Information Gain in SEO and AI Search

What is information gain in SEO?

Information gain in SEO is the new information a page adds beyond what other pages on the same topic already say. The term comes from a Google patent that scores a document by the information it contains beyond documents the reader has already seen.

Picture a reader who has opened three articles about the same question. All three give the same five facts in different words. A fourth article that repeats those five facts gives the reader nothing; one that adds a sixth fact — a measured number, a dated observation, a result from a test — gives the reader something the first three could not. That sixth fact is information gain.

The idea is relative by design. A fact is not “new” on its own; it is new compared with a specific set of other documents. For a searcher, that set is the pages already read. For a search engine building a results page or an AI engine building an answer, it is the other candidate pages retrieved for the same query.

SEO writers adopted the term after Google’s patent on it was granted in 2022, and it now serves as shorthand for a simple editorial test: does this page say anything the pages already ranking do not? The sections below separate what the patent actually describes from what Google has confirmed, then turn the idea into a check any writer can run.

What does Google’s information gain patent describe?

Google's patent US11354342B2, "Contextual estimation of link information gain", describes a machine-learning score for how much new information a document adds beyond documents a user already viewed. It was filed with an October 2018 priority date and granted in June 2022.

A patent protects an invention; it does not report what a company runs in production. A patent application is published before examination, and a grant follows once the patent office accepts the claims. Google Patents, Google’s public patent search service, lists both stages for this invention.

Patent number, inventors and dates

• Granted patent: US11354342B2, “Contextual estimation of link information gain”.
• Assignee: Google LLC.
• Inventors: Victor Carbune and Pedro Gonnet Anders.
• Priority date: 18 October 2018.
• Application published: 5 November 2020, as US20200349181A1.
• Granted: 7 June 2022.
• Status: active, with an anticipated expiration of 14 June 2039, per Google Patents.
• Continuations: US11720613B2 and US12013887B2 share the same title.

Articles that call it “the 2020 patent” are citing the application date; the granted patent dates from 2022.

How the patent scores a document

The abstract states the core idea in one sentence: “An information gain score indicates additional information included in a document beyond information in previously viewed documents.”

Two parts of that mechanism matter for anyone writing content. First, the score comes from a machine learning model: data from the document is applied across the model, which outputs the score. A machine learning model here is a system trained on examples to produce a prediction, not a hand-written rule. Second, the score is contextual. It is measured against the documents that a particular user has previously viewed, so the same page can have high information gain for one reader and low gain for another who has already read its sources.

That second point is why the patent’s title includes “contextual”. The patent does not describe a fixed originality grade stamped on each page. It describes a score recalculated for the reader’s situation.

Where the patent applies the score

The patent names two uses. The first is ranking: “these search results may be ranked at least in part based on their respective information gain scores”. The word “may” is the patent’s own; it describes a possibility the invention allows, not a documented feature of Google Search.

The second is conversational. The patent describes automated assistants, “also referred to as ‘chatbots’”, which present information to a user “in an abbreviated form”. Filed in 2018, the patent already anticipated a system that summarises documents for a user and needs to know which ones add something new. That is close to the job an AI answer engine does today, which is why the patent keeps coming up in discussions of AI citations. The link is conceptual; the patent does not describe AI Overviews, AI Mode or any named product.

Has Google confirmed information gain as a ranking factor?

No. Google has not confirmed information gain as a ranking factor in any public documentation. A patent shows what Google invented, not what Google Search uses. Google's guidance asks instead whether content provides "original information, reporting, research, or analysis".

A ranking factor is a signal a search engine confirms it uses to order results. Neither of Google’s two main content documents uses the phrase “information gain”: not the helpful-content guidance (updated 10 December 2025) and not the AI optimization guide (updated 10 July 2026), checked on 25 September 2026. Claims that information gain became Google’s top ranking signal in 2026 circulate without a Google source, and this page does not repeat them.

What Google does publish is guidance that points in the same direction. Its helpful-content self-assessment asks three questions that describe information gain without naming it:

“Does the content provide original information, reporting, research, or analysis?”

“Does the content provide substantial value when compared to other pages in search results?”

“…avoid simply copying or rewriting those sources, and instead provide substantial additional value and originality?”

The second question is the closest match. It compares a page with “other pages in search results”, which is the relative test the patent formalises. The honest summary is this: Google asks for the property, Google holds a patent on a way to score it, and Google has not said that the patented score is part of ranking.

Why does information gain matter for AI citations?

An AI answer cites a passage because it supports a specific sentence, and a fact found on only one page has no substitute source. Google's AI search guidance favours non-commodity content with "unique expert or experienced takes that go beyond common knowledge".

An AI citation is a link or source reference that an AI answer attaches to a claim. Before an engine can cite a page, it must retrieve it: the engine expands the question into sub-queries (query fan-out), fetches candidate pages and keeps the passages that match. After retrieval, a second gate decides which of those passages gets credited. When five retrieved passages state the same fact, any one of them can support the sentence, and the engine credits whichever it prefers. When one passage states a fact the others lack, only that passage can support a sentence built on it.

