A large amount of money is being spent on answer-engine optimization services built on claims nobody has been able to demonstrate. So when the platform in question publishes documentation saying which of those claims are wrong, that document is worth reading closely, even allowing for the fact that Google has an interest in what it says.
Google's guide to optimizing for generative AI features went up in May 2026 and has been updated since. Its headline is that AEO and GEO, from Google Search's perspective, are still SEO, because the generative features are built on the core Search ranking and quality systems.
The part that is genuinely informative
Google names the two mechanisms. The first is retrieval-augmented generation, which it also calls grounding: the core ranking systems retrieve relevant pages from the Search index, and the model builds a response from what those pages say. The second is query fan-out, a set of concurrent related queries the model generates to gather more material. Google's own example is a question about a weedy lawn producing side queries about herbicides and weed prevention.
Both of those are descriptions of plumbing, not instructions. But they establish something concrete: to appear, a page has to be indexed, eligible to show with a snippet, and included in the Search generative AI control. That is a real eligibility list, and it is short.
The mythbusting list
Google says you can ignore llms.txt and similar machine-readable files, because Google Search does not use them. It says content chunking is not required, and there is no ideal page length. It says you do not need to rewrite content for AI systems, because they understand synonyms and meaning. It says seeking inauthentic mentions is less helpful than it appears. And it says structured data is not required for generative AI search, while remaining worth keeping for rich results.
It also warns against building a page for every fan-out variation, and points at the scaled content abuse policy when it does so. That is the sharpest line in the document, because producing a page per query variation is precisely what a lot of AEO tooling encourages.
On measurement it is equally direct: no third-party tool has access to Google's internal ranking or AI systems, so any tool claiming internal metrics is claiming something that does not exist.
Where we would not just take Google's word for it
Two things sit uneasily. The first is that a platform telling publishers no change in behaviour is required is also a platform with no incentive to describe a change in behaviour that would cost it content. The guidance is accurate about mechanism and self-interested about implication.
The second is scope. All of this is about Google Search. ChatGPT, Copilot and Perplexity retrieve differently and cite differently, and nothing in a Google document constrains them. A business that reads this guide and concludes that answer engines in general need no attention has read a document about one of them.
Our reading is narrower than either the guide or its critics: the tactics on the ignore list really can be dropped, and the effort should move to the two things the guide is quietly firm about, which are content nobody else could have written and a site that is technically legible.
What Google says good looks like
The content section is more specific than it first appears. It contrasts commodity content, and gives an example of a generic list of homebuyer tips, with non-commodity content, and gives an example about waiving an inspection and what the sewer line turned out to be.
That is a distinction most service businesses can act on immediately. The generic version of your expertise is already in the model. The specific version, the case you actually handled and what it cost, is not.
What this means for an operator
Stop paying for the tactics on Google's ignore list and move that budget into content only your business could produce. Then verify the boring eligibility items: indexed, snippet-eligible, and not excluded in Search Console.