Google Weighs In: Updated Understanding on GEO

GEO assumptions that Google says to ignore.

Beyond confirmation of what works, Google also spent time debunking tactics that have sprouted up as assumptions over the past few years, providing their own "Mythbusting" section within the Search Central documentation. According to Google, we can stop worrying about:

llms.txt files.

Google says you don't need machine-readable text files, AI-specific markup, or Markdown versions of your content to appear in generative AI search. While traffic analytics has shown that LLM-assisted tools crawl and ingest file types beyond the standard HTML page — an assumption that led many to promote llms.txt as a GEO tactic — Google claims there is no special treatment.

This is not a total surprise. Even here at Blend, we've positioned llms.txt as an unproven but very low effort future-proofing play. It's not hard to create one of these files, so the benefit of having one ready outweighs the chance that it might not necessarily be worthwhile — especially since Google's guide only speaks for Google, and other AI platforms may handle things differently.

Content chunking and rewriting for AI.

Content chunking is the idea that you should break content into small, self-contained pieces for AI to cleanly extract. With the assumption that LLM-assisted tools pull directly from the body of your webpage, this was a play to help provide the full context of an idea without losing anything from neighboring paragraphs.

Google says this is not necessary — that their systems can understand multi-topic pages, and are able to pull relevant sections without the need to restructure the page into little interchangeable LEGO block paragraphs. They also say there's no need to stuff your content with synonyms and keyword variations — that Google's systems understand synonyms and general meaning enough to provide the correct amount of context.

This all makes sense, honestly. On a personal level, the idea of tearing apart natural reading patterns to benefit robot AI always felt very weird and, to be frank, totally detached from what we’ve always learned about web writing: write like a journalist, with clear headings and inverted pyramid paragraph structure. Turns out, when you write for human understanding, you also write for the tools that are trying to understand and synthesize human understanding.

Inauthentic mentions.

The most obvious "well duh" moment of Google's "mythbusting" is that there is no benefit to gaming paid or fake mentions. Faking influence is the oldest game in SEO: from link swaps and paid footers to comment spam and guest post farms — every era invents new ways to manufacture the appearance of credibility, until Google eventually catches up and says "we've never counted that." This one is the same announcement, just with GEO swapped in for whatever the acronym was that year.

Structured data and schema for AI's sake.

The last point is the hardest, because it's a bit misguided. As we've begun understanding what works and what doesn't in the GEO space, schema markup — structured data that helps search engines understand what your content is about — has been positioned by many as one of the highest-impact things you can do for AI search visibility.

Google's position is that structured data isn't required for generative AI search features, and there's no special schema to add for AI. The idea is that LLM-assisted tools have enough of a shared model of information that they can easily parse the information without explicit structure and context.

Don’t confuse this and start throwing all of your structure out: structured content is still king. We know that vendors like Optimizely are pushing for markdown-based page summaries behind the page itself, continuing to subscribe to the idea that LLMs can understand meta content and will use any shortcut you feed them, not to mention the importance of structured headings and accessibility-focused machine allowances that are already a part of “good web writing.” Semantic HTML and rich-text editors are built for structure, and that structure still matters.

What’s more, Google still recommends schema for rich results eligibility; essentially, Google has said "Our model doesn't look at schema, but schema is still important for other areas where context matters." Recipes, events, articles, questions and answers — all of these still benefit from the structure of schema, but they are not weighed as heavily as we once thought.

At least, for Google. Which brings us to our next point: Google isn't the only game in town.

Where this shifts our earlier GEO research.

If some of this sounds contrary to advice you've heard from other firms — or, even from us at Blend — you are right. The thing with GEO is that, until this small glimmer of transparency, the individual generative AI models were (and, honestly, continue to be) black boxes. Now, there's a little bit of light — but a little bit of light does not suddenly illuminate the entire field. Think of it this way: a newspaper might open up their editorial process to tell you exactly how it decides what goes on the front page … but that doesn't mean you can go into another newspaper and start selecting stories for them.

This is what we're seeing now. ChatGPT, Perplexity, Claude, Google's many versions of AI — all of these are different tools, selecting results and pulling content based on different models. Of course, Google wants to get in and let us know what's behind the curtain — they have a vested interest to improve their entire search experience. What works for Google may not work for the others.

And those others matter. Depending on which report you look at, Google's AI tools represent somewhere between a fifth and a quarter of overall AI chatbot usage — which means hearing their expectations is a great hint into what matters for AI-assisted search, but it is far from the full picture. Google has a vested interest in these announcements, which is to build usage of their own tools rather than lose market share to the new kids on the block.

But it does mean our framing needs to shift. Schema is still valuable for content clarity and for platforms beyond Google, but it's not the Google AI optimization play it was sometimes positioned as. And the editorial work — writing clearly, structuring for humans, providing a real point of view — that holds up everywhere, no caveats needed. Mostly, we've learned that the best thing we can do for GEO is what search companies have been requesting for as long as search has been a thing: instead of writing to game the system, write to help the humans who are reading it.

Why Google, and why now.

Step back from the specific recommendations for a second, because the most interesting part of this might not be what Google said. It's that Google said it at all.

For 18 months, no major AI platform has given us any information about what happens under the hood. OpenAI has a FAQ for publishers, but it's about opting in or out of training data, not about how to get cited in ChatGPT's search results. Perplexity went a different direction entirely — launching a Publisher Program focused on revenue sharing rather than optimization guidance. Anthropic (Claude) has published nothing comparable at all. Instead, the entire GEO conversation has been led by research firms testing content, driven by a kind of weird fear from organizations and agencies. When content and information exchange is crucial to our business, we start to get anxious about how it's being handled, and new tools and techniques lead to panic. Which means these third-parties (agencies, tool vendors, independent researchers) are hitting the black box with hammers, trying to figure out how they work from the outside.

So it's not really an accident that Google has opened up the books. Because Google's AI features depend on the open web — and, because AI Overviews and AI Mode pull from existing web content in real time — they have a desperate need for good, crawlable, structured content. If publishers stop creating good content, Google's AI features lose their source material.

So Google publishing this now, finally, after years, is a reassurance play: keep doing what you're doing, and we'll make it work.

Those other platforms? They don't have the same pressure, and, for all we know, they don't even use the same model. Maybe these platforms don't want to formalize things that might change with the next model, or maybe they just benefit from the ambiguity — if nobody knows exactly how to optimize for ChatGPT, everyone keeps producing content and hoping for the best. It might even mean they simply don't depend on the open web ecosystem the way Google does.

One note on the narrative of this entire thing: by defining GEO as "just SEO," Google hopes to pull the entire conversation back into their yard. If the industry accepts that framing, then GEO budgets and attention stay focused on Google's ecosystem, but if GEO becomes its own discipline — distinct from SEO, spanning multiple platforms — that shifts spending and strategic attention away from Google.

This will keep changing.

Google's new documentation is an important point in our understanding of how AI-assisted search works. But it's a single data point — a significant one, from one company with a vested interest in how the conversation develops.

Other platforms will eventually follow. Spurred by Google or pushed by their own needs, they'll publish guidance, or their approaches will become clear enough through testing that we piece it together ourselves. Models will adapt. New features will launch. Documentation will evolve. Google's own guide nods toward this — their section on "agentic experiences" signals that even Google sees the current moment as early days.

That's normal. This is the web! What won't change is the underlying work: write clearly, build from real expertise, structure content for the person reading it. Be specific, take a position, and say something only you could say. That's what people want — and it turns out that's what Google's AI tools want too.