The SEO industry recommends increasing information gain through first-party data, expert contributions and case studies. The problem is that it often shows neither the data, nor the method, nor the limitations.
In my audit, I retrieved the full content of 82 of the 85 unique URLs visible in Polish and US search results. I recorded 207 claim-occurrence records: no more than one occurrence of a given claim per page within a cluster. The records belong to 31 clusters present in the corpus; the dictionary contains 40 clusters. No occurrence was found in the remaining 9 clusters. Only 2 of the 82 pages contained recognisable signals of first-party data, 18 described a method and 6 disclosed limitations. At the same time, 21 pages recommended using first-party data and 20 recommended conducting original research.
That difference shows what is missing from the discussion of information gain: not another list of ideas for “unique content”, but a way to check whether a new element is:
- aligned with the user’s intent;
- genuinely new in the relevant context;
- supported by evidence;
- useful in making a decision;
- easy to find and verify.
I am publishing public research pack v1, containing the methodology, aggregated results, claim clusters, source verification, charts and the RIG INDEX worksheet. The package integrity can be checked with its SHA-256 checksum. Full records supplied by the external provider are not part of public release v1. RIG INDEX is neither a Google metric nor a measure of indexation status. It is an explicit editorial tool intended to stop me confusing new information with good information.
Public research pack v1 ZIP checksum
What does information gain mean in this study?
The term appears in several different senses. Combining them into one definition leads to conclusions that go too far.
| Layer | What it means | What it can be used for | What it does not prove |
|---|---|---|---|
| Information theory and machine learning | A reduction in uncertainty, or the usefulness of a feature when splitting data | Data analysis, model building and feature selection | It does not, by itself, describe the quality of an SEO article |
| “Contextual estimation of link information gain” patent | An estimate of a document’s information gain relative to information from documents already shown to or reviewed by the user | Understanding the described mechanism and its contextual nature | It does not confirm deployment as a standalone organic-ranking factor |
| Content SEO practice | New and useful value relative to available content | Planning research, examples, tools and better answers | It does not provide access to a hidden Google score or guarantee better rankings |
In the rest of this article I use the third meaning, while keeping it in check with the first two. In editorial practice, information gain is additional, reliable and useful information that the reader would not obtain from comparable material. “Reliable” is deliberate: a false novelty can also be new, but it does not improve the answer.
What does the “Contextual estimation of link information gain” patent actually say?
The most recent verified active US grant in this family is US12326889B2, published on 10 June 2025. The review also covered the earlier active patent US12013887B2, the application publication US20230342384A1 and the PCT document WO2020081082A1. The PCT document is marked as ceased, but remains a source for the family’s description. Patent status alone does not confirm that the mechanism is deployed in Google Search.
US12326889B2 US12013887B2 US20230342384A1 WO2020081082A1
The reference point is information already presented to the user
The gain described in the patent is contextual. A candidate document may be assessed relative to information from documents that have already been shown to or reviewed by the user. This is narrower, and more interesting, than the popular recommendation to “write something that is not in the top 10”.
Why? Because the set of competing results is only an editorial approximation. The patent does not establish a rule under which every page is compared solely with the current organic top 10. The context may depend on what has already been presented to a particular person in the described interaction.
The document also discusses automated assistants and conversational presentation of results. That justifies investigating a relationship with assistant systems. It is not enough to attribute the mechanism to today’s AI Overviews or AI Mode.
An example scale is not a public SEO metric
The patent family includes a scale from 0.00 to 1.00 as an implementation example. It does not follow that Google exposes such a score to site owners, uses exactly that scale in Search or that an external calculator can reproduce it.
I can build my own text-similarity measure, count new entities or use embeddings. That would be a researcher’s model, not a reading of an internal Google score.