Retrieval, relevance, authority and freshness all play a part in AI source selection; information gain affects the last step, where retrieved passages compete for credit. Google’s AI features documentation adds that its models “identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links” than classic search. A more diverse link set leaves more room for the page that adds something.

Google’s AI optimization guide draws the line with two examples. “Commodity content (for example, something like ‘7 Tips for First-Time Homebuyers’) is often based on common knowledge, which could originate from anyone, and typically adds little unique insight for readers. In contrast, non-commodity content (such as ‘Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line’) provides unique expert or experienced takes that go beyond common knowledge and the ordinary.”

The research evidence is narrower than many summaries suggest. The 2023 GEO paper (Aggarwal et al., arXiv 2311.09735) found that adding statistics, adding quotations and citing sources raised a source’s visibility by 30–40% on its Position-Adjusted Word Count metric. That was a lab setup: gpt-3.5-turbo wrote each answer from the top 5 Google results, so every source was already retrieved. The paper tested adding supporting material, not originality itself, and it did not measure live engines.

Which kinds of content add information gain?

6 kinds of content add information gain: first-hand experiment results, original data, dated primary-source quotations, worked examples with real numbers, expert judgement on trade-offs and dated observations of a changing system. Each contains a fact the reader cannot find elsewhere.

Each kind below comes with an example from this site’s own topic, AI search measurement:

• Experiment results: the outcome of a controlled test, such as whether pages with added statistics gained citations against unchanged control pages.
• Original data: numbers collected first-hand, such as weekly citation counts from a fixed prompt set.
• Primary-source quotations with dates: an operator’s exact wording with the date it was checked, such as the update date of Google’s AI optimization guide next to the sentence quoted from it.
• Worked examples with real numbers: a calculation readers can reproduce, such as a cost per tracked prompt worked out from published plan prices.
• Expert judgement on trade-offs: a reasoned choice between options, such as when a manual tracking sheet stops being cheaper than software.
• Dated observations of a changing system: a record of what an engine did on a given day, such as which crawler tokens an operator documented in a given month.

The common thread is production. Every item required someone to run, count, check or decide something, and a writer who only reads other pages cannot produce any of them.

Run a fact-difference check before publishing

Run a fact-difference check in 5 steps: list the facts in the top-ranking pages, list the facts in the draft, mark draft facts missing from every competitor, confirm each marked fact against a primary source, and move the strongest one to the opening section.

The check turns information gain from a vague aim into a list. It is the method used to plan every page on this site.

1. List competitor factsRead the top-ranking pages for the target query and write each factual claim as one line: a number, a date, a name, a definition.
2. List draft factsDo the same for the draft, one claim per line, in the same format.
3. Mark the differencesMark every draft fact that appears in none of the competitor pages. Facts that appear in some but not all are shared, not new.
4. Confirm each marked factTrace each marked fact to a primary source or to first-hand work, and record the date checked. An unconfirmed new fact is a liability, not a gain.
5. Lead with the strongestMove the most useful confirmed new fact into the opening section, where readers and engines meet it first.

Here is the check applied to this page, against the two ranking pages that loaded on 25 September 2026.

4 of 5 facts on this page appear in neither competitor page. The fifth, the 2022 grant, is shared with one competitor, so it counts as coverage, not gain. The strongest new fact, the granted patent number, sits in the centerpiece and the fact card.

What looks original but adds no information gain?

Reworded competitor pages, AI summaries of existing articles, statistics without a named source and "original research" that repackages vendor data add no information gain. Each restates information already present in other pages, so it gives an engine nothing new to cite.

• Reworded competitor pages change the sentences but keep every fact, which is exactly what Google’s guidance means by “simply copying or rewriting those sources”.
• AI summaries of existing articles compress what is already published; a model working from the same pages cannot add a fact none of them contain.
• Statistics without a named source look specific but cannot be checked, so a careful engine or reader has no reason to prefer them.
• Repackaged vendor data presented as “original research” repeats numbers the vendor already published, often with less context than the original.

Each of these passes a quick originality glance and fails step 3 of the check.

A page adds new facts only when someone produced them. Where do this site’s new facts come from?

How is Grow AI Search producing original information?

Grow AI Search produces original information by running dated experiments on how AI engines cite sources. Experiment 01 tracks a new domain on 50 prompts across 6 engines weekly; Experiment 02 is planned to test whether statistics and quotations raise citation rate.

Experiment 01 is running and has reported no results yet. Experiment 02 is planned: it will apply the GEO paper’s top tactics to live pages and compare them with unchanged control pages, which moves the 2023 lab finding onto real engines. Each test is logged publicly with its hypothesis, set-up and dates among the original GEO experiments, so any fact it produces can be traced to how it was measured.

Information gain is one part of earning citations; crawler access, answer-first passages and entity signals are the others, and they are set out together in the AI search content method.