What Google's own documentation says about optimizing for AI

That is not an opinion I formed. It is Google's documentation, published for site owners, in plain language, free to read.

The reason it matters to you is money. A whole product category has grown up around the idea that AI search is a separate discipline requiring separate spend. If the company that operates the system says otherwise, you should know that before you sign anything.

What does Google actually say about AI Overviews and AI Mode?

That they are built on core Search ranking. Both features retrieve from the Search index and use what Google calls query fan-out, described in the documentation as "a set of concurrent, related queries generated by the model." The system breaks a question into several related searches and assembles an answer from those results. The ranking systems underneath are the ones you already work with.

The sentence worth memorizing is Google's own: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO."

The requirement Google states is short. A page must be indexed and eligible to be shown with a snippet. There are no additional technical requirements. That is the whole technical specification, and it fits in two sentences.

Google follows it with an instruction most vendors do not quote back to you: "Prioritize effective SEO strategies over 'AEO/GEO hacks.'" The acronyms appear in quotation marks in Google's own text.

What are AEO and GEO, then?

They are names for the same work, aimed at systems that answer questions instead of listing links. AEO stands for answer engine optimization. GEO stands for generative engine optimization. Neither is a separate technical discipline, and no vendor controls either term.

What is genuinely different is format, not infrastructure. A page that answers the question in its first three sentences is easier to quote than one that warms up for four paragraphs. A self-contained paragraph survives being lifted out of its page. A clear, plainly stated definition gets reused.

That is craft, and it is worth doing. It is also free. It does not require a new tool, a new file, or a new retainer line item.

Are your pages even eligible?

Does structured data get you into AI answers?

No. Google states directly that structured data is not required for its generative AI features and that there is no special schema.org markup for them. If someone is pitching you an AI schema package, they are selling a thing that Google says does not exist.

Structured data still earns rich results in ordinary search, and that is a real reason to implement it. Two facts from the same documentation set are worth carrying into that conversation. Google's current rich results gallery contains no FAQ guide and no HowTo guide. FAQ rich results were restricted to well-known authoritative government and health sites in August 2023, and HowTo results were deprecated the following month. And Google's non-negotiable rule for anything you do mark up is "making sure your structured data matches the visible text on the page."

There is a genuine complication here and it belongs in an honest article. Microsoft says something different about its own surface. Fabrice Canel, a principal product manager at Microsoft Bing, confirmed at SMX Munich in March 2025 that schema markup helps Microsoft's language models understand web content, and recommended pushing updates through the IndexNow API because generative systems value fresh content.

Those two positions are not actually in conflict once you separate the mechanisms. The indirect path is that search engines ingest structured data at index time to resolve what your business is, and that resolved understanding feeds the answer layer. That is well supported and consistent with Canel's statement. The direct-parse path, meaning a model reading your markup and citing you because of it, has no credible published test behind it. Vendors sell the second and cite the first.

What does Google name as ineffective?

Four things, by name.

llms.txt and similar special files. Google's Mythbusting section says it plainly: you do not need to create new machine readable files, AI text files, markup or Markdown to appear in Google Search, and doing so "will neither harm nor help your site's visibility or rankings in Google Search, as Google Search ignores them."

Artificial content chunking. Breaking your page into machine-shaped fragments does not help. Write for a person and the chunking takes care of itself.

Rewriting content for AI systems. There is no separate AI dialect. Content written to please a model reads worse to humans and performs no better.

Seeking inauthentic mentions. Paying for placements, seeding fake discussion threads, and manufacturing brand mentions are called out specifically. Real mentions still count, which is why being written about by small local publications does more than a purchased mention farm ever will.

Independent data supports the first one rather than merely echoing it. Ahrefs examined every domain in its Web Analytics product with traffic in May 2026, 137,210 sites, and found 97% of published llms.txt files received zero requests in that month from any bot or human. Ahrefs sells SEO software and states its own figures are ceilings rather than floors.

