Every big core update spawns a new phrase that SEOs repeat for a year. In 2026 that phrase is "information gain," and unlike most of them, this one actually points at a real, specific Google patent rather than a vibe. It's worth understanding exactly what the patent says, what Google has and hasn't confirmed about using it, and what the pattern-matching around the March and May 2026 core updates actually shows, because a lot of what's being written about it overstates the certainty.

Search "information gain SEO" right now and you'll find dozens of posts stating flatly that this is Google's "number one ranking signal" for 2026. Almost none of them link to anything Google actually said, because Google didn't say it. That gap between confident headlines and what's actually been confirmed is exactly the kind of thing worth sorting out before changing how you write content based on it.

The patent itself

Google holds patent US20200349181A1, "Contextual estimation of link information gain," filed in October 2018 and published in November 2020. It describes a system that scores a document by how much new information it would add for a specific user, compared against a set of documents that user has already been shown on the same topic. The patent's own language calls this an "information gain score," defined as the additional information contained in a given document beyond what's already in the documents the user has already seen.

Mechanically, the system takes two sets of documents, one the user has already viewed and one of new candidates, and runs both through a trained machine learning model using semantic representations (an embedding, a feature vector, or a simpler bag-of-words form) rather than raw text. The output ranks or filters the new candidates by how much genuinely new information each one would contribute. The patent frames a lot of this around conversational, dialog-style interactions, specifically calling out reducing redundant output in voice-based exchanges, which matters once you consider how closely that resembles what a chat-based AI answer has to do: decide what new fact or source actually deserves to be surfaced next, rather than repeating what's already been said.

What Google has actually confirmed, and what it hasn't

This is the part most of the coverage skips. Google has not named "information gain" as a live ranking signal in any public statement, has not confirmed this specific patent is in production, and routinely files patents that never ship as anything, which Google's own search team has pointed out for years whenever a patent story circulates. Holding a patent describing a scoring method is not the same as confirming Search uses that exact method today, at this weight, for this purpose. Treat every claim that starts with "Google's information gain signal now does X" with that in mind; what's true is narrower and more interesting than that framing suggests.

What is real: Google's March 2026 core update produced unusually sharp volatility, and a cluster of SEO trackers and agencies, watching ranking movement during and after it, converged on "pages with genuinely new information gained, pages that repackaged existing information lost" as the best explanation they could construct for the pattern. That's a real, observed correlation worth paying attention to. It's industry inference built from watching rankings move and reaching for the patent that best explains the shape of that movement, not Google confirming a mechanism.

Google's history with patents that never became confirmed features

It's worth some healthy skepticism here, because the SEO industry has been through this exact cycle before. Google has filed and been granted hundreds of patents over the years describing scoring systems, from "reasonable surfer" link weighting to various freshness and query-deserves-freshness concepts, and only a small fraction of them are ever acknowledged, even loosely, as something Search actually uses. Search Engine Journal's own coverage of this patent is careful to frame it the same way: interesting, plausible, consistent with observed behavior, and still unconfirmed. The pattern repeats often enough that "Google filed a patent for X" should never, on its own, be read as "Google now ranks pages using X." It's one data point that becomes meaningful only once real-world ranking behavior lines up with what the patent describes, which is exactly what happened here, but inference from correlation is still inference.

What the March 2026 core update actually looked like

Google's March 2026 core update ran from March 27 to April 8, producing real, measurable volatility, though trackers don't even fully agree on how to characterize it: Digital Applied's tracking put the Semrush Sensor peak around 8.7 out of 10, read as the sharpest reading since the August 2024 update, while Search Engine Roundtable's own coverage cited Sensor scores closer to 9.4-9.5 during peak days, yet described the update itself as "a weird one" that didn't feel as powerful in practice as some earlier broad core updates. That disagreement between two respected trackers, on an update everyone agrees happened and moved rankings, is a useful reminder of how much interpretation is baked into "what this update was about" writing generally, information gain included. Tracking firms that studied the aftermath, Digital Applied among them, reported pages built on proprietary data, first-hand case studies, or original frameworks gaining visibility in the range of 15 to 25 percent, while templated or lightly-rewritten content and low-effort AI-generated pages lost significantly more, in some reports 30 to 80 percent depending on how derivative the page was. Those specific percentages come from one tracking firm's analysis of the update, not from Google, and different trackers report different numbers for the same event, which is normal for any core update and worth remembering before treating any single figure as precise.

The qualitative pattern across multiple independent write-ups agrees more than the exact numbers do: a shorter page built around one real, original data point reportedly outranked longer, more "comprehensive" pages that mostly synthesized what was already in the top results elsewhere. If that holds, it's a meaningful shift in what "good content" means for ranking purposes, away from exhaustive coverage of a topic and toward a narrower question: does this page know something the top results don't already say.

A practical way to audit a page for this

Pick the query a page is meant to rank for, pull the current top three or four results, and read them back to back before rereading your own page. List, honestly, what facts, data points or specific claims appear in your page that don't appear in any of those top results. If that list is short or empty, the page is functioning as a synthesis of what's already ranking, however well-written it is, and that's precisely the category this update pattern appears to penalize. If the list has real entries, specific numbers you measured, a claim you verified against a primary source that the other results didn't bother checking, a case most other guides skip, that's the part of the page actually doing work under this theory, and it's worth making more prominent rather than burying it under more generic framing in paragraph six. This is close to the same discipline behind the technical SEO checklist on this site: prioritize the few things with real, measurable impact over a long list of things that feel thorough but don't move anything.

