Complete Guide

Information Gain in SEO: What the Patent Says, and What to Do About It

Google holds a granted patent describing information gain, and almost every article about it has not read the claims. The score is computed against what a reader has already seen, not against your page, which means a fixed page-level score cannot exist and the value of content changes with where a reader is in a decision. What Google has confirmed, what it has not, and how to apply the principle without claiming a ranking factor.

25 min read 5,405 words Published Oct 2026
Information Gain SEO

The Claim, and the Part of It That Holds

Information Gain is having a moment. The version circulating on social media says Google is obsessed with it, that it is a ranking factor, and that the way to earn it is to add opinions, stories and personal anecdotes to your content.

One of those three things is well supported. The other two are not, and the gap between them matters, because acting on the overstated version produces content that sounds different without being more useful.

Google does hold a patent describing information gain. It is a real, granted, readable document, and most articles about it have not read past the title. When you do read the claims, the mechanism turns out to be considerably more interesting than the advice built on top of it, and it points somewhere the current discussion has not gone.

What the Patent Actually Says

US11354342B2, Contextual estimation of link information gain, is assigned to Google LLC. It was filed on 18 October 2018 and granted on 7 June 2022. The named inventors are Victor Carbune and Pedro Gonnet Anders.

The three independent claims describe the same method. A system identifies a first set of documents on a topic that the user has already viewed. It then identifies a second set of documents on the same topic that the user has not viewed. For each document in the second set it computes an information gain score, and uses those scores to select and rank what to show next.

The score itself is defined as a value indicative of additional information included in a document beyond the information contained in documents the user previously viewed. Claim 1 ties it to a quantity of new information in the new document that differs from the information extracted from each document of the first set. Dependent claims describe applying semantic feature vectors from both sets to a machine learning model whose output determines the score.

Read that definition twice, because everything useful follows from it.

The Patent Family Did Not Stop in 2022

Every article we have read on this subject treats the patent as a finished historical object, and they do not agree on its dates. One calls it “the 2020 Google patent application”. Another says it was filed in 2018 and granted in 2022. A third says filed in 2022 and granted in 2023. The first two are describing different documents in the same family. The third is simply wrong.

Here is the family as the record shows it. The priority date is 18 October 2018. The application published as US20200349181A1, which is the 2020 document. US11354342B2 was granted from it. And there is a third, which most of the articles teaching this subject have missed, though not all: Konrad Dziekan at OptimusPrime.pl identified it in July 2026, along with two further documents in the family.

US12326889B2, the same title, the same two inventors, Victor Carbune and Pedro Gonnet Anders, assigned to Google LLC. Filed 7 May 2024. Granted 10 June 2025.

That matters for a plain reason. A continuation is work. Google wrote this application in May 2024, which is after ChatGPT, after Bard, after AI Overviews began rolling out, and pursued it to grant in June 2025. A company that had lost interest in a mechanism does not keep paying attorneys to extend its claims on it.

The same audit traced US12013887B2, the application publication US20230342384A1 and the PCT document WO2020081082A1, which is the fullest public accounting of this family we have seen anywhere. Be careful about what any of it shows. It is not evidence that the mechanism runs in Google Search. Patents are filed defensively, speculatively, and to fence off territory a company may never build on. What it does establish, and what you can check in about a minute, is that this is a live family that Google was still prosecuting into the AI answers era, not a curiosity abandoned in 2022.

The Score Is Not a Property of Your Page

The information gain score in this patent is not calculated against the whole index, against the other results on the page, or against some notion of how original your writing is. It is calculated against what one specific reader has already read.

That has a consequence the rest of the discussion has skipped. Two people can land on the same URL and the mechanism described here would assign it different scores, because they arrived having read different things. A reader who has read nothing on the subject gets a great deal from a clear definition. A reader who has just read five competing explanations of the same thing gets nothing from a sixth.

So a fixed, page-level information gain score, in the sense this patent describes, is not a thing that can exist. The number depends on the reader, and the reader is different every time.

