The information gain patent is a warning shot

Google's "contextual estimation of link information gain" patent has been sitting in public view since 2017. It was cited again last year. Its US protection was quietly extended to 2039. And most content teams I speak to have never heard of it, or if they have, they've filed it under "interesting but not actionable" — which is the file cabinet where SEO strategy goes to die.

I want to talk about this patent today because Search Engine Journal ran a good breakdown of it this morning, and the implications land differently in mid-2026 than they did in 2017. Back then, "originality" was a soft ranking factor most of us treated as directional. Today, with the commodity content tier being flattened by AI Overviews, image generation baked into the SERP, and ChatGPT vacuuming up informational queries, originality isn't a nice-to-have. It's the thing that determines whether your content gets into the next set of results at all.

The patent describes an information gain score — potentially framed on a 0 to 1 scale — assigned to a document based on how much new information it adds beyond what a user has already seen on the same topic. That's not about ranking the current SERP. It's about ranking the *next* one, based on what a user has already read and what value your document adds on top.

If that sounds abstract, stay with me. Because when you connect this to what's actually happening on the SERP right now, the picture is uncomfortable.

What the patent actually describes

The mechanism is more interesting than the headline. Google isn't just checking whether your document is "unique" against some static corpus. It's modelling a user's journey through a topic — they've read about growing an apple tree, they're now looking for what to feed one — and scoring each candidate document by how much *new* it contributes to that journey.

If your content only tells the user what the AI Overview already told them, why would Google send them to you next?

Documents that add little novel information can be reranked, demoted, excluded, or dropped from results entirely. That last one is worth reading twice. Not demoted. Dropped.

This connects to two systems we already know Google has, both surfaced through public documentation and the leaked Content Warehouse: `OriginalContentScore` and `ContentEffort`. Whether the information gain score is directly deployed in a re-ranker, or whether it operates through engagement proxies, or through some Bayesian predictive layer nobody outside Mountain View has seen — none of us can say with certainty. But the direction of travel is obvious.

Originality is being measured. It has been for years. It's now being measured against a much larger corpus, with better tooling, and with an AI Overview sitting on top of the SERP that itself synthesises the "obvious" answer before the user ever scrolls to your link.

If your content only tells the user what the AI Overview already told them, why would Google send them to you next?

The commodity content problem just got harder

Here's the shift most content teams have missed. In 2020, "commodity content" meant thin listicles or shallow how-tos. You could still rank a competent 1,500-word article on "how to prune an apple tree" if the domain was strong enough and the content was decent.

In 2026, the AI Overview is the commodity content. It's synthesised on demand, in the reader's context, referencing multiple sources, and it appears above your link. If your article is essentially the same information rewritten in your brand voice — congratulations, you're competing with something Google generated for free, in real time, tailored to the query.

The information gain patent describes the mechanism for what's already happening intuitively. Google is asking: what does this document add beyond what the user has already seen? And "what the user has already seen" now includes the AI Overview itself. That's a much higher bar than clearing a set of ten blue links.

I see this constantly with clients whose content programmes were built between 2018 and 2022. The playbook was: identify the top-ranking articles, produce something slightly longer, slightly better structured, with slightly more comprehensive coverage. It worked. It doesn't anymore. Not because the tactics were wrong — they were correct for the market they were designed for — but because the market has moved. The "10x content" doctrine has collapsed into the AI Overview.

What "10% different" actually means

The Search Engine Journal piece flagged something I want to sit with. In their read of the patent, as little as a 10% difference could be the delineator between marketing success and failure. That number isn't in the patent itself — it's an interpretation — but the shape of the claim is right.

Originality doesn't require rewriting the entire information space. It requires a genuine additive contribution. That could be:

• A first-hand data point nobody else has
• A synthesis that connects two things nobody else has connected
• A concrete example from real work
• A counterargument the consensus content is avoiding
• A dataset, benchmark, or measurement
• A specific claim with specific reasoning that a competent reader couldn't get from the AI Overview

What it can't be, in mid-2026, is a rewording of the top-ranking article with better headings. That was the playbook. That playbook is now the thing being demoted.

Where this collides with the measurement blind spot

Here's where it gets genuinely difficult, and where I part company with anyone selling a clean answer.

