How to Optimise Your Website for Google's AI Overviews

Until mid-2026, almost everything written about optimising for AI Overviews was inference. Practitioners ran tests, compared notes, and reverse-engineered patterns — and a whole vocabulary of AEO and GEO tactics grew out of that guesswork. Some of it was sound. A lot of it was confident invention, sold at a premium precisely because nobody could check it against a primary source.

That changed on 15 May 2026, when Google published "Optimizing your website for generative AI features on Google Search" in its Search Central documentation, under a new Generative AI fundamentals section. For the first time there is an authoritative document to check claims against — and the honest reading is that it is far less exotic than the market wants it to be. Google's framing is that its generative features sit on top of the same ranking and quality systems as classic Search, so the same fundamentals carry over.

This is genuinely good news for anyone who has invested in doing the basics properly, and awkward news for anyone who has been sold a separate "AI optimisation" line item that mostly consists of new files and markup. If you want the wider strategic picture of how these layers relate, we covered that separately in SEO vs AEO vs GEO — this post is narrower and more practical: what to actually do about Google's AI surfaces specifically.

Strip the guide back to its load-bearing claims and you get a short eligibility chain. A page has to be crawlable, indexed, eligible to appear with a snippet, and then worth quoting on the merits. Everything else in the document is elaboration on that last point.

The first three gates are unglamorous and technical, which is exactly why they are where most sites quietly fail. A page that is blocked, or crawled but never indexed, or carrying a snippet-suppressing directive, cannot be quoted no matter how good the writing is. Google is explicit that the requirement is that a page "must be indexed and eligible to be shown in Google Search with a snippet" — snippet eligibility is not a formality, it is the mechanism.

That has a specific and under-appreciated consequence for anyone who has ever tightened up their preview settings. The `nosnippet` directive, a `max-snippet:0` setting, or a `data-nosnippet` attribute wrapped around your main content will keep you out of AI Overviews while leaving you apparently fine in classic results. If someone added those years ago to stop scraping, they are now a visibility tax you may not know you are paying. Audit them before you spend a rupee on content.

Google's guidance is reassuring on rendering: "Google is able to process content within JavaScript as long as it isn't blocked." It also advises following JavaScript SEO best practices, noting the approach is "generally more complex" — which is a polite way of saying it works until it doesn't, and when it fails it fails silently.

There is an important asymmetry here that the Google document has no reason to mention, and it is one of the strongest arguments for server-rendered or prerendered pages. Googlebot renders JavaScript. Most other AI crawlers do not. So a client-rendered single-page app can be perfectly visible to Google's AI surfaces while being effectively blank to the crawlers behind other assistants. If your visibility strategy stops at Google, client-side rendering is survivable; if it doesn't, it is a hard ceiling.

This is not a theoretical concern for us — it is why every route on this site is prerendered to static HTML, and why "does this page render with JavaScript disabled?" is a release gate we actually enforce rather than a nice-to-have. The same work also happens to fix your Core Web Vitals, which is a page-experience factor Google's guide names directly. We wrote up the measured results of that engineering in why your website feels slow on mobile.

The single most useful mechanical detail Google has disclosed about these surfaces is the technique it calls query fan-out. Rather than matching one query against an index once, AI Overviews and AI Mode issue multiple related searches across subtopics and data sources, then synthesise an answer from what comes back. Google says this lets the features "display a wider and more diverse set of helpful links" than a classic results page.

Think about what that does to the old model. If a single user question silently becomes a dozen sub-questions, then the page that gets quoted is not necessarily the one that best matches the original phrasing — it is the one that best answers whichever sub-question it was retrieved for. That rewards genuine topical depth and punishes pages that were engineered to match a phrase.

The practical translation is that thin, keyword-targeted pages are now poor value, and one substantial page that actually covers the subtopics around a question is worth several shallow ones. It also means you can earn a citation for a question you never explicitly targeted, which is exactly what we see in practice: the sub-question you answered thoroughly in passing is often the one that gets you quoted.

Note what Google says this does *not* require. You don't need to write in a special style, you don't need to manufacture pages for every phrasing variation, and you don't need to worry about capturing every long-tail permutation — the systems understand synonyms and general meaning. Manufacturing a page per variation to catch them all is not just unnecessary, it is the behaviour Google names as a scaled content abuse violation.

This is the section worth taking to your next agency call. Each of the tactics below is currently being sold somewhere as AI search optimisation, and each is addressed directly in Google's guidance. To be precise about scope: these statements are Google's, about Google's surfaces. Some of these files and formats may have value with other engines or for your own tooling — but if the pitch is "this gets you into AI Overviews," the primary source disagrees.

With the noise cleared, Google's actual recommendations are unfashionably plain — and much harder to fake, which is precisely why they work as a differentiator. The guidance keeps returning to one idea: content that could only have come from you.

It contrasts "a first-hand review [that] provides a unique perspective based on personal experience" against "a summary of existing content [that] simply restates information already available elsewhere," and it warns against commodity content — its own example is an article like "7 Tips for First-Time Homebuyers" — built on common knowledge. The instruction is blunt: "Don't just recycle what others on the internet have already said, or could easily be produced by a generative AI model."

Read that last clause again, because it is the strategic centre of the whole document. If a language model could have written your page from its existing knowledge, you have given the system no reason to retrieve or cite it. Your defensible material is the stuff a model cannot generate: your own numbers, your own tests, your own client work, your own prices, the mistakes you made and what they cost.

Until recently the honest answer was that it barely did. That has improved: on 3 June 2026 Google announced Search Generative AI performance reports in Search Console, giving dedicated views of impressions within generative AI features on Search — AI Overviews and AI Mode — as well as generative AI features in Discover. Google noted it was rolling these out to a subset of websites first, to test and gather feedback before wider availability.

Two things about that report are widely misunderstood. First, the dedicated views are impressions-focused — they tell you that you appeared, not how often people clicked through, so treat them as a visibility signal rather than a traffic report. Second, this is not a separate universe of data: Google states the generative-AI data is also included in the overall performance report, where it continues to be tracked as part of your site's total Search visibility. So if you have been looking at your main Search Console numbers all year, AI-surface activity was already blended into them.

What Search Console will never show you is everything happening outside Google. If your buyers are asking ChatGPT or Perplexity, no amount of Search Console analysis will reveal it — that needs a different measurement stack entirely, which is the subject of our companion post on how to track AI search visibility.

This is the sequence we actually run for a client site, ordered by how often each step turns out to be the real problem. It is deliberately boring at the top — in our experience the technical gates account for most of the surprises, and no amount of content investment can compensate for failing one of them.

Most of the checklist above is engineering, not marketing — which is exactly why it so often goes undone. It sits in the gap between an SEO consultant who cannot change your rendering strategy and a development team with no reason to care about snippet eligibility. We handle both sides because we build the software as well as the visibility strategy, so a finding like "your service pages are invisible without JavaScript" turns into a fixed release rather than a recommendation in a PDF.

Our AI Search Optimization work covers the technical audit, the content architecture, and the measurement setup — and we are straightforward about pricing. A focused AI-search audit and remediation plan for a typical business site starts from about ₹25,000 (~$350); ongoing content and visibility work is scoped per engagement after a free call, on a fixed written quote with no hourly meter. We would rather tell you the four gates are already fine and that your real problem is commodity content than sell you a file Google has publicly said it ignores.

One honest caveat worth stating plainly, because it is the thing most likely to disappoint: none of this is fast. AI surfaces are built on the same ranking and quality systems as classic Search, so they inherit the same timescales. Expect months, not weeks — and treat anyone promising otherwise the same way you would treat a guaranteed number-one ranking.