This does not mean that publishing original, useful information is a bad strategy. It is a very good one, and Google says so in its own documentation, which we come to below. What it means is narrower and more practical: you are not optimising a score on your page. You are increasing the probability that your page says something a given reader has not already been told. Those are different jobs, and only the second one is actually achievable.

Which Makes Information Gain Decision-Stage Dependent

If the measure is relative to prior reading, and prior reading accumulates as somebody researches a purchase, then the same content has different value at different points in a buying decision. This is the practical form of the idea, and it is the part we think is genuinely useful.

Consider somebody evaluating a software category. Their reading history grows at every stage, so what counts as new information shrinks and changes shape.

StageWhat they have read alreadyWhat now carries informationWhat now carries none
Understanding the problemLittle or nothingA clear definition, the distinction between adjacent terms, what the category actually solvesNothing yet. Everything is new.
Comparing optionsTwo or three overviewsWho the real alternatives are, how they genuinely differ, which are not suitable and whyAnother definition of the category
Establishing trustSeveral vendor pagesDocumented controls, named standards, evidence that can be checked independentlyClaims of being trusted or leading
Qualifying against requirementsMost of the vendor marketingExplicit statements of what the product does not do, where it is a poor fit, tested limitsFeature lists, which by now all read the same
Taking actionEnough to have a shortlistPricing structure, what the evaluation involves, what happens nextAny further persuasion

The column on the right is the one most content strategies ignore. A page can be genuinely excellent and carry no information at all for the reader who finds it, because everything in it was already said by the three pages they read first. That is not a quality problem. It is a position problem, and it is fixable by writing for a stage rather than for a keyword.

It also explains a pattern we see repeatedly in AI citation work: a business is cited constantly on educational questions and absent entirely from comparison and qualification questions. The educational content is doing its job. The problem is that nothing in the estate speaks to a reader who has already done the education.

In a Ranked List There Is a Position Four

There is a part of this patent that has almost nothing to do with ten blue links, and for anyone whose customers are starting to ask an assistant instead of a search box, it is the part that matters most.

The claims do not only cover documents. They extend to conversational agents and voice assistants, and in that setting the thing being de-duplicated is not the page. It is the fact.

The mechanism described is this. A first source supplies facts A and B, and those are conveyed to the user. A second source is then considered, containing fact B and a new fact C. Because B has already been conveyed, the system generates output that includes C and excludes B. The second source is used only for the part of it that was new. Its overlap is dropped before the user hears a word.

Sit with what that does to the economics of being ordinary.

In a ranked list, being the fourth page to say the same correct thing is survivable. You are lower down, your click-through is worse, but you are on the page and a share of people still choose you. Redundancy costs you position.

In a synthesised answer there is no position four. The answer is assembled from whichever sources contribute something, and a source whose every fact has already been taken from somewhere else contributes nothing to assemble. It is not ranked below the others. It is absent. Redundancy stops costing you position and starts costing you existence.

This is the honest commercial case for originality, and notice that it does not depend on the patent being live in Google Search at all. It follows from how any system that synthesises one answer from many sources has to behave. Anything that writes a single answer has to decide what not to repeat.

Where the Acronyms Actually Attach

Everything above describes mechanisms: retrieval, de-duplication of facts, synthesis into a single answer, attribution. The field has grown a vocabulary on top of those mechanisms, and the vocabulary is looser than the mechanisms are. It is worth mapping one onto the other, because the terms are not interchangeable and information gain does not mean the same thing to each.

From original information to a useful outcome, where AI sits in between
  1. Information gainWhat genuinely new or useful evidence does this content contribute?
  2. counts only where both of these also hold
    DiscoverabilityCrawlable, indexed, retrievable and relevant to the question asked
    CredibilitySource quality, provenance, identity and corroboration from elsewhere
  3. which together make possible, but never guarantee
    Use by an AI systemAn answer, a citation, a comparison or a recommendation, depending on the system
  4. and only where an interface exists to act through
    A useful actionA qualified enquiry, a trial, a booking or another outcome the business wanted

A conceptual model, not a guaranteed ranking, citation or conversion sequence. Discoverability and credibility are conditions that have to hold alongside original information. They are not produced by it.