You can accept the information gain thesis. You can commit to producing content with genuine additive value. And you still can't cleanly measure whether it's working, because the discovery surfaces have fragmented.

I wrote last week about how AI visibility dashboards are mostly statistical noise — most of the movement they report isn't signal. GA4's AI Assistant channel undercounts real AI-referred traffic and scatters a single source across multiple channels. Server logs are the only reliable source of truth for bot access patterns, and most teams don't have them properly parsed.

So you're being asked to invest in a harder, slower content model — genuine research, genuine synthesis, genuine first-hand contribution — at exactly the moment your measurement infrastructure is least capable of proving it worked.

That's not a reason not to do it. It's a reason to be honest with your CFO about the measurement lag. Which brings me to a related point: the CFO doesn't care about your rankings deck anyway. What they care about is whether you can articulate a defensible content strategy in a market where traditional attribution is degrading. "We're building content Google's own patents say it wants to reward" is a more defensible answer than "we're targeting 340 keywords in the top 20."

The A/B testing footnote matters too

There's a smaller story in today's news that connects to this. John Mueller warned against running A/B tests for excessive durations — specifically, year-long tests where multiple variations are being indexed. Google treats prolonged testing as a signal of deceptive intent.

The connection to information gain isn't obvious at first, but it's there. Both are examples of Google actively penalising content that games the surface without adding genuine value. Both reward stable, coherent, additive content and demote content that's engineered to manipulate discovery.

If your content strategy relies on tactical variance — mass-produced pages, minor variations, A/B tested landing pages left running indefinitely — you're operating in exactly the space these systems are designed to flatten. The systems don't care about your intent. They care about the pattern.

The honest limits

A few caveats I want to be explicit about.

A granted patent proves nothing about deployment. Google files patents defensively — to prevent competitors from using an idea, not necessarily to signal what they're doing today. The information gain patent could be sitting on a shelf. The extension of US protection to 2039 and the ongoing citations suggest active interest, but that's inference, not proof.

Second, "originality" is genuinely hard to define at scale. Machine measurement of novel contribution is a hard problem. Google almost certainly uses proxies — engagement signals, dwell time, cross-document overlap — rather than a clean "gain score." The patent describes one possible mechanism among many, and the real ranking systems are almost certainly a stack of overlapping signals, not a single lever.

Third, this doesn't mean every piece of content needs to be a research paper. Category pages, transactional pages, product content — these operate under different logic. The information gain framework applies most cleanly to informational and mid-funnel content, which is exactly the space AI Overviews are eating anyway.

Reasonable people can look at this patent and conclude it's directionally interesting but not actionable enough to reshape a content programme. I'd push back on that, but I understand it. The confidence I have isn't from the patent alone — it's from the patent plus `OriginalContentScore` plus `ContentEffort` plus the leaked Content Warehouse plus five years of watching commodity content lose ground plus what I see in client audits every week.

What actually changes on Monday morning

If you're running a content programme, here's what this patent should push you toward. Not a rewrite. A recalibration.

Stop measuring content success primarily by keyword coverage. Start measuring it by additive contribution — what does this piece contain that a competent reader couldn't get from the AI Overview or the top three organic results? If the answer is "nothing meaningful," it shouldn't ship.

Reduce volume. Increase depth. This is the least popular advice in content marketing, and it's been the correct advice for at least three years. The information gain patent is another data point in a very long argument that publishing thirty mediocre pieces a month is actively worse than publishing five with genuine contribution.

Invest in the inputs that produce originality: proprietary data, first-hand experience captured well, expert interviews, synthesis across domains, contrarian positions defended properly. These are slower and more expensive per piece. They're also the only thing left that Google's systems can't produce themselves.

And accept that measurement will lag. You'll be doing the right work before you can prove it's working. That's the position everyone serious about content is in right now, and pretending otherwise is worse than sitting with the discomfort.

The information gain patent isn't news. But mid-2026 is finally the moment where it stops being an interesting footnote and starts being the operating framework. Content teams that grasp that early will look prescient in eighteen months. Content teams that don't will spend that eighteen months wondering why their traffic keeps flatlining despite hitting every KPI on the deck.

That's the loop. And we built it.