Information Gain Score: How Google Measures Whether Your Page Adds Something the Index Doesn't Already Have

The information gain score is the metric ROI.LIVE approximates for every article before publication. Google doesn't expose the score in any public tool. You can't look it up in Search Console or Ahrefs. But the patent (US10776471B2) describes the mechanics clearly enough that Jason Spencer built a practical proxy for it: The Delta Audit. This article explains what Google's system measures, how the scoring works at a technical level, and why the most common interpretation of the patent is wrong.

The Most Common Misreading of the Patent

Most SEO articles about the information gain score describe it as "a measure of how much information your content has." That reading is wrong, and the error leads to bad strategy. The patent doesn't measure total information. It measures additional information relative to documents the user has already seen or could see.

The distinction matters because it changes what you optimize for. If the score measured total information, the strategy would be to write longer, more comprehensive articles that cover every angle. That's the skyscraper approach, and it fails the information gain test because comprehensive content that covers the same angles as every other comprehensive article adds nothing new. Two 5,000-word articles covering the same twenty subtopics of a keyword have identical information gain scores: low or zero. The articles are different documents with the same knowledge.

A 400-word article from a craft brewery explaining why they stopped using Cascade hops in their IPA after 2024 because the crop's alpha acid levels have been declining 0.3% per year since 2019, forcing a switch to Centennial for consistent bitterness, has higher information gain than a 5,000-word "Complete Guide to IPA Brewing" that covers the same hop varieties every other brewing guide covers. The craft brewery's article contains knowledge from their direct experience that the comprehensive guide doesn't. That's what the patent measures.

How the Scoring Works

The patent describes a three-step process. First, Google identifies a set of documents the user has already encountered on a topic (the "first set"). These might be pages the user clicked in a search session, or pages Google knows are commonly viewed for that query. Second, the system identifies a set of new documents that could be relevant (the "second set"). Third, for each document in the second set, the system calculates how much information it contains that isn't in the first set. That delta is the information gain score.

A nuance most articles about this patent miss: the second set isn't limited to documents ranking for the original query. The patent describes the system identifying documents relevant to the user's next information need. If someone searches "how to brew IPA at home" and reads three results, the system looks for documents that answer what the user will likely want to know next (water chemistry, fermentation temperature control, hop schedule timing) and scores those documents on information gain relative to what the user already consumed. This is why cluster architecture matters mechanically, not just strategically. Supporting articles that answer the next question in a user's journey benefit directly from information gain scoring because they serve a predicted need.

The calculation can use direct comparison or machine learning models trained to predict information gain based on document features. The patent specifically mentions "semantic feature vectors" and "bag-of-words representations" as inputs. This matters because it means Google doesn't compare words. It compares meaning. Two articles can use completely different vocabulary and still score identically on information gain if they convey the same knowledge. Rewriting existing content in your own words doesn't improve the score. Paraphrasing the same facts that ten other articles contain produces the same meaning, which produces the same score: zero. Jason Spencer demonstrates this to ROI.LIVE clients by showing them two articles about the same topic written by different agencies, with different word choices and different structures, that both scored zero on The Delta Audit because the underlying knowledge was identical.

The patent also states that the ML model can learn to predict information gain without needing the user's browsing history as input. That means Google can estimate information gain at indexing time, not just during a live search session. Jason Spencer at ROI.LIVE believes this is why information gain affects initial rankings, not just the re-ranking of results after a user clicks through multiple pages.

The patent also describes dynamic re-scoring. As a user views additional documents, the system recalculates scores for the remaining documents because the baseline of "what the user has already seen" has changed. A document that scored high before the user read a similar article might score lower afterward. This dynamic quality means information gain isn't a fixed property of a page. It's relative to the rest of the index for that topic at any given time. Your competitors publishing new content can change your page's score even if your page hasn't changed.

How to Approximate Your Score

Since Google doesn't publish information gain scores, ROI.LIVE built a practical proxy: The Delta Audit. The methodology mirrors the patent's logic. For each page, Jason Spencer reads the top 3 results for the target keyword (the "first set" in patent terms). Then reads the client's page (the "second set" candidate). Any paragraph in the client's page that contains knowledge not present in the top 3 gets marked as a unique element. The number of unique elements becomes the score.

High (3+ unique elements): The page contains multiple data points, stories, or perspectives not found in competing results. This page would likely score well on Google's internal measure.

Medium (1-2 unique elements): Some original content mixed with consensus material. Improvable by enriching the unique elements and rebuilding the generic sections.

Zero (nothing unique): The page restates what the top results say. On Google's internal measure, this page provides near-zero information gain regardless of its word count or technical SEO quality.

The proxy isn't perfect. Google's ML models evaluate semantic similarity at a level that human reading can't match precisely. But it's directionally accurate. Every page that ROI.LIVE has scored as High on The Delta Audit has either maintained or gained rankings during the March 2026 core update. Every page scored as Zero has either stagnated or dropped. The correlation is strong enough that Jason Spencer now requires a Delta Audit score of Medium or higher before any article publishes.

What This Means for Your Content

A Scoring Walkthrough

Jason Spencer walks ROI.LIVE clients through this exercise during onboarding. Search Google for your primary keyword. Open the top 3 results. Read them and note what knowledge they share in common. That shared knowledge is the baseline. The first set. Now read your page. Every paragraph that conveys the same knowledge as the baseline, mark red. Every paragraph that contains something the baseline doesn't, mark green. Count the green paragraphs. That count is your approximate information gain score.

For a craft brewery targeting "how to brew IPA," the top 3 results all cover: hop varieties, grain bill, yeast selection, fermentation basics, dry hopping technique. That's the baseline. A brewery article that covers those same five topics with better writing scores zero. A brewery article that adds "we switched from Cascade to Centennial in 2024 because the Cascade crop's alpha acid has declined 0.3% annually since 2019, and our house IPA's bitterness dropped from 65 IBU to 58 before we caught it" scores one green paragraph. If the article also includes fermentation temperature data from their own brew logs showing the 2-degree window where their yeast produces the ester profile they want, that's two green paragraphs. Medium score. Two unique elements that exist nowhere in the top 3.

Three Strategic Implications

First, stop optimizing for comprehensiveness. The patent rewards the delta, not the total. Adding more sections to an article that already covers the same ground as competitors doesn't improve the score. Adding one section with proprietary data does. A dog training company that adds a 500-word section about their specific desensitization protocol for leash reactivity, including the distance thresholds they've calibrated over 200 clients and the failure rate they've observed when trainers skip the baseline assessment, contributes more to the score than adding five generic sections about obedience basics.

Second, monitor competitors. Because the score is relative, a new competitor publishing strong original research on your topic can lower your page's effective score without you changing anything. ROI.LIVE runs quarterly re-audits for this reason. A page that scored High six months ago might score Medium today because a competitor published something that reduced the delta.

Third, the score compounds across a cluster. When multiple pages on the same site all score high on information gain for related topics, Google's system recognizes the site as a source that consistently contributes new knowledge. That recognition accelerates ranking for new pages in the cluster because the ML model has learned to predict high information gain from the site based on its track record. A photography studio with 8 articles that each contain unique shooting technique data, specific gear failure stories, and client session insights builds a pattern that Google's system learns to trust.

Questions About Information Gain Score