With that model in mind, here is where each term attaches, and what it does not cover.

TermWhat information gain does for itWhat it does not settle
SEOOriginal evidence can make a page more useful and more distinct than the results around it.The patent is not evidence that a scoring mechanism of this kind runs in current rankings.
AEO, answer enginesA clearly expressed, distinct answer gives an answer engine something specific to extract.Being extractable does not mean being selected, and selection does not mean attribution.
GEO, generative enginesVerifiable findings give a generated answer something defensible to rest on.Published GEO findings depend heavily on which models, datasets and prompts were tested.
LLM optimisationConsistent, distinctive descriptions improve the material available for retrieval and grounding.Training, retrieval and citation are three different mechanisms and are routinely conflated.
AI visibilityOriginal evidence supports differentiation and accurate representation.Visibility also turns on access, identity, corroboration and each platform’s own behaviour.
Agentic websitesAccurate, specific business facts help an agent judge whether you suit the task.Acting requires interfaces, permissions, validation and transaction boundaries. No amount of originality substitutes for a form an agent can complete.

The distinction worth holding on to is that information gain is a property of content and evidence, while agent readiness is a property of systems and interfaces. A site can carry genuinely original research and be impossible to book through. A site can be booked through perfectly while saying nothing original at all. They are solved separately, and one does not buy the other.

What Google Has Actually Confirmed

Separating the documented from the inferred matters more than usual here, because almost every article on this subject blurs the two.

Documented: Google asks for original information

Google’s guidance on creating helpful, reliable, people-first content contains a self-assessment list. Three of its questions are directly on point, quoted as written:

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

Does the content provide insightful analysis or interesting information that is beyond the obvious?

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

The same document asks whether content clearly demonstrates first-hand expertise and a depth of knowledge, giving examples such as expertise from having used a product or visited a place. It also says Google strongly encourages accurate authorship information, such as bylines, where readers might expect them.

Note the third question in particular. Compared to other pages in search results is a relative test, not an absolute one. That is the same shape as the patent, arriving from the other direction.

Documented: original content systems exist

Google’s guide to its ranking systems describes systems intended to surface original reporting ahead of pages that merely cite it. That is a confirmed system, publicly described, and it is a stronger citation than any patent for the general proposition that originality is rewarded.

Not confirmed: that information gain is a live ranking factor

A granted patent is evidence that a company protected an approach. It is not evidence that the approach is deployed, nor that it carries weight in general organic ranking. Google has not confirmed that this mechanism operates in Search, and no public study establishes a ranking effect attributable to it.

Anyone attaching a percentage uplift to information gain is making it up. Demonstrating one would require a controlled comparison of otherwise equivalent content, holding topical relevance, existing authority, links, search demand and competing results constant. We have not seen that study, and we are not going to pretend a correlation would substitute for it.

Three Things People Are Getting Wrong

1. That there is a page-level information gain score to optimise

Covered above, and it is the root of the other two. The patent’s measure is contextual. A tool offering to score your page for information gain is scoring something else, whatever it has chosen to call it.

2. That adding personal anecdotes improves rankings

A story is not automatically information. “We found this surprising” tells a reader nothing they can use. “We ran this on 40 pages, 12 behaved differently, and here is the characteristic they shared” is information, and it happens to be a story.

The distinction is not between corporate prose and personal prose. It is between novelty of expression, which is saying a known thing in your own voice, and novelty of information, which is saying something a reader could not have got elsewhere. Only the second one survives the comparison the patent describes, and only the second one answers Google’s question about substantial value compared with other pages.

Voice matters for whether people finish reading. It is not the same lever.

3. That AI-written content is inherently penalised

This is not Google’s position and has not been since 8 February 2023, when it published guidance stating that it rewards high-quality content, however it is produced. The same guidance is clear that using automation, including AI, to generate content whose primary purpose is manipulating search rankings violates its spam policies.

