AI publisher licensing deals: 10.2 ChatGPT citations per page against 6.9

Do AI publisher licensing deals increase citations?

The study is a join between two datasets that had not been put side by side before, and the join is the interesting part of the design.

OtterlyAI supplied the citation side: 129.3 million citations across more than 20 million cited URLs, captured on ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot, Gemini and Claude during June 2026. Press Ranger supplied the deal side through what the announcement calls its AI-Publisher Licensing Research, a database of 91 confirmed agreements between AI companies and publishers mapped to 314 publisher domains, with every record stated to be backed by a public source through 28 July 2026. The announcement describes the result as the first analysis to pair large-scale AI citation monitoring with a verified record of licensing deals, which is a claim about novelty rather than about method, and one worth separating from the findings.

The unit of measurement deserves attention before the numbers do, because it is not the unit most people assume. The comparison is citations per cited page, not citations per publisher and not citations per published page. A page enters the denominator only once it has been cited at least once, so the figure describes how often an already-cited page is cited again rather than how likely any given page is to be cited at all. That distinction removes the largest confound available to a study of this shape, which is that big publishers publish more, and it introduces a subtler one, which is that the pages being compared have already cleared the bar this post is about.

On that unit, ChatGPT showed 10.2 citations per cited page for the OpenAI-licensed cohort against 6.9 for unlicensed publishers, a 48 percent premium. Across all seven platforms combined the same cohort averaged 10.7 against 7.3, a 46 percent premium. Those two figures are close enough that the cross-platform number is largely the ChatGPT number carried by weight of volume, which is exactly what the next section is about.

Anyone reading this as a scoreboard should hold one thing in view. A citation count is a measure of appearance, and appearing is not the same as being used: a controlled study covered here found that more citations did not mean more of the page appeared in the answer. It is also not evenly comparable across engines, since 3 of the 8 engines in this repository's registry cannot search the web at all and answer from training data instead. The platform notes on getting cited in ChatGPT start one step earlier still, at whether retrieval happens for the prompt, and AI visibility as this site uses the term is an outcome to be influenced rather than a quantity a deal buys.

The advantage showed up on ChatGPT and on no other platform

The finding that makes the study worth reporting is not the headline premium. It is that the premium did not generalise.

OpenAI was the only licensor whose deals showed a clear advantage on its own platform. The announcement states that Google-licensed publishers were cited on Google AI Overviews at a slightly lower rate than comparable unlicensed publishers, and that Perplexity's licensed publishers landed at parity on Perplexity. Neither of those two results is given a figure, which is a real limit on how far they can be quoted: slightly lower and at parity are the strongest formulations available from the document, and this post will not turn either into a number. Isolating publishers that signed with OpenAI and nobody else sharpens the OpenAI result rather than softening it, at 112 percent more citations per page on ChatGPT than unlicensed publishers.

The distributional finding underneath that is the one a publisher should actually plan against. Publishers who signed with OpenAI now draw 57.9 percent of their entire AI citation volume from ChatGPT alone, while unlicensed publishers keep a more balanced mix across ChatGPT and Perplexity. Read together with the parity and slightly-lower results, that describes a concentration effect rather than a visibility effect: the same body of work becomes more visible in one place and no more visible anywhere else, so the portfolio narrows even as one line in it rises. Thomas Peham, CEO of OtterlyAI, is quoted in the announcement putting it in one sentence, that a licensing deal tilts your citations toward ChatGPT, does not guarantee you show up everywhere, and is not the only way to win.

That shape is consistent with something already published here from an entirely different dataset, which is that platform behaviour does not move together: Meta took the crawl volume while ChatGPT kept 80 to 88 percent of the referrals. It is also a reminder of how thin the downstream layer is, since citations reached 6.8 percent of ChatGPT prompts and the visit landed on the homepage in the measurement covered here in August. If a deal moves where citations concentrate rather than how many exist, then the practical notes for Google AI Overviews and for Perplexity are unaffected by any of this, and so is most of what generative engine optimization means for a site that will never sign anything.

What the announcement does not state, and what that costs the claim

The document this post reads is the study announcement. There is a fuller write-up published as an interactive dashboard, which this post did not open, and the gaps below are gaps in the announcement rather than proof that the work behind it has them.

Four things are absent, and each one narrows what the 48 percent can carry. The announcement does not describe how licensed and unlicensed publishers were matched: the phrase comparable unlicensed publishers appears in the Google result without a stated comparison procedure anywhere. It gives no citation volumes from before the deals were signed, so a single month cannot separate an effect of signing from a difference that already existed. It reports no significance test and no interval around any figure. And it does not state the size of the unlicensed comparison set, while giving 314 domains for the licensed side.

