How to Rank in AI Overviews, According to Google

Somebody has probably pitched you on generative engine optimization by now. The deck is always roughly the same: AI Overviews are a new channel, they need their own playbook, and for a monthly retainer they will chunk your content, add an llms.txt file, and go build brand mentions on the sites the models like to quote.

In May 2026 Google announced its own guide to optimizing for generative AI features, and most of that pitch does not survive contact with it. The guide is worth reading in full, but the short version is that the tactics being sold are the ones Google specifically names as unnecessary.

That is not the same as saying AI search does not matter. One thing did genuinely change in 2026, and it is worth your attention. It just is not the thing you are being sold.

The AI Overviews eligibility bar is one you already cleared

Start with the requirement, because it is shorter than you would expect. In Google’s words:

To be eligible to be shown in generative AI features on Google Search, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements.

That is the whole gate. Indexed, snippet-eligible, technically sound. If your pages already rank for anything at all, you cleared it some time ago and nobody told you.

There is no second index, no separate submission, no AI-specific eligibility track. Google documents the AI surfaces themselves in AI features and your website, and the story is the same there. Which means the most productive thing you can do about AI Overviews is the least exciting: make sure Googlebot can fetch, render and index the pages you care about, and that none of them are quietly carrying a noindex or a nosnippet directive. A technical crawl answers that in a few minutes and it is the same crawl you should already be running.

Five things Google says you don’t need to do

The guide has a mythbusting section, and it is the most useful page of SEO writing published this year, mostly because of what it rules out. Every quote below is Google’s.

Read that last row twice, because an entire service category has grown up around it. The pitch is that models cite what the web talks about, so you should manufacture the web talking about you. Google’s position is that manufactured mentions are worth less than they look, and that its spam systems are the reason.

One clarification worth making, since this is where people overcorrect: none of this says structured data is useless or llms.txt is harmful. Schema still earns rich results in regular search, which is a good enough reason to keep it. The claim is narrower and more specific. These things buy you nothing extra in AI features, so a plan built on them is a plan built on nothing.

What the guide says to do instead

The affirmative advice is short, and it is the same advice Google has given for a decade: foundational SEO, plus content that is worth citing. The one line with real teeth in it is this:

Don’t just recycle what others on the internet have already said, or could easily be produced by a generative AI model.

That is a harder brief than it sounds. A model summarizing a topic already has the commodity version of every answer. It has no reason to reach for your page unless your page contains something the commodity version does not: original data, a first-hand account, a number you measured, a position someone would disagree with.

Notice that this is the one piece of advice in the entire guide that cannot be productized. You cannot buy it as a retainer, which is exactly why the market sells you file formats instead.

The thing that actually changed is measurement

Here is the part of 2026 that genuinely matters. In June, Google shipped a Generative AI performance report in Search Console. For the first time you can see impressions from AI Overviews and AI Mode as their own line rather than folded invisibly into your Web search totals.

Take it. It is free and it is the only first-party data anyone has. But go in knowing its three limits, because they define the shape of what you still cannot see.

It counts impressions, not clicks.The report tells you how many times links to your site were shown in an AI feature. It does not tell you whether anybody came. For a surface whose entire design question is “does the answer replace the visit”, that is the number you most want and do not get.

Not everyone has it.Google is explicit that it is “rolling out this report to a subset of website owners”. If it is not in your property yet, there is nothing to do but wait.

It only covers Google. This is the big one. Search Console reports on Google Search and Discover, and that is the correct scope for a Google product. It will never tell you that ChatGPT recommends your competitor when someone asks it to compare tools in your category, or that Perplexity has been citing a three-year-old review of you that gets a key fact wrong.

That third gap is why we built AI Visibilityinto Reachavo: it runs your brand against both Google’s AI surfaces and ChatGPT, reports which prompts mention you, and shows which sources the answers were built from. It measures. It does not claim to optimize, because after reading Google’s guide it is hard to argue there is much to optimize.

Why “ranking” is the wrong mental model

The phrase “rank in AI Overviews” imports an assumption that does not hold. Ranking describes an ordered list with room for ten results. An AI answer names two or three sources and stops.

Go back to the eligibility quote. The bar is indexed and snippet-eligible, not top three. Google set the floor low and said nothing about position being the selector. So the two assumptions people carry over from classic SEO both come apart: holding position one does not entitle you to a citation, and you do not have to hold position one to earn one.

Which changes the question you should be asking. Not “what position am I”, but “across the questions my buyers actually type, how often am I in the answer at all, and who is there instead of me?” That is a presence-and-share question, and your rank tracker was never built to answer it.

What to actually do this quarter

Deliberately boring, in order:

1. Confirm eligibility. Are your money pages indexed, renderable and free of nosnippet? That is the entire technical requirement, and it is the only one with a hard pass/fail.
2. Check Search Console for the Generative AI report.If you have it, note today’s impressions as a baseline. There is no historical backfill to go and find.
3. Find your commodity pages and fix them. Anything a model could have written without you is a page no model needs to cite. This is slow work and it is the only lever the guide actually names.
4. Baseline the surfaces Search Console cannot see. Pick the fifteen or twenty questions a buyer would genuinely ask an assistant before choosing a tool like yours, and find out what the answers currently say.

Nothing on that list is a growth hack, and step three will take a quarter rather than an afternoon. That is roughly the point. When the platform owner publishes a document saying there are no special requirements, the correct response is to stop looking for the special requirements.

The short version

Google’s guide is short, and I would rather you read it than take my summary of it. But if you only take three things: eligibility for AI features is ordinary Search eligibility; the file formats, chunking and mention-building being sold as AI optimization are named in Google’s own documentation as unnecessary; and the only real work is content a model has a reason to reach for.

What is left after you delete the hacks is measurement, and half of it is now free in Search Console. For the half Google will never report on, which is every assistant that is not Google, that is what Reachavo’s AI Visibility reports are for. Run one against your brand and see whether the answers your buyers are getting mention you at all.