The target is scaled, low-value content, not the tool used to write it. An article produced with AI assistance that contains a measurement nobody else has published can carry more information than an entirely human-written article that restates the first page of results. The production method is not the variable.

We Tried to Source Every Number in the Articles Teaching This

A page arguing for evidence ought to test the evidence around it, including when the result is awkward for us.

The best work on this has already been done, and not by us. In July 2026 Konrad Dziekan at OptimusPrime.pl audited 85 unique URLs from Polish and US results, retrieved the full content of 82, and recorded 207 claim occurrences against a dictionary of 40 clusters, publishing the dataset with a checksum. His central finding is the one worth repeating: 21 of those 82 pages recommended first-party data and 20 recommended original research, while he could detect first-party data on 2, a stated method on 18 and disclosed limitations on 6. Recommending evidence is roughly ten times more common than showing any. If you read one other thing on this subject, read that.

We did something narrower and complementary. His audit measured whether an evidence signal was present. Ours takes the specific named statistics being passed around in the articles ranking today and tries to follow each one back to the study it names. Run 15 November 2026, repeatable in an afternoon.

We want to be careful about the finding. Not traced means we could not find the source, not that the source does not exist and not that anybody invented anything. A study can be real and badly cited, behind a login, or renamed. But a statistic a reader cannot check is not evidence, whatever lies behind it, and three of these are load-bearing claims in an article about the importance of original data.

ClaimSource givenOur resultWhat we found
“Research from Semrush (2024) found that pages with original data outrank commodity content by an average of 3.2 positions”Semrush, 2024Not tracedNo Semrush study we could find reports this. The only page on the open web carrying the figure is the article making the claim.
“Analysis by Moz (2023) showed that pages featuring original research received 2.8x more citations in AI-generated answers”Moz, 2023Not tracedMoz has published real research on AI Mode citations, including that 91% of AI Mode responses draw on ten or more unique URLs. We could not find the 2.8x figure in it or anywhere else.
“HubSpot’s 2024 Content Strategy Report found that 67% of top-performing B2B content includes proprietary data, compared to just 18% of pages ranking below position 10”HubSpot, 2024Not tracedWe could not locate a HubSpot report containing these figures.
Proprietary assets predict which sites gained trafficCyrus Shepard, 400+ sitesTraced, but reported inconsistentlyThe study is real and the finding is strong. One article attributes it to the December 2025 core update, another to March 2026. Neither carries the caveats.
The patent datesVariousThree versionsFiled 2018 granted 2022, a 2020 application, and filed 2022 granted 2023. The first two describe different documents in one family. The third matches no document we can find.

Two things are worth drawing out of that table.

The first is that the strongest real finding in the category is being reported worse than it deserves. Cyrus Shepard’s analysis of more than 400 sites found that 92% of traffic winners had proprietary assets against 57% of losers, and that sites with none of the five features he measured won 13.5% of the time while sites with all five won 69.7%. That is a genuinely useful pair of numbers. It is also reported with Spearman correlations between 0.206 and 0.391, which is moderate, using third-party traffic estimates rather than Search Console data, and the study does not establish causation. The articles citing it carry the conclusion and drop every one of those qualifications. The caveats are the first thing to fall off a statistic as it travels.

The second is the shape of the thing. The subject is how to add information that can be checked. Several of the pages teaching it are resting on numbers that cannot be. We are not claiming bad faith, and we would rather be shown the sources and correct this section. But if you are deciding whose method to follow, the question to ask of any page here, including this one, is whether you can get to the underlying source from what is on the page.

How to Apply the Principle Honestly

Without claiming to manipulate a ranking factor nobody has confirmed exists.

The test to apply before publishing

Read the first five results for the query you are targeting. Then ask one question about your draft: could a reader get this from those five pages?

If the answer is yes, the piece is not finished, however well written it is. If the answer is no, identify the specific sentence or section that is not available in those five and make sure it survives editing. It is usually the first thing cut for length.

