Google Says There Is No Such Thing as GEO. It Is Mostly Right.

Google published its first consolidated guidance on optimising for generative AI features in search on 15 May 2026, and its position has not softened since. There is no separate discipline. What people are selling as GEO or AEO is, in Google's framing, foundational SEO applied to a different surface.

The specifics are blunter than the summary. Google says llms.txt files can be ignored, that they are treated like any other text file with no special indexing pathway. It says there is no need to break content into small pieces for AI systems, because its systems can understand multi-topic pages and extract the relevant passage without help. John Mueller's framing has been to focus on actual audience behaviour rather than on the mechanics.

The obvious objection

Google is not a disinterested party. It is the company whose AI Overviews removed a large share of the clicks that used to reach publishers, and it has an interest in telling everyone that nothing has changed and no new work is required.

That is a fair thing to hold in mind and it is not a reason to dismiss the guidance, because the guidance happens to line up with what the underlying systems plausibly do. A model deciding which source to draw on is not consulting a manifest file you wrote for it. It is working from whatever it has learned about which sources are reliable on a subject.

Where I was too generous to the tooling

“The honest version of the pitch is that these tools tell you what is happening and cannot tell you what to do about it.”

I have written twice this month about AI visibility, including a recommendation to audit how assistants describe you. I stand by the audit, which costs an hour and reliably surfaces something wrong. I was too soft on the surrounding category.

A large amount of GEO tooling is measurement dressed as strategy. Knowing your mention rate across five assistants is interesting. It is not actionable in the way a rank tracker was actionable, because there is no lever that reliably moves it. The honest version of the pitch is that these tools tell you what is happening and cannot tell you what to do about it.

The tactics that sit underneath the category are weaker still. llms.txt is the clearest case: a file that a meaningful number of consultants charged to implement, which the largest system it was aimed at says it ignores.

The honest version of the pitch is that these tools tell you what is happening and cannot tell you what to do about it.

What genuinely is different

Two things, and they are narrower than the category implies.

The first is specificity. A model needs something concrete to attribute to you. A page that argues a position without a number, a date, or a name gives it nothing to carry, whereas a page containing a specific claim that can be checked is repeatable. That is not a new technique so much as an old standard of writing that ranking never strictly required.

The second is that the description of you sitting in the model's understanding may be badly out of date, and there is no equivalent of a rank check that will tell you. That is the actual asymmetry with search, and it is why the manual audit is worth doing even though the tooling around it is oversold.

The practical position

Write things worth citing, keep the plain description of what you do current and unambiguous, and stop buying tactics aimed at a parsing behaviour that the systems have said they do not have.

Then take Mueller's actual point seriously, which is the least convenient part of the guidance. If your business depends on referred traffic, the thing to reconsider is the dependence, not the optimisation. That is a strategy question rather than a marketing one, and no amount of doing the mechanics well answers it.

See ranking first and invisible in AI answers for the audit, where the AI search losses land for why the dependence is the real risk, and the citation economy for what a model can actually carry. The marketer track covers the rest.