Information Gain SEO: What Google's Patent Claims and What the Tests Show

Google owns 4 granted US patents built on one idea: score a document by how much new information it adds beyond what the reader has already seen. The SEO industry calls this information gain and often treats it as a confirmed ranking factor. It is not confirmed. The patent is real, the claims are specific, and the gap between what they say and what gets repeated about them is the useful part.

The short answer. Information gain SEO is the practice of making a page add something a reader cannot get from the pages they have already read. Google patented a method for scoring that, but has not confirmed using it in ranking. Its published guidance asks for original value anyway, so the practice is sound even though the patent proves nothing about live search.

What Google's information gain patent actually covers

The parent patent is US 11,354,342 B2, "Contextual estimation of link information gain". Google LLC is the assignee. The inventors are Victor Carbune and Pedro Gonnet Anders. It was filed on 18 October 2018 as an international application and granted on 7 June 2022.

Much SEO commentary dates the patent to June 2024. That is the grant date of a continuation, US 12,013,887, not the original. The family now has 4 US grants, each with the same 2018 priority date and the same title. The mix-up matters only because it makes a six-year-old idea read like a fresh policy signal.

The patent defines its central object plainly. In its own words, the score "is indicative of additional information that is included in the given document beyond information contained in other documents that were already presented to the user". Read that sentence twice. The comparison set is what this user has already seen.

It is framed around an automated assistant

The description spends most of its length on an automated assistant, which it defines to include chatbots and virtual assistants. One of its figures is a dialog interface where the assistant answers, the user asks for more, and the next answer is chosen to add the most new material. A second figure shows a search results page.

That framing is easy to skip, and skipping it causes most of the overreach. The patent is a method for deciding what to show next in a conversation or session. It is not, on its face, a method for ranking the open web against itself.

The 2024 continuation changes the shape slightly. Its claim 1 ranks a first set of documents for a topic and provides part of the most relevant one. When the user asks for more on that topic, it scores a second set by what each adds beyond that first document. Its system claim begins with a query that includes a topic, and it does not recite an assistant.

So the family moved from an assistant setting toward a generic query-and-follow-up setting. That is a reasonable thing for a patent attorney to do with a good idea. It is not evidence about which Google product, if any, runs the method.

How the information gain score is calculated in the claims

Claims are the legally binding part of a patent. The description can speculate freely, and the claims cannot. If you want to know what Google actually fenced off, read the claims and ignore the summaries.

US 11,354,342 has 16 claims, with three independent ones: two methods and a system. The first set is what the user has seen, and the second set is the candidates. Claim 1 bases the score on "a quantity of new information included in the given new document of the second set that differs from the information extracted from each document of the first set".

A model, not a word count

Claims 3 and 9 apply data from the documents across a machine learning model and take the score from its output. Claim 15 specifies that the model comprises a neural network. Claims 4 and 16 compare documents using semantic feature vectors or semantic representations rather than raw text.

This is the detail most "information gain checklists" get wrong. A semantic comparison does not reward a page for using different words. Rewording a competitor's article produces a near-identical vector. Only a different claim, number, example or conclusion moves it.

Claim 2 ranks the candidate documents by their scores and selects from that ranking. Claim 7 covers a search results page shown when a user returns from a document they opened. Claim 8 excludes the already viewed document from those results, and the description says viewed results may be "significantly demoted."

Put together, the method is session reranking. You read result one, you come back, and the next results are reordered toward whatever adds the most beyond result one. That is a narrower and more plausible mechanism than a sitewide originality grade.

What Google has confirmed about information gain, and what it has not

Three categories get blurred in almost every post on this topic. Some things are documented in the patent. Some things Google has said publicly. Everything else is inference, and most of the confident advice lives there.

What Google has said on the record

Google's helpful content guidance, was last updated 5 October 2026. One of its questions reads: "Does the content provide original information, reporting, research, or analysis?" It never uses the phrase information gain. It also asks whether content drawing on other sources avoids simply rewriting them. Such content should provide "substantial additional value and originality".

On AI features, Google is more explicit than SEO vendors would like. Its AI features documentation says: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary". The same page says both features may use a "query fan-out" technique, issuing multiple related searches across subtopics and data sources.

The Search Central blog post of 21 May 2025 on AI experiences opens with a familiar instruction: "Focus on your visitors and provide them with unique, satisfying content". It describes the goal as helping people find "original content that adds unique value". That is the closest Google has come, and it is a statement of intent, not a disclosed signal.

Google staff have warned against reading patents as product documentation for over a decade. Matt Cutts said in 2013 that a patent issuing with a search engineer's name on it "doesn't necessarily mean we are using that patent at that moment," as reported by Search Engine Land. John Mueller later made the same point in a Search Central hangout, saying a patent "doesn't mean that we actually use it in search".

When Google's internal API documentation leaked in May 2024, a spokesperson confirmed the material was authentic. The company cautioned against "inaccurate assumptions about Search based on out-of-context, outdated, or incomplete information," per The Verge. Neither the leak nor that statement confirmed an information gain score.

Status reflects public sources available on 23 October 2026. "Speculated" means no primary Google source was found, not that the claim is false.

What the tests on unique content in AI search show

If the patent cannot settle the question, measurement might. There are two kinds of evidence, and they point in different directions.