Where does Google contradict itself, and does it matter?

In one place, and it is worth knowing because a vendor will use it on you.

Google Search says skip llms.txt. Chrome shipped a Lighthouse audit that checks for one. Pressed on that inconsistency by Lily Ray, John Mueller of Google said the Lighthouse audit is "not done for search" and described the file as a "temporary crutch, perhaps to save some tokens" for AI coding tools reading developer documentation.

His broader observation is the more useful sentence: "AFAIK none of the AI services have said they're using LLMs.TXT (and you can tell when you look at your server logs that they don't even check for it)."

Read that as an instruction rather than an opinion. Mueller is telling you the answer is in your own logs. That is the same method the Ahrefs study used, and it is available to you for free.

The contradiction is real and it is narrow. One Google product ships a developer-tooling check. Google Search ignores the file. Both statements can be true, and neither of them is an argument for buying one.

What does Google publish about how you are doing?

Less than you would expect, and the omissions are the honest part.

Since June 2026 Search Console has carried Search Generative AI performance reports, described by Google as dedicated views of impressions within generative AI features such as AI Overviews and AI Mode.

What you get: impressions, pages, countries, devices and dates. What you do not get: clicks, click-through rate, position, and the user's prompt. AI Mode and AI Overviews are not separated from each other. And it is a view onto data already inside the overall performance report, so it must not be added to your existing totals.

That list is the ceiling for everyone. No third-party tool has better access than the platform's own reporting, which is why Google attaches its own warning: "Be wary of third-party tools that promise ranking success or claim to use 'internal' Google metrics. No third-party tool has access to our internal ranking or AI systems."

Where the documentation deserves scepticism: Google is a self-interested narrator, and the guidance conveniently keeps you inside the practices that serve Google. Its negative claims hold up because outside log data agrees with them. Its silences are the part to watch. It does not tell you what share of impressions convert, it does not separate its two AI surfaces, and it does not publish anything about how sources are chosen.

What does this mean for a local business specifically?

Google addresses local directly. Its generative responses can include information about local businesses, and Google states that a Business Profile helps your services appear "in both AI responses and other Google Search results." The same documentation points at Business Agent, Google's own layer for business information inside these surfaces. That is unusually plain language from a company that normally hedges.

So the AI search work that pays for a local operator is a correctly configured Business Profile, accurate details that agree everywhere the web mentions you, and an active flow of reviews. Whether your name, address and phone still need to match exactly across the web is a fair question with a more nuanced answer than the old advice suggested, but the direction has not changed.

One thing that does not work, tested and published: stuffing keywords into the reviews you request. A controlled test found keyword text inside reviews does not move map pack rankings. The review itself matters. The words you coach into it do not.

What to do this week

1. Read the Google guidance yourself. It is one page, it takes fifteen minutes, and it will make you a harder person to sell to. 2. Run the three-number eligibility audit above. Submitted, indexed, snippet-eligible. Fix the gap before you fix anything else. 3. Open your three most important service pages. Check whether each answers its core question in the first three sentences. Fix the ones that do not. 4. Check your markup against what is on the page. Anything marked up but not displayed is a violation, not a shortcut. 5. Take any AI search proposal you have received and mark each line item as ordinary SEO, genuinely new, or named as ineffective by Google. Most lists collapse. 6. Before you measure anything, understand why a single AI visibility score is not a real number. Buying the wrong metric is how good work gets cancelled.

Questions about what you are being sold as AI search?

Send me the single AI search tactic you were told you needed, in one line, to eric@seod.com. I will tell you where Google's documentation actually lands on it and quote the relevant part back to you. No charge, no pitch, and if the tactic is legitimate I will say so.

I offer this because the documentation is public and almost nobody reads it, which lets the category price vagueness. Reading it on your behalf takes me five minutes.

Otherwise keep reading the rest of the AI search library.