What actually counts as a contribution

  • A measurement you took, with the method stated, including the ones that went the wrong way
  • A limitation of a product or approach that its own documentation does not mention
  • A comparison on a dimension nobody else has compared on
  • A primary source read properly, where everyone else has cited a summary of it
  • A negative result, which almost nobody publishes and which saves a reader real time
  • Structure that makes an existing body of information usable, where the facts exist but are scattered

The last one is worth dwelling on, because it is the contribution most often dismissed. Assembling, reconciling and sourcing material that exists only in fragments is original work, even when no individual fact in it is new.

Make the contribution checkable

An unverifiable original claim is worth less than a verifiable one, to readers and to the systems that cite them. This is where our CITATE framework already does the work: a specific, verifiable data point, attributed to a named source. No new criteria are needed for information gain. Contribution is the what, and attribution is what turns it into something a reader, an editor or a model can rely on.

For the same reason, write the contribution so it survives extraction. A finding buried in the ninth paragraph of a narrative is harder to retrieve than one stated plainly under a heading that matches the question it answers. The entity and structure work that makes a page extractable is what makes its original information reachable.

Where Information Gain Sits in the Order of Work

Information gain is not a lever you pull. It is a property of being worth citing, and it has prerequisites that almost every article on the subject skips, because the prerequisites are less interesting than the idea.

The order matters more than the idea does.

OrderWhat it meansWhy it comes first
1. RetrievableCrawlable, indexable, renders without JavaScript the crawler will not run, loads fast enough to be fetched.A contribution on a page that cannot be retrieved has not been made.
2. RecognisedThe organisation exists as an entity, corroborated by sources that are not you.A novel claim from an unrecognised source reads as noise rather than insight. Corroboration is what turns one into the other.
3. CompleteYou cover what every page on the subject covers.This is the cost of entry, not the differentiator. You cannot usefully add to a conversation you have not joined.
4. ContributingYou add the measurement, the limitation, the negative result, the thing only you know.Only now does it pay, because only now is there something to attach it to.

Here is the part that will not be popular, and it is the honest answer to the question most people are actually asking when they ask about information gain.

For most sites, information gain is the last thing to work on, not the first. Commissioning original research for a page that a crawler renders as an empty shell, published by an organisation that search engines do not yet recognise as existing, on a topic the site has not otherwise covered, is money spent at the wrong end of the problem. The research will be good. It will not be found, and if it is found it will not be trusted, and if it is trusted it will have nothing around it to make it credible.

That is the opposite of what the rest of this category will tell you, so weigh it accordingly. But it follows directly from the mechanism. A score computed against what a reader has already seen can only be computed for a document the system has retrieved and is willing to use. Everything upstream of that is not optional, and it is usually cheaper.

The sensible exception is a site that already has the first three in place and is stuck. That site has a genuine information gain problem, and for that site this is the highest-leverage work available.

Where the Contribution Comes From When You Do Not Have a Dataset

The examples held up in this field are a home services platform with a database of millions of real project costs, a video company with ninety million uploaded videos to analyse, and a recipe writer who spent twenty nine hours on a single curry. All three are real and all three are useless as a model if you are a six person firm with a full diary.

Nobody writing about information gain seems to address the ordinary case, which is the one most businesses are in: no research budget, no proprietary dataset, no content team. That case is not hopeless. It is just that the evidence is sitting in the business rather than in a data warehouse, and it has never been written down.

What you already holdWhat it becomesWhat to check before publishing
The quotes you have issued for one job typeA real price range with the method stated, rather than a guess copied from a national averageAggregate it, never identify a client, and say how many jobs and over what period
The questions customers actually ask, in their wordsHeadings that match how people phrase it, and answers nobody else thought worth writingThat you are describing a pattern you have seen, not one example
The jobs that went wrong and what caused itThe failure modes and limitations missing from every competitor pageAnonymised, and fair to the people involved
The work you turned down and whyAn honest account of who a service is not for, which is rare and disproportionately trustedThat it reads as judgement, not as a dig at a competitor
Photographs of the actual work, before and duringFirst-hand evidence that the work was done by the people claiming itPermission, and that nothing identifying is in shot
What changed after you did the work, measuredA result with a method attachedThat the figure is yours to publish and that you can show how it was measured

None of that requires a study, a survey panel or a budget. It requires someone to write down what the business already knows and is currently throwing away, and it is unavailable to every competitor, because they did not do your work.

