The GEO Cat-and-Mouse Game Has Already Begun

I have been looking at a few different datasets about ChatGPT Search over the last couple of days, and something interesting seems to be happening.

Reddit citations dropped. A lot.

Promptwatch measured Reddit’s average share of citations in ChatGPT Search at 3.83% between July 18 and August 7. On August 14, it fell below 1%. Between August 14 and 17, the average was only 0.52%. (1)

Around the same time, researchers looking at what ChatGPT actually searches before giving an answer started seeing another change.

More searches.

More specific searches.

And, at least according to one dataset, many more searches narrowed to particular websites or including words such as “official” and “gov.” (2)

Then, on August 18, Google started rolling out another spam update globally and across all languages. (3)

Before I accidentally start a conspiracy theory on LinkedIn, there is no evidence that Google’s spam update has anything to do with what is happening with ChatGPT citations. 😅

But I think they are trying to solve a very similar problem.

They both need to give their users the best information they can find.

And we, marketers, have a long history of making that job harder.

We have seen this before

If you have worked in SEO for long enough, you probably know how this story goes.

Google produces a certain type of search result. People analyze what ranks. Someone discovers something that seems to work.

Then we give it a name.

Then someone turns it into a framework.

An agency starts selling it.

A software company automates it.

A LinkedIn post appears explaining how you can do the same thing in seven steps.

Six months later, everyone is doing it.

The signal becomes less useful.

Google changes something.

And we start again.

Obviously, I am simplifying decades of SEO history into a few sentences, but I think the pattern is familiar.

Google’s current spam policies specifically address scaled content abuse, meaning large amounts of unoriginal content created primarily to manipulate rankings rather than help users. (4)

Nobody wants to search for something and land on 2,000 words of AI-generated slop that exists only because someone identified a keyword with search volume.

Google doesn’t want that either.

Because if the results become useless, eventually people stop trusting Google.

I think exactly the same incentive exists for ChatGPT and other AI assistants.

The interesting difference is speed.

The SEO cat-and-mouse game played out over years. Sometimes decades.

With AI search, we can observe a new citation pattern today, analyze millions of responses tomorrow, publish a “GEO strategy” the day after, and have thousands of marketers trying it shortly afterwards.

That feedback loop is much, much faster.

Reddit might be the perfect example

Earlier this year, Peec analyzed five million query fan-outs collected across ChatGPT, Perplexity and Grok.

A query fan-out is basically an additional search an AI system performs after you ask it something.

You might ask:

“What’s the best project management software for a remote team?”

But the system doesn’t necessarily search only for that sentence. It can break the question into several searches looking for comparisons, reviews, features, different use cases and so on.

Peec found that ChatGPT was frequently adding words such as “best”, “reviews”, “comparison” and the current year to these searches. (5) Neil Patel’s analysis of the same research also found that fan-outs explicitly mentioning Reddit had increased significantly between January and May 2026. (6)(7)

Why Reddit?

I think the answer is quite simple.

Reddit has something that most company websites don’t.

People.

Messy people with opinions.

Someone saying that a vacuum cleaner was great for six months but the battery became terrible after a year.

Someone explaining that the software everyone recommends becomes a nightmare once you have 100 employees.

Someone saying the shoes actually run half a size small.

Someone else completely disagreeing.

These discussions are subjective. They can be wrong. Sometimes they are ridiculous.

But they are useful precisely because they are not another polished paragraph written by the company selling the product.

And of course, once marketers noticed ChatGPT using Reddit, the conclusion was almost inevitable:

ChatGPT likes Reddit, so we need to be on Reddit.

Then “being on Reddit” slowly turns into creating threads, buying old accounts, generating comments, manufacturing reviews and trying to make the whole thing look organic.

Which is quite funny when you think about it.

Because we are destroying the exact reason Reddit was useful in the first place.

Now Promptwatch is seeing Reddit’s citation share in ChatGPT collapse. The researchers themselves are careful not to overstate why. They explicitly say the chart tells us when the change happened, not why, and that even a data collection issue cannot yet be completely ruled out. (1)

So I don’t think the conclusion is “Reddit is dead.”

That would be making exactly the same mistake again.

The lesson is about authenticity and information quality, not Reddit.

The dangerous part is confusing observations with truth

There is another dataset I found particularly interesting.

Nectiv took roughly 4,000 prompts from an earlier study and ran them again. In its previous dataset, ChatGPT averaged 2.17 fan-out queries. In the new one, the average was 7.61.

They also found site: appearing in 64% of the fan-outs they analyzed, with terms such as “official” and “gov” appearing frequently too. (2)

That is interesting.

It does not mean “ChatGPT now trusts official websites.”

It doesn’t mean site: is a ranking factor.

It doesn’t mean someone at OpenAI changed a setting called Authority from 4 to 8.

Maybe the direction is real. Maybe the behavior differs depending on the question, model, category or hundreds of other things we cannot see.

The point is, we don’t know.

We can observe the system. We can run experiments. We can collect millions of outputs and connect findings together.

Then we can make deductions.

But the deduction is not necessarily the truth.

That distinction is especially important in a non-deterministic system that keeps changing.

Maybe we should create information, not just content

This is where I think the conversation becomes much more interesting for brands.

Let’s say I run a skincare company.

I can write another article explaining the benefits of vitamin C.

But why should an AI system use me as the source for that?

I am not a medical institution. I am not a dermatologist. There are researchers, hospitals and specialist publications that are probably much better sources for a general question about skin health.

