Information gain is how much new, useful information a piece of content adds beyond what AI engines have already seen, and it is the missing metric in most GEO content strategies. Paris Childress walks through his brightonSEO keynote of October 8, 2026, covering GEOforge's information gain scoring, a GlycanAge case study, grounded drafting with retrieval-augmented generation, and self-reported attribution.
AI engines reward knowledge they have not seen before. Most content teams are stuck between slow manual production and scaled AI content that only recycles what is already public, and neither feeds AI what it actually prefers: original, proprietary evidence. Paris Childress, founder of Hop AI and co-founder of GEOforge, calls that evidence truffles, and argues that information gain is the metric that tells you whether a piece of content contains any.
This episode is Paris's own walkthrough of the keynote he gave at brightonSEO on October 8, 2026.
Information gain is how much new, useful information a passage adds beyond content that has already been covered and absorbed into AI training data. Google filed a patent on an information gain score in 2018, when the recipient was a human searcher. Today the recipient is AI crawlers and the models they feed.
"AI, first and foremost, is a truffle hunter, and truffles are original, useful knowledge that only your brand possesses." Paris Childress
GEOforge scores drafts on a scale from 0 to 100%. At the low end the content is common public knowledge. At 80 to 100% it carries internal metrics, proprietary processes, unpublished research or strategic insight an LLM would not have encountered. Paris suggests that content scoring below about 20% probably is not worth publishing yet.
In the GlycanAge case study, yes. The biotech company published 52 pages with GEOforge:
Paris attributes much of the lift to GlycanAge's willingness to share proprietary studies, and frames the real question every brand must answer: how much of its secret sauce it is willing to share with AI in exchange for visibility.
Google's guidance on generative AI content accepts its use but warns that models predict words rather than retrieve facts, so outputs need careful fact-checking. Grounding supplies the facts before drafting: proprietary files, transcripts and documents go into a vector database as semantic chunks, and the content engine retrieves from it through retrieval-augmented generation. The result is a grounded draft that cites its sources, can be scored for accuracy, information gain and privacy, and needs far less editorial cleanup than a generic draft.
Every brand has truffles. Some are buried deep, but the job of content marketing in the AI age is to dig them up and make them available to the engines that will reward them.
[00:00:00] Paris Childress: Hi, everybody. I'd like to walk you through the presentation that I gave at brightonSEO on October 8th on information gain. So I'm gonna be walking through the same slides that I presented in my keynote at Brighton, which was really a great time, great audience, and a great event. My topic is about information gain, which I believe is a missing metric in most content marketing strategies, particularly the ones that are now making this important transition from SEO-led to GEO-led. As marketers now in the age of AI, we are stuck between a rock and a hard place.
Producing content the manual way, even though it delivers high quality, it really is too slow and not competitive enough. So we can't continue on with 100% manual processes here. And if we lean too heavily on AI to generate content, then it gets kind of risky. We don't wanna scale pure AI content, and we've all heard the warnings about AI content getting flagged or getting fingerprinted with SynthID, perhaps getting penalized.
So we don't want to over-index on generic AI content. We've gotta squeeze in the middle somehow. We've gotta find this sweet spot, middle ground. And we know that AI is content enough to feed on the trough of AI slop, which is mostly public information that they find on the web, and this is the large majority of AI-generated or even AI-assisted content that is feeding AI engines right now.
But really what AI prefers are truffles. AI, first and foremost, is a truffle hunter, and truffles are original, useful knowledge that only your brand possesses that most likely hasn't surfaced yet on the web, so it hasn't been discovered, and it's buried deep inside of your company. And our goal is to unearth those truffles and make them available for AI so that they can then reward us with greater visibility. And that brings us to the concept of information gain, which is how much new useful information a passage adds beyond existing content that's already been covered.
So you can imagine the vast body of information that's already out on the internet, that's already been crawled and ingested and baked into AI's training datasets. And then you have your content, which hasn't surfaced yet, and it contains new and useful evidence about what makes your brand unique. And the degree to which you are willing to share that new useful evidence, that information gain with AI, that area that is overlapping, that shared area there is called information gain.
