On 15 May 2026 Google published its first standalone guide to appearing in AI Overviews and AI Mode, and most of it tells you what to stop doing.
The document is called Optimizing your website for generative AI features on Google Search. John Mueller announced it on the Search Central blog the same day.
It runs to about 2,000 words and has been revised since: a note on llms.txt files went in on 15 June, according to Google's documentation changelog, and the page now carries a 10 July last-updated stamp.
We read the whole thing against what agencies and tool vendors have been selling as "AI Overviews SEO" for two years. This article covers what Google says, what it does not say, and what a UK business with a modest budget should do about it.
What the guide covers
The guide covers AI Overviews and AI Mode on Google Search, and nothing else. It names both features in its opening paragraph and describes how they work: retrieval-augmented generation, which Google also calls grounding, pulls pages from the normal Search index, and query fan-out fires off several related searches at once to widen the pool.
The fan-out example Google gives is "how to fix a lawn that's full of weeds", which might spawn "best herbicides for lawns" and "how to prevent weeds in lawn". For a small business that means a page about one narrow service can be pulled into an answer to a much broader question.
What Google says you should do
Google's main instruction is to produce content that could not have been written by anyone else. The guide calls this "non-commodity content" and says it will "likely influence your website's presence in generative AI search in the long run more than any of the other suggestions in this guide".
The example is pointed. "7 Tips for First-Time Homebuyers" is commodity content because the information "could originate from anyone". "Why We Waived the Inspection and Saved Money: A Look Inside the Sewer Line" is not, because it comes from someone who was there.
The line worth pinning above a content marketer's desk is this one:
"Don't just recycle what others on the internet have already said, or could easily be produced by a generative AI model."
Google Search Central, Optimizing your website for generative AI features, May 2026
Beyond that, the content advice is familiar. Organise pages with headings people can follow and add images and video where they help. Do not spin up a page for every fan-out query you can think of; Google says doing so "primarily to manipulate rankings or generative AI responses" breaks its scaled content abuse spam policy.
The technical section is equally plain. A page must be indexed and eligible to show a snippet in ordinary Search. That is the whole requirement.
There is one extra condition that is easy to miss: the site must be "included in Search generative AI features" in Search Console. If someone on your team switched that control off last year, your pages will not appear in AI Overviews regardless of quality.
For local and ecommerce retailers, the guide says Merchant Center feeds and a maintained Google Business Profile can get products and services into AI responses. For a Northampton plumber or a Milton Keynes shop, that is probably the most useful sentence in the document.
What Google says you can stop doing
The guide has a section headed "Mythbusting generative AI search" that dismisses five tactics by name. Here they are alongside the pitch each one usually arrives with.
| Tactic being sold | What Google says | Typical pitch |
|---|---|---|
| llms.txt files and "AI markup" | Not needed. Google Search "doesn't use them". The 15 June note adds they do not affect rankings. | "Make your site readable to AI" |
| Chunking content into short blocks | "No requirement to break your content into tiny pieces." No ideal page length. | "AI needs bite-sized answers" |
| Rewriting pages for AI, chasing long-tail variants | Not needed. Systems understand synonyms and meaning. | "Answer-first formatting for AI Overviews" |
| Buying or seeding "mentions" across the web | "Isn't as helpful as it might seem." Spam systems apply. | "Brand mentions are the new backlinks" |
| Adding special schema for AI | "There's no special schema.org markup you need to add." Keep it for rich results. | "AI-ready structured data package" |
The guide is direct about the vocabulary too. It says that while terms like Answer Engine Optimisation and Generative Engine Optimisation "are common online, many suggested 'hacks' aren't effective or supported by how Google Search actually works". If you have paid for any of the five items above as a line on an invoice, ask the supplier what evidence they had that it worked. The independent evidence is just as thin: a survey of 45 generative engine optimisation studies found no tactic with a proven, cross-platform effect.
What the guide leaves out
The guide says nothing about how sources are chosen once pages have been retrieved. It describes the retrieval step (core ranking systems pull candidates from the index) and the display step ("prominent, clickable links to relevant web pages that support the information in the response"). The bit in between, where a model decides which of perhaps twenty candidates gets quoted, is not described.
That gap is the question every business owner asks. Ranking fourth used to mean a predictable share of clicks. Appearing in the AI Overview is binary, and the guide does not say what tips a page from one side to the other.
The guide covers only Google. ChatGPT, Perplexity, Claude and Microsoft Copilot each retrieve and cite pages in their own way, and Google's advice on llms.txt or chunking is scoped to Google Search. The June note concedes that keeping an llms.txt file "for other services is fine".