What the source says and what the industry adds
| Claim | What the source supports | Assessment |
|---|---|---|
| Gain is calculated relative to information from documents already presented or reviewed | The patent describes that context | Supported |
| The mechanism can reduce repeated information | The patent lists reducing repeated or redundant information as a benefit | Supported |
| The patent also concerns automated assistants | Assistants and conversational presentation appear in the description and claims | Supported |
| Google compares an article solely with the current top 10 | The patent imposes no such restriction | Unsupported |
| The number of unique entities is a direct Information Gain Score | The source defines no such measure | Unsupported |
| The patent confirms an active Google ranking factor | Patent publication describes an invention, not its production deployment | Unsupported |
| This mechanism selects sources for AI Overviews | A relationship is a possible hypothesis, but is not confirmed by the reviewed sources | Hypothesis |
This separation does not make industry interpretations useless. It means only that they should be labelled as interpretations.
What do Polish and English-language search results repeat?
The study covered 18 query–market combinations on desktop and four mobile checks. I recorded 349 SERP observations, including 196 organic results. After normalisation, the corpus contained 85 unique URLs from 72 domains. Full content was retrieved for 82 pages; three returned HTTP 402 or 403 errors.
This is not a sample of the entire internet. It is a dated view of results returned on 18 July 2026 for a defined set of Polish and English-language queries.
Recommendations are more common than demonstrations
The most frequently recommended ways of creating new value were:
| Recommendation | Number of pages | Share of 82 retrieved pages |
|---|---|---|
| First-party or proprietary data | 21 | 25.6% |
| Original research | 20 | 24.4% |
| Subject-matter expert contribution | 16 | 19.5% |
| A new perspective or analysis | 16 | 19.5% |
| Specific examples and useful tools | 16 | 19.5% |
| Case study | 15 | 18.3% |
| First-hand experience | 7 | 8.5% |
Recommendations are more common than demonstrations. Source: OptimusPrime.pl audit, 18 July 2026; numerator depends on the category, denominator: 82 retrieved pages.
The comparison with execution looks different: I detected first-party data on 2 pages (2.4%), a method on 18 (22.0%) and limitations on 6 (7.3%). A qualifying source appeared on 44 pages (53.7%), and a primary source on 26 (31.7%). On 38 pages, or 46.3% of the corpus, I found no qualifying evidence signal under the adopted rules.
In this audit, a qualifying source meant a link to a primary source, a Sources, References or equivalent section, or an explicit attribution such as “according to” combined with an external non-social link. This is an audit heuristic, not a judgement of whether the whole page is true. Primary sources included patents, official Google documentation and specified publications and institutional domains.
Evidence coverage in the corpus. “Primary source” is a subset of “qualifying source”. Source: OptimusPrime.pl audit, 18 July 2026; denominator: 82 retrieved pages.
This does not automatically mean that the remaining articles are false. It means that the reader has more difficulty establishing where the information came from and where fact ends and the author’s interpretation begins.
Poland and the US: differences exist, but should not be overstated
The direct language comparison included 34 Polish pages and 45 English-language pages. Three pages marked as Swedish or of undetermined language remained in the overall corpus.
In the Polish set, 22 of 34 pages (64.7%) had qualifying sources and 9 of 34 (26.5%) had primary sources. In the English-language set, the corresponding figures were 22 of 45 (48.9%) and 17 of 45 (37.8%). The English-language corpus contained more tools and marketing measurement formulas. Polish results more often mixed local articles with English-language leaders.
This is an observation about the collected corpus, not a claim about the quality of the entire Polish or English-language internet.
Mobile does not create one permanent picture
For some queries, desktop and mobile results were almost identical. For the Polish query “information gain w SEO”, however, only 5 URLs overlapped in the union of 13 URLs across both devices. That variation is a reminder that a desktop-only audit is a snapshot, not a universal list of competitors.
New, true, useful — or merely different?
The simplest test of a new passage does not begin with “does a competitor already have this?”. It begins with four separate questions.
| Test | Control question | Example failure |
|---|---|---|
| Novelty | Does the element add anything beyond information already available to the reader? | A paraphrase of popular advice in different words |
| Truth | Does the claim have a source appropriate to its strength? | A patent presented as proof of an operating algorithm |
| Relevance | Does the information help fulfil the intent? | An interesting statistic unrelated to the user’s decision |
| Usefulness | Does the reader know what to do or how to change their assessment after reading it? | A list of entities with no explanation of their significance |
A content gap passes only the first test, and not always. A missing topic may be unnecessary, false or too remote from the intent. Likewise, E-E-A-T, completeness, original phrasing and entity counts can support quality, but they are not synonyms for information gain.