The peer-reviewed study: specific edits raised visibility

The Princeton and IIT Delhi paper GEO: Generative Engine Optimization, accepted to KDD 2024, built a benchmark of 10,000 queries. It edited one source per query and measured how much of the generated answer drew on it. The abstract reports that the methods "can boost visibility by up to 40%" in generative engine responses.

The useful detail is which edits worked. Adding statistics, quotations and source citations lifted visibility on the paper's position-adjusted word count metric. Keyword stuffing scored below the unedited baseline. In plain terms, adding evidence helped and repeating terms did not.

The limitation is large and the paper is open about its setup. The main experiments used a pipeline that fetched the top 5 Google results and generated answers with gpt-3.5-turbo. That is a simulated engine, not production AI Overviews. A follow-up on Perplexity reported gains up to 37%. A wider survey of this literature is in the review of 45 GEO studies on what actually gets cited.

The vendor test: no originality edge for cited pages

On 15 July 2026, On-Page.ai Research captured Google AI Overviews for 50 keywords across 10 verticals on a single day. It scored the ranking pages with its own originality metric, which it calls an Information Gain Score, measuring what each page adds beyond the pooled content of the other ranking pages.

Cited and uncited pages came out "statistically indistinguishable on originality," with cited pages slightly lower. Legal content was the one vertical where cited pages looked markedly more original, on small samples. Treat all of it as vendor research. The metric is proprietary, the company sells the tool that reports it, and the keyword set was hand-built rather than random.

The same study found most AI Overview citations pointed at pages outside Google's first page of results, led by YouTube. That echoes the pattern in the analysis of how ChatGPT citations ignore Google rank. Citation and ranking are drifting apart, which is why a ranking patent is a weak guide to citation behaviour.

Why content differentiation matters more once answers are synthesised

Here is where I part company with both camps. The patent does not prove Google rewards originality, and the vendor test does not prove it ignores it. The stronger case for content differentiation comes from how synthesis works, not from any single signal.

An answer engine that fans one query out into many searches collects overlapping pages. Whatever those pages agree on becomes the answer, and the answer does not need to cite any particular one of them. A page that says what ten others say is interchangeable, and interchangeable pages are the first to go uncited.

A page that holds the only copy of a number, a worked example or a reasoned dissent is different. If the engine uses that material at all, it has one place to attribute it. That is a structural argument, and it holds whether or not any information gain score exists in production.

Commodity sections versus differentiated sections

The practical question is which parts of a page are commodity. Most posts are 70% restatement by design, because readers need context. The test is whether the remaining part exists and is findable.

The last row is the honest one. Differentiated content costs more and carries risk. For a short definitional query, the commodity version is often the right call, and pretending otherwise wastes budget. The economics of choosing between volume and depth are covered in the breakdown of content flood economics.

Google's scaled content policy targets pages made at volume without adding value. Differentiation is the positive version of the same rule. The cases where that policy bit hardest are traced in the review of scaled content abuse penalties, and most of the sites hit had no information a reader could not get elsewhere.

Where the information gain argument is weakest

This post argues for differentiation while denying that the patent proves anything. That position has its own soft spots, and they deserve the same scrutiny.

Absence of confirmation is not absence of use

Google rarely confirms individual signals, and it has every reason not to. A method that reorders results after a user returns from a page would be invisible to outside testing. The fact that no one has confirmed information gain ranking does not mean Google does not run something like it. My position is agnostic, not dismissive.

The evidence on citation is thin on both sides

The KDD study used a simulated engine and edited one source per query. The July 2026 test used a vendor's own metric on a single day and 50 hand-picked keywords. Neither is strong enough to bear much weight. Agency posts quote multipliers for "data-rich" pages from large citation datasets, and I could not trace those figures to a primary method, so none appear here.

Differentiation can mean being wrong in a new way

The cheapest way to add new information is to add claims nobody else makes. Some of those claims are new because they are false. A page can maximise novelty and minimise accuracy at once. An information gain score, if one exists, would measure difference, not truth, and the reader's trust is spent on the second.

A content differentiation test you can run on any draft

You do not need a patent model to check whether a draft adds anything. You need the same comparison set the patent uses: the pages your reader has probably already seen.

Start by collecting that set. Take the top 5 organic results for the target query and the AI Overview or ChatGPT answer for it. List every claim, figure and recommendation that appears in at least 2 of them. That list is the consensus, and the answer engine already has it.

Then mark your draft against the list. Every paragraph that only restates a consensus item is context. Every paragraph that does not appear on the list is your information gain. I would not ship a post where that second group fits in a single paragraph.

Finally, check the second group for the four things the GEO study and Google's guidance both point toward. Those are a sourced number, a primary document read in full, a real example, and a stated position with its own limitation. Pages with all four rarely read like the other ten. The discipline of producing your own numbers is set out in the case for original research as a content moat.

I would skip any tool that sells an "information gain score" as a ranking forecast. At best it measures distance from the current top results, which is a useful editing prompt. Sold as Google's number, it is a guess dressed as a metric.

Frequently asked questions

Where to start this week

Pick the one post on your site that brings the most search traffic. Run the consensus test above against its target query, and count how many paragraphs survive as genuinely new. If the answer is fewer than three, that post is exposed to every answer engine that can summarise its competitors.

Then add one thing to it that nobody else has: a number from your own systems, a primary document quoted accurately, or a position you are prepared to defend. Date the change, and check the AI answers for that query again in six weeks.

Related reading