If AI Stops Searching Google, Does SEO Still Matter?

Google, and many others (including this publication) have always maintained that SEO best practices are required to do GEO/AEO. Google has even gone so far as to say that they are essentially the same thing and not separate from traditional SEO. In large part this is because most AI systems and chat interfaces still rely heavily on search engine result pages to collect information.

But there are signs on the horizon that all of that might be changing. For example some users are reporting changes to ChatGPT’s fan-out queries. Some of those changes are an increase in branded terms, as well as an increase in the use of the site: search operator. We have also seen a dramatic drop in reddit citations. All while the number of query fan-outs per prompt is dropping.

All of this could mean many different things. OpenAI could be retooling their recent models in a way that saves resources. Or maybe they have greater confidence in the new models’ ability to process responses before RAG. But because many of these changes are also taking place at the same time that Google is implementing more tactics to block OpenAI from it’s results, it seems likely that OpenAI, Anthropic, and others might be exploring other ways to supplement their responses with out relying on Google’s results.

I imagine in the most likely future scenario, search engines, in some form or anther, still exist. The major difference is that AI tools, such as ChatGPT or Claude, would not use them to search the web and collect information. What would they use instead? My best guess is a combination of their existing LLM entity embedding, with their own index of the web like Google has.

Relying too much on the LLM embedding without RAG would be dangerous. However, it might be a good starting point to best define how the future RAG pipeline would work. For example when I send the following prompt What is the best soccer shoe for a 10 year old boy? to ChatGPT, these are the query fan-outs:

Notice all of the brand mentions above? Those aren’t coming from search engines, they are embeddings inside the LLM. At the moment ChatGPT uses these as branded search terms when retrieving results from Google. They are relying on what the model already knows about soccer shoes for boys to generate more focused search terms to Google. If they one day didn’t use Google, and instead used their own internal index, these types of brand embeddings could be the first step to querying that internal index with targeted terms.

But, what would that internal index look like? How would it store data? My best guess is that it would be similar to Google’s index but far less sophisticated and probably smaller. Which means it would probably allow for them to do semantic search on things like title tags, meta descriptions, h1s, and body content. If they are using branded queries or site: level search operators, I doubt they would need much of a link analysis, because authority would be implied from the LLM embeddings.

If the above scenario came true, would we need to update our SEO For LLMs (GEO/AEO) Checklist? If ChatGPT and others no longer used search engines, would SEO even matter for GEO/AEO?

If search engines become less important to AI retrieval, the first thing I would focus on is how strongly your brand is associated with the topics you want to be known for.

In the example above, ChatGPT already associated Nike, Adidas, and New Balance with youth soccer shoes before it searched the web. Those associations matter because they can influence what the system decides to search for next. If that same process were eventually pointed at an internal index instead of Google, getting your brand into that initial consideration set could become even more important.

That starts with making sure your content is actually accessible to the systems that may be used to train future models. OpenAI, for example, distinguishes between GPTBot, which site owners can use to control whether content may be included in future model training, and OAI-SearchBot, which is used for ChatGPT search. Anthropic similarly documents ClaudeBot as a crawler that collects public web content that could contribute to model training.

That does not mean allowing an AI crawler today guarantees that your brand will appear in the embeddings of a future model. We have no idea how any individual piece of content will be selected, weighted, filtered, or represented during training. But blocking training crawlers entirely removes at least one possible path for that information to make it into future models. The content itself matters too.

Brands should consistently use their brand and entity names alongside the terms they want to be associated with. If you sell accounting software for small businesses, for example, it should be extremely easy for a system processing your content to connect your company with "small business accounting software." That relationship should not require an LLM to make five inferential jumps.

This is also where simple "is" statements can become surprisingly useful.

"Acme is an accounting platform for small businesses" creates a much clearer relationship between the entity and the category than a homepage filled with copy like "Take control of your financial future." The second version might sound better to a marketing team. The first one tells a machine what the company actually is.

This does not mean every page should become repetitive keyword soup in an attempt to manipulate embeddings. In fact, that would probably create the same problems it creates in traditional SEO. The goal is clarity and consistency. Your brand, products, categories, audiences, and use cases should repeatedly appear in sensible relationships across the content you publish.