The selection problem those gaps leave open is the obvious one and it runs in the direction that flatters the finding. OpenAI has signed large, established, English-language news organisations. Those publishers were plausibly more cited before any deal existed, for reasons a deal did not create, and a design that compares cohorts in one month cannot tell the two stories apart. The announcement's own subheading is careful about this, saying the deals track with a citation advantage, which is a correlational verb. The quoted commentary in the same document is not as careful, describing what a deal does. Both sentences sit in one press release, and the difference between them is the whole question.

Two further things are worth naming plainly. The study was published by two companies with a commercial interest in its conclusion, one selling press release distribution and one selling AI search monitoring, and its practical recommendation is to pitch trade publications and monitor the results. That does not make it wrong, and the underlying citation panel is the largest thing of its kind this site has seen reported, but it is a fact a reader should hold. It is also one month of data on a system that moves: a separate study found that AI visibility took seven runs per prompt to settle before a figure stopped drifting, and another found that 16 percent of the sources four AI search engines cited were AI-generated, which is a reminder of how loosely filtered these citation pools are. Reading a vendor study for its evidence rather than its conclusion is the same discipline as reading a category page for its criteria, which is how the two machines sold as AI visibility tools were separated here on the same day.

News was 7.2 percent of citations, and unlicensed trade outlets carried most of it

The part of the study least likely to be quoted is the part with the widest application, because it is about everybody who is not a national newspaper.

News is a small share of what these systems cite at all: 7.2 percent of the 129.3 million citations in the dataset. Within that small share the concentration is severe, with five media groups, named in the announcement as Future plc, Forbes, People Inc., Conde Nast and Hearst, capturing 69 percent of all citations to licensed publishers. So the licensing premium described above is a premium inside a thin slice, distributed unevenly within it, and a publisher outside those five groups gains little by knowing the aggregate.

Then the finding that cuts the other way. Across 16 US industries, niche and trade outlets, most of them unlicensed, carried the majority of news citations in 15 of them, and the announcement reports those outlets collecting 213 percent more AI citations than mainstream media. The most-cited news domains in the data are given as including unlicensed outlets such as NerdWallet, Healthline and Bankrate. Alongside that, 46.9 percent of licensed publishers' citations went to best-of lists, buying guides and product reviews, which says the content type these systems reach for is commercial and evergreen rather than breaking.

Put those three together and the practical reading inverts the headline. The premium is real on one platform and belongs to a set of publishers almost nobody reading this belongs to, while the volume sits with trade and niche outlets that signed nothing. Steve Beyatte, Founder of Press Ranger, is quoted making the same point in the announcement, that the larger opening is with the trade and niche outlets nobody licensed. It is a self-interested observation from a company selling PR distribution, and it is also what the numbers in the same document say.

For a brand rather than a publisher, that turns the exercise into a mapping problem rather than a negotiation. Which outlets already earn citations for your topics, what content types do they publish, and can those pages be fetched and parsed at all. That is answer engine optimization as a supply chain question, and it is the same reasoning behind publishing our own comparison of the monitoring category rather than treating a citation count as a single score.

A licensing deal does not decide whether your page can be read

Here is the boundary this site exists to draw, and the reason a study about commercial agreements belongs on a blog about crawlers.

A citation requires a retrieval, and a retrieval requires a fetch that succeeded and a page whose text survived it. Nothing in a licensing agreement changes any of those three. The vendors publish this themselves and the documents are short. OpenAI's crawler documentation separates GPTBot for training from OAI-SearchBot for surfacing a site in ChatGPT's search features and from ChatGPT-User for user-initiated fetches, so which token a site allows decides which of those can happen, whatever contracts exist. Perplexity's bots guide does the same for its own tokens, and Anthropic's page on its crawler for Claude. A publisher that signs a deal and blocks the wrong token has bought a premium on a platform that cannot reach the page.

The second gate is the one a contract cannot touch either. If the prose only exists after JavaScript runs, then what a non-rendering client receives is markup with the article missing, and that gap is prose parity, the single largest component of the score this scanner produces. It is invisible from the outside to anyone counting citations, because a citation dataset records what was cited and never why something was not. If you want to see what an unauthenticated non-browser client currently gets from one of your own URLs, that observation is what GPTBot sees, and it costs nothing to run.

None of that is an argument that the study is unimportant. It is an argument about ordering. This scanner cannot tell you whether signing with OpenAI would raise your citation rate, has no data that could answer it, and would be inventing something if it claimed otherwise. What it can tell you is whether the pages you would be signing about are currently readable by the fetchers that produce citations, which is a precondition rather than a strategy and is the only half of the problem with a definite answer.

The last thing worth saying is the least comfortable one for a vendor of measurement. If 46.9 percent of the citations going to licensed publishers land on best-of lists, buying guides and product reviews, then a great deal of what these systems quote about a brand sits on somebody else's domain, where no scanner the brand runs has any authority and no deal the brand signs has any effect. The honest scope of any tool here, this one included, stops at the pages you control and the requests they answer. Everything past that boundary is other people's publishing, and this study is a useful measurement of how unevenly that terrain is distributed rather than a lever anyone can pull.