“Negative information gain”: an editorial term for novelty that makes an article worse
In the analysed corpus, an explicit warning that novelty does not guarantee truth or usefulness appeared on only 2 pages. Yet this is where the greatest risk arises.
In this article, “negative information gain” is an editorial term, not a formal information-theory measure. I use it for a new element that:
- introduces false certainty, for example by giving an invented “Google score” to two decimal places;
- extends the scope of a source, for example by turning a patent description into certainty about ranking deployment;
- only appears current because it has a fresh date while repeating old data;
- adds detail that is irrelevant to the user’s decision;
- increases cognitive load without changing the answer;
- hides the limitations of original research and therefore implies broader representativeness.
SEO makes it easy to reward difference because difference can be compared quickly. Truth and usefulness are much harder to check. That is why an evidence gate is needed.
RIG INDEX: an explicit model for assessing real value
RIG, or Reliable Information Gain, is an original editorial framework. INDEX denotes five assessment dimensions, and RIG INDEX is both the result and the name of the model. RIG INDEX v1.1 is used to compare proposed elements of an article.
RIG INDEX does not measure indexation status or internal search-engine signals. It assesses content against an explicitly defined corpus, methodology and five editorial dimensions.
Each candidate receives between 0 and 3 points in five dimensions, checked in a natural sequence: is it aligned with the intent, new, useful, evidenced and easy to use?
- I — intent fit: alignment with the user’s intent;
- N — novelty: newness relative to the appropriate comparison corpus;
- D — decision utility: the effect on the reader’s decision or action;
- E — evidence: the quality and suitability of the evidence;
- X — eXtractability: how easy the information is to find, understand and verify.
The base score is calculated as follows:
RIG INDEX_base = 100 × (I × N × D × E × X) / 243
Multiplication is deliberate. Evidence is a critical gate: E=0 reduces the base score to zero. A new claim with no evidence should not be rescued by a well-presented table.
Explicit penalties are deducted from the score:
| Error | Penalty |
|---|---|
| Unsupported factual claim | −20 |
| Patent presented as proof of deployment | −15 |
| No date for time-dependent data | −10 |
| Novelty unrelated to the intent | −10 |
| False precision | −10 |
| No limitations disclosed for original research | −10 |
RIG INDEX = max(0, RIG INDEX_base − P)
P is the sum of explicit penalties. The final score cannot fall below zero. Its value does not lie in a magic publication threshold. It makes visible why an editor accepts, revises or rejects an element.
The model has material limitations. The 0–3 scale and penalties are editorial decisions, not parameters supplied by Google. RIG INDEX has not been validated against rankings or search visibility. The result depends on the assessor’s judgement, so important decisions should be reviewed by a second reviewer. OptimusPrime.pl assessing its own material is not independent validation of the model.
RIG INDEX assesses not whether content is merely different, but whether it adds new, useful, documented and readily usable information relative to a defined corpus.
Example candidate assessment
The idea “publish a dated Polish–English dataset, method and source files” received 3 points in every dimension and a score of 100. The idea “build a calculator for a hidden Google Information Gain Score” lacks appropriate evidence and also pretends to precision and ownership of someone else’s metric: score 0, decision: reject.
An element scored I=3, N=3, D=3, E=2, X=3 receives 66.7 points before penalties. That is a signal to strengthen the sources, not to write a more forceful headline.
The complete candidate assessment register and RIG INDEX worksheet are included in public research pack v1.
A before-and-after example: what I removed from this article
The passage below comes from a working draft prepared for this article. It is neither a quotation from a competitor nor a test of ranking impact.
