Google has held a patent called "Contextual estimation of link information gain" since June 2022, and large parts of the SEO industry still cite it as proof that unique content outranks and outcites generic content. Read the patent itself, filed in October 2018 by Google engineers Victor Carbune and Pedro Gonnet, and it describes something narrower: scoring how much new information one unread document adds compared with what a single user has already seen in one browsing session, not how a page compares to its competitors across the web. That gap between the patent's actual claims and what gets repeated about it matters, because the GEO research that followed found uniqueness really does predict AI citations, just not through this mechanism.
What the information gain patent actually describes
The patent, granted to Google as US 11,354,342 B2 in June 2022 and followed by a related continuation granted in June 2024, trains a model to compare a candidate document's meaning against documents a user already viewed in that session, then scores how much new information the candidate would add. A low score means the content mostly repeats what the session already covered.
The application frames this around automated assistants and chatbots recommending what to read or hear next, with search ranking mentioned only as a secondary use. Search Engine Journal's Roger Montti read the full claim language in 2024 and made the point that keeps getting lost: nothing in it says a page with more information than a competing page should rank higher. The score only ever compares a candidate document to what one person has already seen, not to the rest of the web, and Montti found no public evidence Google applies it inside AI Overviews, organic ranking, or any other live product.
What independent GEO research actually finds about uniqueness and citations
Two 2026 studies, run independently of Google's patent, both tie uniqueness to real citation outcomes. On-Page.ai scored 150 top-three-ranking pages for originality on a 0-to-100 scale and found pages with fifteen or more unique data points averaged 62, against 40 for pages with one or none. Positions one through three averaged about 51.4 on the same scale, with little separating the top spots on originality alone.
Separately, the paper that coined the term Generative Engine Optimization (GEO), by Aggarwal and colleagues at Princeton, built a benchmark of generative-engine queries and found that rewriting content with real statistics, direct quotations, and named sources lifted visibility in AI answers by as much as 40%. Keyword-heavy rewrites did worse than making no changes at all.
Ahrefs' Ryan Law frames the resulting gap simply:
Information gain is Google rewarding content for being different, not just better.
Why word count and out-covering a competitor aren't the lever
Two more 2026 findings rule out the explanations people reach for first. Ahrefs measured a correlation of just 0.04 between word count and AI Overview citations across 174,048 pages, confirming a deeper look at content length and AI citations that found almost the same thing. And when a site runs several pages that answer close to the same question, AI engines tend to group them and cite only one, which is rarely the longest page in the group.
That second finding is the part most information gain explainers skip. The gain an AI system gets from a passage is measured against what the model could already produce or had already retrieved elsewhere in that answer, not against a word count or a tally of facts. Research into query fan-out, the practice of splitting one search into dozens of background sub-queries, found retrieval rank and relevance predicted citations far better than topic coverage did, because each passage gets retrieved and judged on its own. Searchbloom, a GEO agency, argues from its own unverified 400-page audit that insight density should be counted per section rather than per page for exactly that reason.
The AI-generated content problem this actually points to
A 2026 position paper from the GEO firm Passionfruit argues retrieval systems cite a source in proportion to the information it adds beyond what the model could generate on its own, something it calls a redundancy problem rather than a quality problem. AI-written text, shaped by similar alignment training across many tools, tends to converge toward similar phrasing and framing, leaving less of that gap to fill.
The paper leans on separate, peer-reviewed work to make that case: a 2024 Nature study on model collapse from repeatedly training on AI-generated data, and ICLR research showing reinforcement learning from human feedback narrows how varied a model's outputs are. Passionfruit's own predictions from this theory are untested by the paper's own account, and the firm sells the GEO services its thesis recommends. Treat it as a plausible explanation worth watching, not a confirmed mechanism.
What actually changes if you want AI engines to cite you
None of this says length or coverage is worthless, only that neither is the mechanism the information gain patent describes. The lever both real studies found independently of each other is the same one: original data a competitor cannot easily copy, named sources cited close to the specific claim they support, and passages that answer a query's exact phrasing on their own.
That's the same lesson the broader shift from SEO to GEO keeps teaching: there's no evidence the patent itself runs inside Google AI Overviews or any assistant brands can audit from the outside, so a strategy built only around "satisfying information gain" optimizes for a mechanism nobody has confirmed is live. The research that is confirmed, built on actual citation data rather than patent language, keeps landing on the same practical list: real numbers, real sources, and content that doesn't just restate what a dozen other pages already said.
The stakes for getting this wrong are already visible in how people use AI summaries at all. Pew Research found Google users clicked a traditional result in just 8% of visits where an AI summary appeared, against 15% where none did, and only 1% of visits with a summary ended in a click on a cited source inside it.
Frequently Asked Questions
What is Google's information gain patent?
It's a patent Google filed in 2018 and was granted in June 2022 as US 11,354,342 B2, with a related continuation granted in June 2024. It describes scoring how much new information a candidate document adds compared with documents a specific user has already viewed in one session, mainly for recommending what an automated assistant or chatbot should surface next.
Does information gain predict whether AI engines cite your content?
Not directly. The patent never claims to compare a page against its competitors, and there's no public evidence Google uses it in AI Overviews or organic ranking. Separate 2026 research does find that pages with more original data and named sources get cited more often, but that's an independent finding, not proof the patent itself is the mechanism.
Does word count affect information gain or AI citations?
Barely. Ahrefs found almost no correlation, 0.04, between word count and AI Overview citations across more than 174,000 pages. What correlates instead is original data: pages with fifteen or more unique data points scored noticeably higher on an independent uniqueness study than pages with one or none.
How can I tell if my content has enough original information?
Check whether each section states a fact, figure, or finding that a competitor's page on the same topic doesn't already say, and whether that fact traces to a named, checkable source. If a section only restates what the top-ranking pages already cover, it's unlikely to add the kind of information gain that independent GEO research ties to citations.
Does using AI to write content hurt your chances of getting cited by AI engines?
It might, according to a 2026 theory from the GEO firm Passionfruit, because AI-written text trained with similar alignment methods tends to converge on similar phrasing, leaving a retrieval system less new information to cite. The theory draws on real model-collapse research, but its own predictions are still untested, so treat it as a caution rather than a proven rule.