall, keep search open, and see what others do on the AI crawler census.
3. Do the rights work first. Sort staff, freelance, wire, image and user content; get new grants where contracts are old; record what is cleared. The rights page and the rights check list what a buyer's lawyers ask.
4. Join the non-exclusive programmes. PCM, ProRata, TollBit or Cloudflare, your trade association's collective deal, and Google's pilots if invited. Track what each actually pays before you renew anything.
5. Value the deep archive before anyone quotes. The valuation gives a range from public prices, the rates and price board show the evidence, buyer match finds the buyers whose rules fit, and the guide to what buyers check before they buy shows what to prepare.
6. Keep licences non-exclusive and bounded. A defined term for the copy, limits on reproducing your text in outputs, a floor where you can get one, and no exclusivity without a premium and an end date.

AIData.Broker values an archive against the public record and, for a holder based in the United States, the United Kingdom and Canada, describes it without the holder's name to every buyer that fits at the same time; the holder licenses directly to the buyer it chooses and pays our fee only out of a sale. For the rest of the market, see how AI companies get training data and AI training data prices; every figure here is on the records page, and how we source every number explains the method.