A playbook for ecommerce and marketing teams — SEO and AEO
Information gain measures how much new information a page adds beyond what a searcher has already seen. It rewards content for being different, not merely better — and that distinction is the whole point. Most teams optimize for "better": more comprehensive, more words, more sub-headings. Comprehensiveness is now the cheapest thing in the world to produce, which means it has stopped being a differentiator. What remains scarce is information no one else has.
Four properties of the mechanism dictate everything downstream.
1. It is scored per-user, per-session. The score compares a page against the documents a specific user has already been shown. Someone who has just read five near-identical pages finds a sixth redundant; someone arriving cold finds the same page genuinely useful. Gain is contextual, not absolute — the same page can be worth a lot or nothing depending on what preceded it.
2. It compares meaning, not wording. Each page is read as a semantic representation: what it means, converted into numbers. Two pages with almost no shared vocabulary can be recognized as saying the same thing and land at the same coordinates. The practical consequence is decisive and routinely ignored: rewording a competitor's page adds nothing. A rewrite occupies the same semantic position as the original. Publishing original data is one of the only reliable ways to move a page to genuinely new coordinates.
3. It is dynamic. Scores recalculate mid-session. As a user consumes pages, the remaining unseen pages are re-ranked based on what has already been covered. This re-ranking is distinct from the initial ranking every searcher sees on the results page.
4. It prunes duplicate facts, not just duplicate pages. This is the property that bridges SEO and AEO. An answer engine takes one fact from a source, then draws on the next source only for what it newly adds, deleting the overlap before the user ever sees it. This behavior extends explicitly to chatbots, conversational agents, and voice assistants. The implication is the single most important idea in this document:
In a ranked list, the fourth page saying the same thing still holds position four and can still earn a click. In a synthesized AI answer, there is no position four. A page either supplies information that has not been said yet, or it is cut from the answer entirely.
Two ways to read the mechanism. The literal reading is that each page is scored, then results are re-ranked. A second, more strategic reading treats information gain as an output rather than an input: search ranks a page on authority and entity coverage, observes whether real users are satisfied (click-through, dwell time, pogo-sticking), and once the page has proven itself, harvests its claims and folds them back into its own understanding of the topic — as raw material for future AI answers, as another data point in an entity's coverage, or as the new baseline for what counts as consensus next time. Both readings point to the same conclusion: gain lives on the fringe of a topic — the concepts, data, and perspectives the results set hasn't picked up yet. The core of any topic is saturated because everyone has already written it. Value sits at the edges.
Is a literal "information gain score" confirmed? No, and it's worth being precise about this. The mechanism is described in a granted patent, but no live score has ever been confirmed, and many patents never ship. What is certain is that the underlying preference is real and consistently rewarded, for three grounded reasons: official quality guidance repeatedly asks whether content is original and adds substantial value compared to other results; sites that gain from core updates reliably over-index on original assets like proprietary data and tools; and engines already re-rank what a user sees mid-session based on real behavior, which is mechanically how the patent is described as working. The objective is therefore not to chase a score. It is to be the source of something not yet consensus.
No category on the web is more duplicative than product content. Retailers paste manufacturer spec sheets verbatim, resell identical SKUs with identical descriptions, and mirror each other's category copy word for word. On marketplaces, hundreds of listings for the same item converge on the same bullet points. Measured semantically, most of the shopping web sits directly on top of itself.
That creates a two-sided situation. The downside: the sameness penalty falls harder here than anywhere else, because so little product content adds anything. The upside: in a field where almost everyone publishes the same thing, the brand that publishes something genuinely original stands out by a margin no content-saturated vertical can offer.
And the stakes are commercial, not academic. The queries that decide revenue — "best [product] for [use case]," "[brand] vs [competitor]," "is [product] worth it," "which size am I" — are increasingly answered by engines that cite a small handful of sources and drop everyone else. Whether a brand is named in those answers depends almost entirely on whether it holds information the consensus does not. Information gain is no longer a blog-team nicety. It is the mechanism that determines whether a brand exists inside a shopping answer at all.
Gain applies to every page type, and each has a different fringe to fill. Mapping this first tells a team where original data will pay off, before deciding what to produce.