One caution, and it is the one that matters most. An invented figure is worse than no figure. A made-up percentage, an anonymous case study that cannot be checked, or a range that sounds authoritative and came from nowhere does more damage than a page that honestly says it does not know. If you have not measured it, do not publish a number for it. Describe what you have seen instead, say how many times you have seen it, and leave it there.

Gain Gets You Considered. Being Citable Gets You Used.

Everything above is about whether what you added was worth adding. That is the right first question and the published work on it is good. It is not the last question, because a contribution that cannot be found, trusted or lifted out of the page does not get used, however original it is.

There is evidence for that gap sitting on this very query, and it is uncomfortable for everyone writing in this space, ourselves included.

On 15 November 2026 the AI Overview for information gain seo strategy drew on Ahrefs and Search Engine Land. The page with the 85 URL audit, the published dataset and the SHA-256 checksum was on the same results page and was not among the cited sources. The most rigorous research in the result set was not the research being used. We are reading one snapshot and these features vary by session and by day, so treat it as an observation rather than a measurement. But it points the same way as the order of work above: originality that is not yet recognised does not get picked up.

So the sequence is two steps, not one. Information gain decides whether you have something to contribute. Citability decides whether a system assembling an answer can take it. They fail independently. A page can be genuinely novel and completely unusable, because the finding is buried in paragraph forty, phrased so it cannot stand on its own, and attributed to nobody.

CITATE is our framework for the second step. Six criteria, all of them about whether a claim can survive being taken out of the page it lives on.

CriterionWhat it asks of a passage
C1 Standalone opening answerDoes the section open with an answer that still makes sense with everything around it removed?
C2 Explicit definitionIs the thing being discussed actually defined, rather than assumed?
C3 Statistic with contextDoes every figure carry its numerator, denominator, date and method?
C4 Named sourceIs the source named and reachable, rather than implied?
C5 Named entityAre the people, organisations and documents named, so the claim attaches to something real?
C6 Attributable claimIf a system quoted this sentence, would the attribution be unambiguous?

Hold this to the same standard we held everyone else to. CITATE is an editorial framework of ours, not a Google metric, and we have not validated it against rankings or citation rates. It does not measure information gain and it is not a scoring model for novelty, which is a different job done well elsewhere. What it does is force the question that the gain discussion skips: having written something worth taking, have you made it takeable?

The two tests also fail in opposite directions, which is why running only one of them is a mistake. Pass novelty and fail citability and nobody can use what you found. Pass citability and fail novelty and you have produced a beautifully structured, perfectly attributed restatement of what the reader already knew, which is the commodity content this entire subject exists to argue against.

Where This Leaves the Original Claim

Google values original information. That is documented, repeatedly, in its own words, and it has publicly described ranking systems intended to favour original reporting. Building a content strategy on it is sound.

Information gain, as the patent defines it, is a contextual measure of what a document adds relative to what a reader has already seen. It is a useful way to think and a poor thing to claim you can optimise, because the number is not yours to control.

The practical version is simpler than either. Work out what your reader already knows, find the thing they cannot get from the results already ranking, say it plainly, and show where it came from.

