Information Gain Is the Missing Metric in GEO Content

BrightonSEO Keynote Recap · GEO Strategy

When AI can produce unlimited content, what makes yours worth citing? Information gain, the metric missing from most GEO content strategies.

By Paris ChildressCo-founder & CEO, GEOforge5 min read

That was the question I opened my BrightonSEO keynote with. Below is the argument, the data from a real GlycanAge programme, and the workflow we use to build original knowledge into content production.

Why does AI content all sound the same?

Most content teams are choosing between two bad options. Generic AI content is fast, but it’s often interchangeable. Manual production can capture real expertise, but it takes time you don’t have. What we actually want is the efficiency of AI and the distinctive knowledge of the business.

The reason the first option disappoints is simple. AI mostly feeds on public information. If every brand in your category starts from the same public sources, it’s unsurprising that every brand’s content sounds the same. You can change the wording all day without adding anything a reader (or an LLM) couldn’t already find somewhere else.

What is information gain?

AI is a truffle hunter, and your business has something more valuable than public information: original knowledge. Customer experiences, expert explanations, research, and lessons from actually doing the work. Those are the truffles. The opportunity is to capture that knowledge and make it available.

The value that knowledge adds has a name: information gain. Picture what existing content already covers, then picture what your passage says. Some of it overlaps. Information gain is the new, useful information your passage adds on top.

This isn’t a GEO-era invention. Google’s 2018 patent describes generating an information gain score with a machine learning model, then presenting documents in a way that reflects the likely information gain a user would get from viewing them. We can apply the same principle editorially.

We do that with an AI-assisted scoring rubric. It’s our assessment method, not a Google metric, and it sorts content into five bands.

Higher scores mean the piece gives an AI engine something it can’t get anywhere else. The obvious next question is whether that shows up in the results.

What happened when we scored real content?

GlycanAge is a biological-age testing company, and we ran a content programme for them from 19 May to 7 Sep 2026. The median information gain score across the published pieces was 8.5 out of 10. The results varied by engine and topic:

• 35 of 52 published pieces were cited in measured AI answers.
• The median time to first AI citation was 5.9 days.
• Google AI share of voice moved from 27.8% to 32.3%, a 16.1% relative increase, comparing 5-6 Jul with 7 Sep. Share of voice is the share of tracked AI answers that mention the brand.

The competitive picture matters too. As of 28 Sep, GlycanAge held 31.11% share of voice against 7.74% for Function Health, 3.7% for NOVOS and 2.61% for Tally Health. It has held a large lead over those larger competitors for the whole tracking window.

Then we looked at how citation rates related to information gain scores.

How do you ground AI drafts in your own knowledge?

The content playbook itself is familiar: measure, create, distribute, re-measure. AI speeds up the creation step, but it introduces a quality-control problem. Google’s guidance on generative AI content, updated on 2 Oct 2026, says as much: generative models don’t retrieve facts, they predict a likely sequence of words, so outputs may contain inaccuracies, and all AI-generated content should be manually fact-checked and reviewed before publishing.

Human verification stays. But could we reduce the burden by fixing part of the problem before drafting? What if the AI started with your evidence?

That’s the missing step: ground. Instead of discovering missing facts after the draft, you supply the relevant evidence before it. The mechanics are retrieval augmented generation (RAG): retrieve the relevant passages from your own knowledge base, supply them as context, then generate.

The difference shows up in the draft. A generic draft says AI is changing search and brands should create useful content and publish consistently. It’s accurate, and it’s also what every other brand’s draft says. A grounded draft says the same thing with a sourced fact in it: Pew Research Center (22 Jul 2025) found users clicked a traditional result in 8% of Google visits with an AI summary, versus 15% without.

Sourcing improves the substance and makes checking easier, because every claim points back to evidence. The grounded example also carries a content quality audit that scores accuracy (9.8 out of 10), information gain (8.5 out of 10) and privacy (pass). The question that follows is where the most distinctive evidence comes from.

Where are your truffles buried?

The most distinctive evidence is already being spoken out loud inside your business. Customers explain their questions and their experiences. Your experts explain causes, trade-offs and solutions. Both are original knowledge that no LLM has read.

The accessible place to start is one recorded interview. A customer call, a sales conversation or a subject matter expert walking through how they solve a problem. If your team already records every meeting, ask what happens to those recordings. If the answer is that they sit in a folder, that’s your first truffle patch.

Capture that evidence, put it in a knowledge base, and the loop starts to compound:

1. A customer asks a question.
2. You capture evidence that answers it.
3. You publish the answer.
4. That answer prompts the next customer question.

Each useful source adds one more entry to the knowledge base per cycle. I pitched this as AI as your best salesperson, and the point is commercial, not cosmetic. The goal is better answers for buyers that contribute to opportunities.

You can check that it does. Add “How did you hear about us?” to your forms, with an AI chatbot as an option, then match the responses to opportunities in your CRM. That gives you a share of the pipeline influenced by AI search from your own data, not from a guess.

Start with one truffle

If unlimited AI content is the baseline, being worth citing comes down to what you know that nobody else has published. Information gain gives you a way to score that, grounding gives you a way to build it into every draft, and your own customers and experts supply the raw material.

So here is the one thing I asked the room to do. Identify one proprietary source in your business. Capture it. Use it to answer one customer question better than anyone else has.

Sources & method. GlycanAge figures come from a GEOforge case study: programme 19 May to 7 Sep 2026, share of voice compared across 5-6 Jul and 7 Sep 2026, competitor share of voice as of 28 Sep 2026. Results varied by engine and topic. The information gain rubric is GEOforge’s own AI-assisted assessment method, not a Google metric. Also referenced: Google’s 2018 information gain patent, Google Search’s guidance on generative AI content (updated 2 Oct 2026) and Pew Research Center (22 Jul 2025).