Before
Google compares an article with the top 10 and assigns it an Information Gain Score. To increase that score, add more unique entities, first-party data and an expert opinion. This gives the content a better chance of ranking highly and being cited in an AI Overview.
The passage is short and sounds specific, but combines at least five different claims. The two most important — comparison solely with the top 10 and the availability of a specific score — have no support in the reviewed sources. The relationship with AI Overviews is presented as an effect, although the available evidence permits, at most, a hypothesis.
RIG INDEX assessment: I=2, N=1, D=1, E=0, X=2. Base score: 0. Further penalties would apply for an unsupported claim, a patent used as proof of deployment and false precision.
After
In my audit, 21 of 82 pages recommended first-party data, but I found recognisable signals of such data on only 2 pages. The patent describes gain relative to information from documents already presented to or reviewed by the user; it neither restricts comparison to the organic top 10 nor confirms a public SEO score. Before adding data, therefore, state its date, method, limitations and the decision it is intended to support.
The revised version adds an observation from the original corpus, gives both numerator and denominator, narrows the interpretation of the patent and ends with a practical rule. Every claim can be checked in the data or a primary source.
RIG INDEX assessment: I=3, N=3, D=3, E=3, X=3. Score: 100 before any penalties.
This is an example of an editorial decision, not evidence of ranking impact. Its value can be assessed after publication through user behaviour, query visibility and use of the published materials.
Workflow: from a search result to an evidence register
Competitors’ headings are a useful starting point, but a poor final unit of analysis. The unit should be the claim.
- Define the question and corpus. Record the queries, market, language, device, date and result depth. Without this, “the competition” becomes an arbitrarily selected list.
- Retain the raw results. Do not rely solely on memory or the current SERP preview.
- Extract atomic claims. One sentence can contain a definition, a claim about a mechanism and a promise of an outcome. Separate them.
- Normalise repetitions. Paraphrases of the same claim should enter one cluster.
- Assign a type. Mark an observed fact, an external provider’s estimate, a derived result, an interpretation and a hypothesis.
- Verify the source and its scope. Ask not only “is there a link?” but “does the link support precisely this claim?”.
- Identify the decision gap. Look for information that changes the reader’s action, not merely a missing keyword.
- Score the idea in RIG INDEX. A low score indicates whether novelty, usefulness, evidence or a clear form is missing.
- Record the editorial decision. Add, revise, hold for further research or remove, and retain the reasoning.
- Disclose the method and limitations. Original research without these elements can look like strong evidence when it is not.
This register takes more work than analysing headings, but it makes correction possible. When a source changes or better evidence appears, the affected claims can be identified.
Does information gain help with AI Overviews and AI Mode?
It may improve the value of the material for the user, but there is no basis for promising a citation. Google’s guide to optimisation for generative Search features, updated on 10 July 2026, recommends useful, reliable, non-commodity content with its own point of view, experience or knowledge beyond an easy summary of what already exists. This is qualitative guidance, not a disclosed ranking factor or a formula for information gain.
Google describes retrieval-augmented generation (RAG), which retrieves current pages from the Search index and uses their information to build answers with supporting links, and query fan-out, which runs related queries in parallel to collect additional results for subtopics. The AI features documentation states that AI Overviews and AI Mode may use different models and techniques, and that their answers and links will vary. None of this shows that the mechanism described in the patent family selects sources for these features.
Google also states explicitly that third-party tools do not have access to Google’s internal ranking or AI systems. Their indicators may support a workflow, but do not reveal a hidden metric. Visibility can be monitored in the Search Console generative AI performance report, being rolled out to some sites; it includes, among other things, URL impressions in AI Overviews and AI Mode.
In the dataset used for this study, DataForSEO recorded the presence of an AI Overview element, but the interface used did not return the answer text or a source list. I could not therefore check whether domains cited by AI differed from organic results. The absence of those data is not evidence that the feature or its effect is absent.
The practical conclusion remains cautious: clearly separated claims, appropriate evidence and a readable structure serve the reader and make verification easier. They do not guarantee selection by a generative system.
