A study released on August 20, 2026 put a number on something the industry has assumed without evidence: publishers with an OpenAI licensing agreement get cited more by ChatGPT. The AI citation premium is 48%. The more useful finding is buried further down, and it is the one that determines whether any of this is actionable for a publisher that will never be offered a deal. This piece covers what was measured, why the platform-by-platform split makes the result more credible than it first appears, how concentrated the benefit is, what it does to citation optimization advice, the caveats that belong on it, and what is left to actually do.
What the study measured
Press Ranger and OtterlyAI published an analysis of 129.3 million citations across more than 20 million cited URLs, drawn from June 2026 across seven platforms: ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot, Gemini and Claude.
The licensing side came from a Press Ranger database of 91 confirmed agreements between AI companies and publishers, mapped to 314 publisher domains.
The headline result is straightforward. On ChatGPT, pages from publishers with an OpenAI deal averaged 10.2 citations against 6.9 for pages from publishers without one. Across all seven platforms together the gap narrows to 46%, at 10.7 against 7.3.
| Measure | Licensed | Unlicensed | Gap |
|---|---|---|---|
| ChatGPT citations per page | 10.2 | 6.9 | 48% |
| All seven platforms | 10.7 | 7.3 | 46% |
The platform split is what makes the finding credible
The obvious objection to a result like this is confounding. Publishers who sign licensing deals are disproportionately large, established, high-authority domains. Those publishers would be cited more often anyway, deal or no deal, so a raw comparison tells you about publisher size rather than about licensing.
That objection is correct in principle, and the study’s own numbers are what answer it.
If the premium were purely a size and authority effect, licensed publishers would show it everywhere. They do not. The same set of publishers with deals was cited on Google AI Overviews at a slightly lower rate than comparable unlicensed publishers, and landed at parity on Perplexity.
That asymmetry is difficult to explain by publisher quality. A New York Times page does not become more authoritative when ChatGPT reads it and less authoritative when Google AI Overviews does. Something platform-specific is happening, and the platform where it happens is the one where the commercial relationship exists.
Thomas Peham, OtterlyAI’s CEO, puts it more narrowly than the headline does, telling reporters: "A licensing deal does one clear thing: it tilts your citations toward ChatGPT."
Note the absence of the reverse claim. Google-licensed publishers did not get a Google premium. Perplexity-licensed publishers did not get a Perplexity premium. Whatever produces the effect on ChatGPT is not a general property of licensing deals.
Five groups take 69% of it
Here is the number that should reframe the whole conversation for anyone running a site below the top tier.
The five largest media groups captured 69% of all licensed publisher citations. Not 69% of citations overall, but 69% of the benefit accruing to the 314 domains that have deals at all.
Separately, news accounted for only 7.2% of all citations across the study. The category everyone is arguing about is a small slice of what these systems actually cite.
Stack those two facts. Ninety-one agreements exist. They cover 314 domains. Five groups take roughly seven-tenths of the resulting citation benefit. The pool of publishers meaningfully advantaged by licensing is not the 314; it is a handful of them.
The ai citation premium is not available to most publishers
This is the part nobody writing about generative engine optimization wants to say plainly, so we will.
If citation share on the largest AI platform tracks a commercial agreement, and those agreements are extended to a few hundred domains with the benefit concentrated in five groups, then a publisher outside that set is optimizing inside a ceiling. The techniques still work. They work within a band that has been set by something other than content quality.
We wrote our guide to writing for AI citation on the assumption that structure, clarity and sourcing drive citation. Nothing in this study contradicts that for the comparison it can actually make, which is between unlicensed publishers. Our piece on citations and rankings decoupling argued that traditional ranking signals no longer predict AI citation. This adds a signal that is not editorial at all.
That is a real qualification to our own advice and it deserves stating rather than burying. The practical guidance does not change. The expected ceiling does.
It also sharpens what we covered in the traffic collapse piece. A small publisher losing search traffic to AI summaries has been told the answer is to optimize for citation instead. That remains the best available answer. It is not a level playing field, and pretending otherwise has been the weakest part of the GEO conversation.
What a licensing deal plausibly grants
The study establishes an association and stops there, which is the correct scope for it. But the mechanism question is worth thinking through, because the answer determines whether the ai citation premium is a durable structural advantage or a side effect that erodes.