And the term specifically means that AI is gaining new knowledge through your sharing of your proprietary brand data. Information gain is not a brand new concept. In fact, Google filed a patent for information gain back in 2018, and they described it back then as machine learning models that generate an information gain score.
And in this context, the recipient of information gain was a human user. And now in today's AI-driven landscape, the recipient of information gain now is AI. More specifically, it's AI bots and AI crawlers.
What we do within GEOforge is we have created a scoring rubric for information gain on a scale of zero to one. And at the low end of the scale, it is entirely common public knowledge. And as you get to the higher end of the scale, 80 to 100%, this is highly proprietary knowledge that we are contributing, offering, and making available to AI systems.
It's entirely original with internal metrics, proprietary processes, unpublished research, or unique strategic insights that an LLM would not have encountered. This is the end of the spectrum where we really wanna focus, and I would even suggest that if a piece of content has a low information gain score, let's say below 20%, it's probably not worth publishing, and you need to keep working on that content before publishing. We have several successful case studies in delivering high information gain and observing what that does to citation rates and AI visibility.
This client is GlycanAge. It's a biotech company, and with them, we have published a little more than fifty-two pages so far. Thirty-five of them have already received citations in under six days on average. And the share of voice, particularly within Google AI properties like AI Overviews and AI Mode, has risen from about twenty-eight percent to just over thirty-two percent, so a sixteen percent relative increase.
And I think that that lift is in part due to the high information gain that is baked into every piece of content. And they're sharing a lot of proprietary studies, and this really is the crux of the argument. The degree to which you are willing to share your secret sauce or your alpha with AI, despite the fact that this may contain risks of, let's say, making that available to competitors, that, that is the question that every brand needs to answer.
To what extent are we comfortable sharing our proprietary data with AI, knowing that we're gonna receive a reward for that with AI visibility? And GlycanAge has been very generous and not extremely conservative in this respect. What we saw in particular is when we break down the articles published by information gain score, most of them have been seventy percent or higher.
But the ones that really outperform in terms of their median citation rate are the ones scoring ninety percent or above. There we really see a spike in citation rates. And overall, their share of voice, they have enjoyed a large lead over much larger and better-funded competitors over time. And this is, I think, primarily due to the fact that they are more willing to share their proprietary data and more focused on information gain delivery through their content. So let's pull back now and try to put information gain into the perspective of the broader content marketing playbook.
And this is a familiar playbook that we've been running since the early days of SEO, and we're now all trying to adapt this content playbook to a GEO-first world. It starts with measurement and measuring some sort of baseline of where we're starting from so that we know what good looks like later on.
And then we create content, publish it, we distribute that content, observe any changes, and then remeasure the impact of that content.
Google has come out very recently, just a couple of weeks ago, with a very, very important update that they published in their webmaster blog about using generative AI content on your website, and specifically, they said this: "Keep in mind that generative models don't retrieve facts, but predict a likely sequence of words based on their training data.
Because of this, generative AI outputs may contain inaccuracies, also known as hallucinations. It is critical to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing." So this is a definitive statement by Google saying two things. First of all, we know you all are using AI to create content, and it's okay.
We're not going to penalize you for that. If we read between the lines, they are in no way saying that this is against the rules or this is gonna be penalized, but they are secondly cautioning us that because AI is probabilistic and predictive in nature, this type of content often produces inaccuracies and hallucinations, and it's very, very important to fact-check and perform a very rigorous post-editorial review on that content.
And this is suggesting a pretty tedious editorial burden on content marketers. So now I wanna pose this question: Could supplying reliable evidence, such as information gain, before drafting or during the initial drafting process, could that reduce these errors that Google is alluding to and relieve this editorial burden that is suggested by this new process?
And the answer is yes, and that missing input is proprietary knowledge, and that needs to come before the creation of content, and this is called grounding. When you ground content in your proprietary knowledge and ingest or infuse it with high information gain, you're setting yourself up for a more successful piece of content, and you're also going to alleviate the editorial burden that comes with fact-checking because these facts are gonna be correct from your first draft.