Even if I produce the most beautifully SEO-optimized version of the same information, I am still mostly repeating something the internet already knows.

Now let’s look at another scenario.

I open our Google Analytics data.

I check Google Search Console.

I look at the searches people make inside our own website. I talk to customer support. I look at returns, product reviews and the questions customers repeatedly ask before purchasing.

And I notice something.

Maybe a lot of visitors search for one particular problem, but visitors searching for another, more specific problem are three times more likely to purchase a certain product.

Now I have something potentially interesting.

Not something I should immediately throw into an article and call “research.”

First I need to understand the sample size. I need to see whether the difference holds across a meaningful period. I need to make sure I am not exposing personal information. I should probably understand whether there are other variables explaining the behavior.

Correlation is not causation just because I put it in a nice chart.

But once I have done that work, I may have something the internet genuinely did not know before.

I am no longer just creating content.

I am creating information.

And interestingly, Google’s new guidance for generative AI search says something very similar.

Google recommends what it calls “valuable, non-commodity content.” It specifically contrasts firsthand experience and unique viewpoints with content that simply summarizes information already available elsewhere. It also warns publishers against creating pages for every imaginable fan-out variation in an attempt to manipulate generative search results. (8)

I don’t think that means traditional content is suddenly useless.

And I definitely don’t think GEO is useless.

Quite the opposite.

There is important work to do.

Your website needs to be crawlable. Your information architecture should make sense. Pages should connect to each other properly. Schema should describe what is actually on the page. Product specifications should be complete. Machines should be able to understand what you sell, where you sell it and the differences between your products.

Google’s own AI search guidance still emphasizes those technical foundations. (8)

But there is an important order here.

Create something worth finding.

Then make it easy to find.

Not the other way around.

Visibility is only the beginning

There is another reason I think this matters.

For most of SEO’s history, the job of the search engine ended with sending you somewhere.

AI is starting to move further into the decision.

OpenAI itself says ChatGPT Search can rewrite a user’s question into one or more targeted searches before returning an answer. (9)

For shopping, it goes much further.

People can ask ChatGPT what they should buy, compare products and narrow their choices without opening ten browser tabs. OpenAI is also encouraging merchants to provide structured product feeds so ChatGPT can access more complete and up-to-date product information. (10)

This means there are several different things we can easily mix together.

A brand can be mentioned but not cited.

It can be cited but not recommended.

It can be recommended, but with the wrong product information.

It can be recommended correctly, but the product might not actually be available to that shopper.

And eventually, the whole journey can happen without the traditional website visit we used to measure.

This is why I see AI visibility as the first step, not the final objective.

At Herm, the framework we use to think about this is AI Commerce Readiness.

Be found. Be sellable. Be chosen. Be proven. Then scale on AI.

I don’t want to turn this article into a Herm sales pitch. The reason I am mentioning the framework is that I think it is useful for separating the stages.

Visibility matters enormously.

If you are not in the consideration set, everything after it is irrelevant.

But winning a citation dashboard is not the same thing as winning commerce.

So what should brands actually do?

If I were responsible for a brand today, I would absolutely pay attention to all these changes.

I would track citations.

I would study fan-outs.

I would look at which sources appear when people ask questions in my category.

I would invest in GEO fundamentals and make sure machines can actually understand my website and products.

But I would be very careful about turning today’s observation into tomorrow’s strategy.

If ChatGPT is searching Reddit today, the strategy shouldn’t be “hack Reddit.”

If site: searches are increasing, the strategy shouldn’t be “how do I make ChatGPT use site:mycompany.com?”

Watch the system.

Learn from it.

Don’t worship its current quirks.

And spend at least as much time asking a harder question:

What does my company know that nobody else can write?

Maybe it comes from customer behavior.

Maybe it comes from years of product development.

Maybe it is hidden in support tickets.

Maybe you tested 42 different versions before launching the final product.

Maybe you know that customers consistently misunderstand one feature.

Maybe an assumption you made while building the business turned out to be completely wrong.

Maybe you have 10 years of sales data showing how a category changed.

There is information inside almost every company that has never made it onto the public internet.

That is where I would start.

What I think happens next

I expect citation patterns to keep changing.

I expect AI systems to get better at choosing different types of sources for different types of questions.

If I ask when a product launched, the company itself might be the best source.

If I ask whether that product actually lasts three years, I would much rather hear from someone who owned it for three years.

If I ask about a medical issue, I would hope the system prefers medical expertise over a cosmetics brand that happens to rank well.

There probably won’t be one universally “trusted” type of website.

Context will matter.

I also think AI visibility and AI commerce will continue moving closer together. Discovery, comparison, recommendation, product information, availability and transaction are already starting to collapse into the same interface. OpenAI’s shopping work is a very visible example of that direction. (10)

And yes, people will try to manipulate all of this.

We did it with Google.

We will do it with ChatGPT, Gemini, Perplexity and whatever comes next.

The platforms will respond.

Then we will respond to their response.

The cat-and-mouse game isn’t going anywhere.

It is just getting much faster.

Which is why I think the long-term answer is actually quite boring.

Create information worth retrieving.

Tell the internet something useful that it didn’t know before.

Explain how you know it.

Make it understandable for people.

Make it accessible to machines.

Then do it again.

Algorithms will change.

Citation patterns will change.

The websites AI systems prefer will change.

Useful information has a much better chance of surviving all three.