Three explanations fit the data, and they have very different implications.
| Explanation | What it would mean |
|---|---|
| Preferential retrieval for licensed sources | A deliberate ranking input. Durable, and not reachable by anyone else |
| Better content access from the agreement | Cleaner text, no paywall truncation, fuller crawl. Reachable in part by anyone with good technical access hygiene |
| Training data overlap | Licensed corpora are better represented, so the model reaches for them. Decays as models are retrained on different mixes |
The second one is the interesting possibility for a publisher without a deal, because it is partly a technical problem rather than a commercial one. A licensing agreement typically comes with cleaner access to full article text. A publisher who makes their content equally accessible to crawlers, without paywall interstitials or JavaScript-dependent rendering, may capture some of the same effect without any agreement.
Nobody has shown which of the three is operating, and it may be some combination. Media Copilot’s coverage and OtterlyAI’s own writeup both stop short of claiming a mechanism, which is to their credit.
The caveats that belong on this
Three, and they are substantial enough that nobody should treat this as settled.
It is a vendor-adjacent study. Press Ranger and OtterlyAI both sell AI search visibility services. Neither has an interest in concluding that visibility is unwinnable, and the finding as framed supports a product narrative about tracking citations. That does not make the numbers wrong. It does mean the framing should be read with the same skepticism you would apply to any vendor research.
It is correlation. The platform asymmetry argues strongly against a pure size effect, which is why it is worth taking seriously, but no mechanism has been demonstrated. Nobody has shown that a licensing agreement causes preferential retrieval. Several other explanations survive, including that licensing grants cleaner or fuller content access which independently improves retrievability.
It is one month. June 2026, a single snapshot. Retrieval behavior on these platforms changes with model versions, and AI Mode has been served by a different model for some queries since. A one-month window cannot distinguish a durable structural effect from a temporary artifact of one model generation.
What to actually do about it
The advice is short, because most of the honest answer is that this is not a lever most publishers can pull.
Keep doing the citation work. Between unlicensed publishers, which is the comparison that applies to nearly everyone, structure and sourcing are still what separate cited pages from uncited ones.
Set expectations to the right benchmark. Comparing your citation rate to a large licensed publisher’s is measuring the wrong gap. Compare against publishers in your own tier.
Do not over-index on news. At 7.2% of all citations, news is a small share of what gets cited. Reference, explanatory and technical content are most of the rest, and that is where a small publisher’s returns are.
Watch whether the asymmetry holds. If the ChatGPT premium persists across several months and model versions while Google and Perplexity stay flat, the structural reading gets stronger. If it decays, this was a snapshot.
Treat platform diversity as risk management. The effect is concentrated on one platform. Content that performs across Perplexity, Copilot, Gemini and Claude is less exposed to whatever ChatGPT is doing.
Frequently Asked Questions
What is the AI citation premium?
The gap in citation frequency between publishers with an AI licensing agreement and those without. In this study, pages from publishers with an OpenAI deal averaged 10.2 citations on ChatGPT against 6.9 for unlicensed publishers, a 48% difference.
How large was the dataset?
129.3 million citations across more than 20 million cited URLs, from June 2026, spanning ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot, Gemini and Claude. Licensing data came from 91 confirmed agreements mapped to 314 publisher domains.
Does the effect appear on Google or Perplexity?
No. Google-licensed publishers were cited on Google AI Overviews at a slightly lower rate than comparable unlicensed publishers, and Perplexity’s licensed publishers were at parity. The effect is specific to ChatGPT.
Does this prove licensing causes more citations?
No. It establishes a correlation. The platform asymmetry makes a pure publisher-size explanation hard to sustain, which is the strongest argument for taking it seriously, but no mechanism has been demonstrated and other explanations remain open.
Should I stop optimizing for AI citation?
No. For the comparison that applies to most publishers, which is against other unlicensed publishers, the usual work still determines outcomes. What changes is the ceiling you should expect, not the method.
Why does the concentration figure matter so much?
Because the five largest media groups captured 69% of licensed publisher citations. Even among the 314 domains with agreements, the benefit is not evenly spread, so “get a licensing deal” is not a strategy available to almost anyone reading this.
How reliable is the study?
The dataset is large and the platform breakdown is the kind of detail that makes a finding checkable. It is also published by two companies that sell AI search visibility tools, covers a single month, and shows correlation rather than mechanism. Weight it accordingly.
Is news content still worth producing for citations?
News made up only 7.2% of citations in this dataset, so it is a smaller share of AI citation volume than its share of the debate would suggest. Reference and explanatory content account for far more, which is where a smaller publisher’s effort tends to return more.
