What Google actually says about AEO and GEO

There is a guide on Google’s own developer site called Optimizing your website for generative AI features on Google Search. It was last updated on 10 July 2026. Almost nobody selling AEO or GEO services quotes it, and reading it explains why.

The three sentences that matter

On machine-readable files, under a heading that literally reads “things you can ignore for Google Search”:

You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them.

On structured data:

Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.

On the category itself:

From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.

That last paragraph ends by pointing the reader at Google’s guidance on evaluating third-party SEO advice. Google does not put that link there by accident.

What the measurements say

Vendor claims are one thing. Two studies with published methods say the same as the guide.

llms.txt. Ahrefs looked at 137,210 domains with traffic in May 2026. Twenty-eight per cent published a valid llms.txt. Ninety-seven per cent of those files received zero requests in the month measured. Of the few that were fetched, the largest single category was SEO audit tools, at 21.7%; AI retrieval bots accounted for 1.1%. Requests for missing llms.txt files came from no AI bot at all, which tells you nothing is out there looking.

Structured data. The same team ran a matched-control experiment: 1,885 pages that added JSON-LD between August 2025 and March 2026, matched against 4,000 control URLs on different domains with similar prior citation counts, measured 30 days either side. AI Overviews fell 4.6%, significant at roughly one in 2,500 by chance. AI Mode rose 2.4% and ChatGPT 2.2%, neither significant. The widely repeated statistic that cited pages are three times more likely to carry JSON-LD is the correlation this experiment was designed to test, and it did not survive.

I have a stake in saying this clearly: I build structured data into every site I ship, and this site’s JSON-LD is a single @graph per page with no dangling references, because a build check fails if there is one. I keep doing it because it is cheap, it keeps an identity unambiguous, and it earns rich results where those still exist. I do not sell it as the thing that gets you cited, because the only controlled test anyone has published says it is not.

Why the bolt-ons sell anyway

Because they are the only part of this that can be done to a website in an afternoon and invoiced.

The things that do move AI visibility are slow and partly outside your control. Assistants corroborate: they lean on the places that talk about you, not the place where you talk about yourself. One of the more honest studios in this market, Space & Story, puts it on their own services page: most AI citations come from third-party sources, not your own site. That is a difficult sentence to put next to a price.

So the market sells the fast half. A file you can upload, a block of markup you can paste, a “GEO audit” that produces a map. A survey of 60 agencies in September 2026 found 13% promised a result or a refund, 32% stated plainly that results are not guaranteed, and 50% said nothing at all about whether their work would work.

What is left, once you take the bolt-ons away

Google’s guide is unusually specific about this, and its examples are worth reading rather than paraphrasing. It contrasts commodity content, its example being “7 Tips for First-Time Homebuyers”, with non-commodity content, its example being “Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line”. One of those can be produced by anyone or by a model. The other requires having done the thing.

That is the whole instruction, and it is not a trick. If your page restates what is already on the web, there is no reason for an assistant to prefer it over the forty other pages that restate the same thing. If it contains something only you could write, because you measured it or you lived it, then it is the page worth citing.

The rest is hygiene, and hygiene is still worth having:

• The page must render without JavaScript, or a crawler that does not execute scripts gets an empty shell. This is the single most common fault I find, and it is invisible in a browser.
• It must be crawlable and reasonably fast, because retrieval happens before any of this matters.
• Your name, description and category should be the same everywhere a machine can read them, so mentions in different places resolve to one entity instead of scattering.
• Structured data, because it is cheap and it makes the identity unambiguous. Budget it as hygiene, not as a lever.

None of that is a new discipline. Google says as much in a sentence.

The honest version of the sale

If someone offers to make you visible in AI search, three questions sort the field quickly.

What exactly will you ship, and will it be live? The gap in this market is not diagnosis. Tools diagnose perfectly well and cost under $300 a month. The recommendations then die in a development queue. Ask who deploys.

What do you say cannot be controlled? Anyone who does not volunteer the third-party-citation problem either does not know it or is hoping you do not.

What is the baseline and when is it remeasured? A fixed set of prompts, recorded before anything changes, rerun on a schedule. Without that first measurement nobody can tell you afterwards whether anything happened, which is convenient for everybody except you.

I would rather lose the work than sell a file that 97% of the internet is not reading.