And if model training is part of how those relationships are eventually learned, this is not something I would wait to begin.

Search engines can recrawl a changed page tomorrow. Model training works on a very different timeline. Even when your content is available for training, there may be a substantial delay before that information could influence a future production model. That makes this one of those rare optimization problems where starting early may matter more than optimizing aggressively.

The second half of this hypothetical system looks much more familiar to SEOs.

Even if an LLM knows that your brand is relevant to a topic, it still needs current information. Prices change. Products launch. Inventory disappears. Policies get updated. No serious AI assistant can depend entirely on what was captured during model training.

Some kind of retrieval layer would still be necessary.

If AI companies build larger indexes of their own, the technical requirements for appearing in those indexes are likely to look a lot like the requirements we already deal with in search. Your pages need to be accessible. They need to render. The important content needs to be easy to extract. And the system needs enough information to understand what each page is about.

We can already see pieces of that infrastructure forming. OpenAI tells publishers that content needs to be accessible to OAI-SearchBot to be included in ChatGPT search summaries and snippets. Its crawler guidance also calls out familiar technical failure points such as robots.txt restrictions, WAF rules, JavaScript challenges, authentication, redirects, and other systems that can prevent automated access.

So even in a world where Google becomes irrelevant to ChatGPT, crawlability does not suddenly become irrelevant. If anything, it may become more complicated.

SEO teams are accustomed to testing what Googlebot can access and render. A future where several AI companies operate independent retrieval systems could mean monitoring multiple crawlers with different capabilities, crawl budgets, rendering behavior, and tolerance for complicated JavaScript.

That makes crawl and rendering efficiency worth improving now, regardless of which architecture wins.

I would also continue using traditional SEO elements such as title tags, H1s, headings, and prominent body copy to establish what a page is about. Not because we know that some future OpenAI index will weight a title tag the way Google does. We don't.

But if a system is performing semantic retrieval across a large collection of documents, clearly describing the primary entity and topic of a page gives it better information to work with.

A title like "Acme Small Business Accounting Software" leaves very little ambiguity about the relationship between the brand and the category. That is useful to Google today. It would probably also be useful to a much simpler retrieval system tomorrow.

This is one reason I am skeptical of the idea that GEO/AEO eventually replaces SEO. The search engine might disappear from the middle of the process. The underlying infrastructure problems probably won't.

Someone still has to crawl the page. Someone still has to render it. Someone still has to understand it. And someone still has to decide whether it is relevant enough to retrieve.

Those are SEO problems, even if Google is no longer involved.

Maybe not. There is a perfectly reasonable version of the future where ChatGPT, Claude, and other assistants continue using traditional search engines for a large portion of their retrieval. Building and maintaining a high-quality web index is enormously expensive. Google has spent decades doing it. Outsourcing at least part of that problem may continue to make more sense than rebuilding everything from scratch.

There is also a danger in optimizing too aggressively for an architecture that does not exist yet. SEO has a long history of people spotting one signal, extrapolating an entire ranking system from it, and then creating a new checklist that everyone supposedly needs to follow.

I don't think we should do that here. But I also don't think brands should ignore the possibility.

The useful thing about most of these recommendations is that they do not require a speculative GEO project. Making your site easier for legitimate AI crawlers to access is useful today. Making your pages faster and easier to render is useful today. Clearly connecting your brand with its products, categories, and areas of expertise is useful today. Writing title tags and headings that accurately describe the page is useful today.

These are mostly good SEO practices with a potentially larger payoff if retrieval changes. The one area where waiting could create a real disadvantage is model training.

If brand associations inside the model eventually play a larger role in deciding which sources get retrieved, you cannot necessarily fix that problem with a few weeks of optimization. You may have to wait for new data collection, new training runs, new models, and eventually widespread adoption of those models.

By then, the brands that established clear relationships between their entities and important topics years earlier may already have an advantage.

None of this means search engines are disappearing. It means our job is to think about what happens if they become optional.

And when you strip Google out of the middle of the system, most of the things that remain still look a lot like SEO.

Reinforcement Learning: Training through rewards and penalties based on actions.

Learn More

One more thing: AI is only as good as its operator, and if you are reading this newsletter, you’re better than most!

Till next time,

PS: Let me know what you think of this issue, or anything else here: [email protected]