Ofcom counted 1.8 billion UK visits to ChatGPT in the first eight months of 2025, so the other engines are not a rounding error.
Measurement is thin. The guide points to the Generative AI performance report in Search Console, which Google launched on 3 June 2026 for a subset of sites. That report shows impressions, pages, countries, devices and dates. It does not show clicks or queries. You can see that a page appeared inside an AI feature; you cannot see what the person typed or whether they visited.
There is no guidance on click behaviour. The closest Google gets is a line on a separate AI features page saying clicks from results with AI Overviews have been "higher quality". The only independent data we have found is from Pew Research Center, which tracked 68,879 Google searches by 900 US adults in March 2025.
xychart-beta
title "Share of visits where the user clicked a result (Pew, US, March 2025)"
x-axis ["Standard result, no AI summary", "Standard result, AI summary shown", "Link inside the AI summary"]
y-axis "Per cent of visits" 0 --> 20
bar [15, 8, 1]
In Pew's sample, users clicked a standard result on 15 per cent of pages without an AI summary and 8 per cent with one. Links inside the summary were clicked on about 1 per cent of visits. That is US data and over a year old. Google has not published anything to contradict it.
Nor does the guide say how often AI Overviews appear in the UK. Ofcom's Online Nation report, published in December 2025, put it at about 30 per cent of UK searches, with 53 per cent of adults saying they see the summaries often.
A Peec AI study of 500,000 business-intent queries in April 2026 found an AI Overview on 91.7 per cent of UK commercial searches. The two figures measure different things, all searches against commercial ones, but together they say the feature is now the norm for the queries a business cares about.
Where that leaves a UK small business
A UK business should read the guide as permission to stop spending on AI-specific SEO products and put the money into the two things Google named: original content and a clean site. This is the order we would work in.
- Open Search Console and check the "Search generative AI features" setting is on. Two minutes, and it is the only hard gate in the document.
- Run the URL Inspection tool on the ten pages that bring in enquiries. If a page is not indexed or has no snippet, nothing else matters.
- Audit your last twelve blog posts against Google's commodity test. For each one, ask whether a competitor or a chatbot could have written it. Most "5 reasons to choose a local accountant" posts fail.
- Replace the failures with pages only you could write: a real job with real numbers, a mistake you corrected, a comparison from your own purchase records, a pricing explanation for your own area.
- Keep your Google Business Profile and, if you sell products, your Merchant Center feed current. Google named both as inputs to AI responses.
- Cancel or renegotiate anything billed as llms.txt setup, AI schema, content chunking or mention seeding. Google has said in writing that none of it affects Google Search.
- Record a baseline now: Search Console impressions and clicks for your key pages, plus AI impressions once the generative AI report reaches your property. You will not get query data, but you will see direction.
Steps 1 and 2 are technical SEO and belong in any site review. Steps 3 and 4 are the hard part, because someone inside the business has to spend time on the content.
One thing the guide does not resolve is the trade-off between being quoted and being visited. If Pew's figures hold, a page that appears in a summary earns fewer clicks than one ranking under no summary at all.
For some businesses a brand mention inside the answer is still the right goal. For others, a page that ranks for queries with no AI Overview may be worth more. The guide gives you no way to tell which queries those are, so you have to check the results pages yourself.
Reading the guide against the industry
The guide reads as a correction to the market. Semrush's write-up of the launch noted that much of it consolidates things Google staff had already said at conferences and in interviews. What changed in May is that the statements now sit in official documentation with a date, which means they can be quoted back to anyone selling the opposite.
Three weeks later, on 5 June, Google added a page on third-party SEO tools and advice to the same documentation set, per the changelog. The timing looks deliberate.
An "AI visibility" dashboard can tell you whether your domain appeared in a sample of AI responses on a given day. It cannot tell you why, and it cannot see what Google's retrieval step saw. Use them to spot a change, then look for the cause somewhere else.
There is a risk in over-reading it, though. Google says you do not need llms.txt for Google. It does not say the file hurts you. Google says chunking is not required. It does not say long, well-structured pages are penalised. The guide removes obligations; it does not create new ones.
The most useful way to hold the document is as a test for any proposal that lands on your desk. If a supplier recommends something the guide says Google ignores, ask which engine it is for and how they will show it worked there. If the recommendation is about content nobody else could produce, or a site that indexes cleanly, it is in line with Google and worth doing anyway.
If you want a second opinion on whether your pages pass Google's commodity test, or what the Search Console setting and index coverage look like for your site, our AI Overviews optimisation team can run that check. Get in touch and tell us which pages bring in the enquiries.