Evidence ledger, last reviewed 15 November 2026
What is documented
US11354342B2, Contextual estimation of link information gain, assigned to Google LLC, filed 18 October 2018, granted 7 June 2022, inventors Victor Carbune and Pedro Gonnet Anders. Its claims compute an information gain score for unseen documents relative to documents the user has already viewed. Google’s helpful content guidance asks whether content provides original information, reporting, research or analysis, insightful analysis beyond the obvious, and substantial value compared to other pages in search results. Google’s ranking systems guide describes systems intended to surface original reporting. Google stated on 8 February 2023 that it rewards high-quality content however it is produced, while treating automation used primarily to manipulate rankings as a spam policy violation. US12326889B2, the same title and the same inventors, assigned to Google LLC, filed 7 May 2024 and granted 10 June 2025, which shows the family was still being prosecuted after AI Overviews began. Cyrus Shepard’s analysis of more than 400 sites reports proprietary assets present on 92% of traffic winners against 57% of losers, with Spearman correlations between 0.206 and 0.391, using third-party traffic estimates rather than Search Console data. Our own source audit of the articles teaching information gain, run 15 November 2026 and repeatable.
What we infer
That because the patented measure is computed against a reader’s prior reading, no fixed page-level information gain score can exist in that sense, and that the informational value of a page therefore varies with where a reader is in a decision. This follows from the claim language but Google has not described it this way.
What is not established
That the mechanism in this patent operates in Google Search today, that information gain is a ranking factor in general organic results, or that any percentage uplift can be attributed to it. We have found no controlled study isolating it, and we are not aware of any public Google confirmation.

Frequently Asked Questions

Is Information Gain a confirmed Google ranking factor?

No. Google holds a granted patent describing an information gain mechanism, US11354342B2, but a patent shows that a company protected an approach rather than that it is deployed. Google has not confirmed that this mechanism operates in Search, and no public study has isolated a ranking effect attributable to it. What Google has confirmed is a preference for original information, stated in its helpful content guidance, and the existence of ranking systems intended to surface original reporting.

Can a page have an Information Gain score?

Not in the sense the patent describes. The score there is computed for a document relative to the documents a particular user has already viewed, so the same page would score differently for different readers. A tool that assigns your page a single information gain number is measuring something else. The useful version of the idea is to increase the chance that your page says something a given reader has not already been told.

Does adding personal stories and opinions improve rankings?

Not automatically. The distinction that matters is between novelty of expression, which is saying a known thing in your own voice, and novelty of information, which is saying something a reader could not have obtained elsewhere. A story that carries a measurement, a limitation or a negative result is information. A story that carries only tone is not. Google’s own question is whether content provides substantial value compared to other pages in search results, which is a test of what is said rather than how it sounds.

Is AI-written content penalised by Google?

No. Google stated on 8 February 2023 that it rewards high-quality content however it is produced. Its spam policies target automation used primarily to manipulate search rankings, and scaled low-value content, not the use of AI as a writing tool. An AI-assisted article containing original measurement can be more valuable than a human-written article that restates existing results.

Should information gain be my first priority?

For most sites, no. A contribution only counts once the page can be retrieved, the organisation is recognised by sources other than itself, and the subject is otherwise covered. Original research published on a page a crawler cannot render, by an entity search engines do not recognise, is money spent at the wrong end of the problem. The exception is a site that already has those three in place and is stuck, where this is the highest-leverage work available.

What if I have no data of my own to publish?

Most businesses have more first-hand evidence than they think, it is simply unwritten: the real spread of quotes you have issued, the questions customers actually ask in their own words, the jobs that went wrong and why, the work you turned down, what you measured after doing the work. None of that needs a research budget and none of it is available to a competitor. The one rule is that an invented figure is worse than no figure. If you have not measured it, describe what you have seen and say how often, rather than publishing a number.

How do I know whether my content adds anything?

Read the first five results for the query you are targeting, then ask whether a reader could get your draft from those five pages. If they could, the piece is not finished. If they could not, identify the specific section that is unavailable in those results and make sure it survives editing, because it is usually the first thing cut for length.

Sean Mullins

Founder of SEO Strategy Ltd with 20+ years in SEO, web development and digital marketing. Specialising in healthcare IT, legal services and SaaS — from technical audits to AI-assisted development.

Ready to improve your search visibility?

Book a free 30-minute consultation and let's discuss your SEO strategy.

Get in Touch