And the mechanics of this is a process called RAG, or Retrieval-Augmented Generation. It starts with a vector database. So what you're doing is you're storing all of your proprietary knowledge, your files. These could be transcripts, webinars, documents that you haven't shared publicly. You're dumping all that into a vector database, and what that vector database does is it chunks all that information into semantic-- what are called semantic chunks.
They're similar pieces of information, and then they're storing them in this vector database, and these are called embeddings. And then later, when you are creating content, when you have a, a content engine that's AI-generated that is creating the content, it is grounded in that RAG vector database. And just like what AI does when it goes and retrieves an answer in real-time or provides an answer in real-time via retrieval through real-time search, it is augmenting that response through what it finds.
We're doing the same thing with our content generation process. We're generating content that is retrieving facts and evidence from our knowledge base, and it is embedding those into our content from the very first draft in a way that is traceable and easy to fact-check. And the result is that instead of getting a generic draft, which probably has low information gain and a very unlikely chance of great success, you are creating a grounded draft.
And this is a comparison here. So with a grounded draft and with the help of some AI tools that we've built into GEOforge, you can have a draft that specifically cites passages and facts from your knowledge base in that draft. And you can also score it on the basis of accuracy, information gain, and also privacy.
If you have things that are too sensitive, you can have certain rules that would flag anything that you do not wanna share publicly. And that grounded draft is gonna be much, much faster for you to then get published and reduce that editorial burden. So where does this knowledge come from? I think the best source of knowledge for brands that maybe have a hard time finding this data is to capture the voices of your customers, of your prospects, and your experts.
And that means, first and foremost, recording everything. Recording sales calls, recording meetings with customers, recording customer support logs, recording, of course, all the webinars, podcasts. Every time that someone is speaking about your brand, whether it be a customer asking questions or an in-house expert sharing their detailed knowledge, you wanna capture that, transcribe it, and flow it into the knowledge base.
This is probably the richest and most persistent source of supplying that knowledge base day in and day out with high information gain.
And what we're trying to create here, particularly with capturing sales call transcripts, is we wanna transform AI into your best salesperson. So imagine that someone asks a very detailed question in a demo from one of your top sales people, and they get a great detailed answer to that. That gets captured and transcribed and flowed into a knowledge base.
And then very soon after that, that same question and answer comes out in a piece of content. It gets crawled by AI. AI updates its training data on your brand relative to this question. And then a few days later, a prospect comes along, and before booking the demo, they go straight to ChatGPT with that very same question.
And now ChatGPT can answer that question by essentially lifting out that specific answer that your expert salesperson delivered, and they can deliver that quality answer to that prospect right then and there, so that hopefully when that prospect downstream shows up for that demo, they're more educated, they have accurate answers to their questions, and they're closer to making a purchase.
So I wanna also now connect this to attribution because if we are able to get AI saying accurate things about our brand, promoting our brand, recommending our brand in the right conversations, we wanna be able to ultimately see what impact this has and even what impact that information gain effort has had downstream on the sales pipeline.
And the best way to try to triangulate ROI for GEO is to ask every prospect when they fill out that first form, "How did you hear about us?" And you offer them among these options, make sure that AI chat, AI chatbots like ChatGPT is one of those selections. Most likely they're gonna remember that chat session because it's gonna be memorable.
It's probably gonna be fairly long and containing several questions and answers. It's gonna be memorable. I think this is a pretty accurate way of getting people to self-attribute. This is called self-reported attribution. And then later when you capture those responses, and you can match them in your CRM to what downstream will become opportunities, you're able to then see what portion of your sales pipeline is influenced by AI as a first touch.
And I would consider this mostly first touch attribution because AI chatbots are the new front door to the customer journey. It's often the very first starting point. And if you're able to do this over time, first by benchmarking this with some kind of a baseline and then observing the lift in your AI search referred or AI search influenced sales pipeline, then you can really see what impact your efforts are having on actual revenue.
So in closing, I wanna bring this back to the truffle analogy. Every brand has truffles. Some of them are buried very deep and need to be dug up with some degree of effort. But I'm telling you that it's worth it because our goal as content marketers now is to help AI find our brand's truffles. And that's the end of this presentation, and I hope you all can go forth and hunt and dig up your brand's truffles, and AI will reward you generously for that.
Thanks.