Google guide to generative Search optimisation AI features and your website Search Console generative AI performance reports
Checklist before adding a “unique” element
- Does it answer the question the user actually brings?
- Against which corpus is novelty being assessed, and when was that corpus collected?
- Is it a fact, derived result, interpretation, hypothesis or recommendation?
- Does the source support the full strength of the claim?
- Are both numerator and denominator provided rather than a percentage alone?
- Do time-dependent data have a date?
- Does original research describe its method and limitations?
- Which reader decision will this element change?
- Can it be found and verified quickly?
- What happens to the article if the element is removed? If the answer is “nothing”, it probably adds no value.
What this article does not claim
- It does not claim that the patent confirms an active organic-ranking algorithm.
- It does not claim that RIG INDEX is a Google metric or a measure of indexation status.
- It does not promise better rankings or citation by AI.
- It does not equate a content gap, entities, E-E-A-T, linguistic originality or completeness with information gain.
- It does not present estimated keyword volume as actual traffic.
- It does not generalise the results from the 82 retrieved and analysed pages to the whole internet.
FAQ
What is information gain in SEO?
In editorial practice, it is additional, reliable and useful information that the reader would not obtain from comparable material. A different form or a new word is not enough. The information must fit the intent and have appropriate support.
Is information gain a confirmed ranking factor?
I found no confirmation in the reviewed official Google documents of a standalone ranking factor by that name. Google recommends content containing original information, reporting, research and analysis, and delivering substantial additional value. That supports a qualitative direction, but discloses neither a separate score nor its weight.
What does the Google patent actually describe?
It describes possible ways to estimate a document’s information gain in the context of information from documents previously presented to or reviewed by the user. The contexts include automated assistants. It is not an announcement that the mechanism is deployed in organic Search.
How can information gain be measured without pretending it is a Google metric?
Intent fit, novelty, evidence, decision utility and clarity can be assessed explicitly. The comparison corpus, date and rules must be stated. The result should be described as an editorial model, not a Google score.
Should information gain be high or low?
In ordinary language, more useful new value is beneficial. Difference itself should not be maximised. New information that is false or irrelevant makes the material worse.
How does a content gap differ from information gain?
A content gap identifies a topic or element absent from the compared material. Information gain also requires checking whether the missing element is relevant, true and useful. Every valuable gain may fill a gap, but not every gap is worth filling.
Does first-party data always improve content?
No. Data without a date, method, suitable denominator and limitations may only create an appearance of credibility. They should answer a specific user question and allow the conclusion to be checked.
Does information gain help with AI Overviews and AI Mode?
No simple causal relationship has been confirmed. Original, well-documented and readable material can be valuable to people and machines, but does not guarantee citation in AI features.
When does new information make an article worse?
When it is false, irrelevant, lacks context, presents a hypothesis as fact or increases length without changing the answer. This is what this study calls negative information gain.
Method, data and limitations
I conducted the audit on 18 July 2026. It covered 18 query–market combinations on desktop and four mobile checks for Poland and the United States. Results were retrieved to a top-10 depth. Of 85 unique organic URLs, I analysed the content of 82. Each record in claims.csv denotes a detected occurrence of a canonical claim on a page; I recorded the same claim no more than once per page. Records were grouped into clusters, and key claims were verified against patents and official Google documentation.
The methodology, limitations, aggregated results, claim clusters, source verification, publication gaps, candidate assessments and the RIG INDEX worksheet are in public research pack v1. Its contents are described by the public manifest and its integrity is confirmed by the SHA-256 file. Full records supplied by the external provider are not part of public release v1.
Public research pack v1 Public manifest Checksum file
The principal primary sources are US12326889B2, US12013887B2, US20230342384A1, WO2020081082A1, Google’s Creating helpful, reliable, people-first content and AI features and your website.
US12326889B2 US12013887B2 US20230342384A1 WO2020081082A1 Creating helpful, reliable, people-first content AI features and your website
A similar process for a website can be combined with an SEO audit, ongoing SEO work and a broader information-resilience approach for the age of AI.
SEO audit SEO services Information resilience in the age of AI