Why this connects directly to AI answers, not just classic SEO

The information gain concept maps unusually well onto how an AI answer engine, ChatGPT, Gemini, an AI Overview, actually has to decide what to cite. These systems are explicitly trying to avoid restating the same fact five different ways from five different sources; when a model is synthesizing an answer, a source that adds a genuinely new fact or angle is more useful to cite than the sixth source saying the same thing the first five already said. That's the same underlying math the patent describes for a search results page, applied to citation selection instead of ranking position. It's a reasonable explanation for why pages built from proprietary data and original research tend to show up disproportionately often as cited sources in AI Overviews and chat answers, something we've seen anecdotally across the roundups and reviews on this site that lean on our own pricing research rather than repeating someone else's numbers. It's also consistent with how LLM seeding and query fan-out both already describe AI search behavior on this site: a model answering one question actually runs several related sub-queries behind the scenes and stitches the results together, and a page that only repeats a well-known fact is redundant across most of those sub-queries, while a page with one specific, verifiable detail can end up being the unique answer to one of them.

Why this is a problem specifically for AI-generated content at scale

A lot of the volatility trackers reported after March 2026 concentrated in exactly the kind of pages an AI writing tool produces by default: take a target keyword, pull the themes from the current top 10, write a comprehensive piece covering all of them. That workflow is, almost by definition, optimized to maximize topical coverage of what's already ranking, which is the opposite of information gain as the patent describes it. None of this means AI-assisted writing is inherently penalized; a tool used to draft around a founder's own data, a company's own usage statistics, or a genuinely new argument doesn't have this problem, because the gap is in the sourcing, not the sentence-generation. The problem is specifically the common pattern of using AI to synthesize what other pages already say faster than a human would, which is exactly the content category multiple trackers reported losing the most ground.

What to actually do about it

The practical response isn't complicated, even if the signal's exact mechanics are uncertain. Stop producing pages whose only contribution is reorganizing information that's already sitting in the top 10 results; a well-written synthesis of other people's research doesn't carry the same weight it used to, and after March 2026 it may carry meaningfully less. Build pages around something only you have: your own usage data, a survey you actually ran, pricing you actually verified against vendor pages rather than copied from a competitor's roundup, a case study with real numbers. If a page genuinely can't offer anything new on a topic that's already well covered, the more honest move is to not write it, or to fold the point into a page that does have something original, rather than publishing a 2,000-word restatement hoping volume compensates for the lack of a new idea.

This also changes how to think about competitive content audits. The old version of that exercise asked "what are competitors covering that we're not," which rewards comprehensiveness. The information-gain-aware version asks "what does the current top-ranked content collectively fail to say," which is a narrower, harder, and more valuable question, and it's closer to how our own competitive gap analyses on this site try to work: find the specific angle nobody else has covered rather than restating what twenty other roundups already said.

The honest limitation here

Nobody outside Google can say with certainty how heavily information gain, as this patent describes it, actually factors into the March or May 2026 updates, or whether the volatility those updates produced has a different primary cause that the SEO industry hasn't correctly identified yet. Google doesn't confirm signal weights, and post-update forensic analysis from outside trackers, however consistent across firms, is still inference from ranking movement, not a leaked algorithm. Treat "information gain" as the best current working theory for a real, measurable shift in what wins, not as a confirmed mechanical fact you can optimize against with precision.

Is information gain the same thing as E-E-A-T?

No, though they overlap. E-E-A-T (experience, expertise, authoritativeness, trust) is about whether a source is credible to write about a topic at all. Information gain, as the patent describes it, is about whether a specific document adds something new relative to what's already been shown, regardless of how credible the source is. A highly credible source that only restates known information scores well on E-E-A-T signals and poorly on information gain; the two aren't the same test.

Can an AI content tool help with information gain, or does it work against it by definition?

It depends entirely on what the tool is doing. An AI tool used to synthesize and rewrite existing top-ranking content, the exact use case most AI SEO content tools are built around, works against information gain almost by construction, since synthesis is the opposite of adding something new. An AI tool used to help analyze your own first-party data, draft a case study from your own usage numbers, or structure original research you've already gathered doesn't have the same problem, because the novelty comes from the underlying data, not from how the sentences got written.

How would I know if my own content has low information gain?

Read your top pages next to the three results currently outranking you for the same query and ask honestly whether your page says anything those three don't already say between them. If the answer is no, that page is a synthesis, not a source, and it's a reasonable candidate for the kind of original data, first-hand experience, or specific claim that would give it something to add.

Should I rewrite my whole site around "information gain" right now?

No. Treat it as a direction to lean toward over time, not an emergency migration. Prioritize it on pages that are already competing in a crowded, well-covered topic where a synthesis genuinely has nothing left to add; a page that already serves a specific, narrow query well doesn't need to be rebuilt around a patent nobody has confirmed is live. Spend the effort on new content and on the pages currently underperforming for reasons you can't otherwise explain.

Did the May 2026 core update reinforce the same pattern as March?

Multiple trackers reported the May update behaving consistently with the March pattern rather than reversing it, which is part of why "information gain" became the dominant explanation industry-wide rather than a one-update theory. That consistency across two consecutive updates is itself the strongest evidence for the pattern, even without Google confirming the mechanism directly.