Skip to content
Have a project in mind?
The Gloria JournalSEO

Optimising your site for AI search: what works

Screenshot of a dashboard measuring a site's visibility in AI search engines

Your next customer may put their question to ChatGPT, read a Google AI Overview or use AI Mode before opening a single website. That changes where the first impression happens, and what you can measure afterwards.

Published figures on this shift are scarce, and most of them are national. In France, the Baromètre du numérique 2026 (Crédoc, for Arcep, Arcom, the CGE and the ANCT) reports that 48% of French respondents say they use a generative AI tool, that looking for information is the most cited use, mentioned by 73% of those users, and that ChatGPT is the main tool for 63%. The survey ran from 5 to 21 June 2025 among 4,145 people aged 12 and over, in France only. It records stated habits in one country, not traffic to your site, and is not a British or global figure.

Around that shift, advice has multiplied: an llms.txt file, content "chunked for AI", markup said to be decisive. This article separates what Google and OpenAI actually document, as of September 2026, from what remains a working hypothesis, then sets out an order of priorities for a small or mid-sized organisation.

Three surfaces, three sets of rules

Google's AI Overviews and AI Mode, ChatGPT search and the other assistants do not use the same crawlers, the same rules or the same controls. Treating them as a single block is the first mistake: the lever that works on one has no effect on another.

What each provider documents for websites, September 2026
Surface Documented condition of access Control available to the site Published selection criteria
AI Overviews and AI Mode (Google) Page indexed and eligible for snippets, as in ordinary Search robots.txt, nosnippet, max-snippet, noindex, and the Search generative AI setting in Search Console Ordinary SEO good practice, with no additional requirement
ChatGPT search (OpenAI) Access granted to OAI-SearchBot robots.txt, per crawler, applied within roughly 24 hours None published
Perplexity, Claude, Copilot Each publisher runs its own crawlers robots.txt, with exceptions for user-initiated fetches None published

On the Google side, both AI Overviews and AI Mode may use a technique the company calls query fan-out: the system runs several searches on sub-topics of the original question, then composes an answer from those results. A page can therefore be cited for one narrow sub-question without ranking for the query the user typed. Google's documentation on AI features and your website describes the mechanism. Availability differs from one country to the next, so check what is live in the markets you sell to.

On the OpenAI side, the page on its crawlers states that sites which disallow OAI-SearchBot are not shown in ChatGPT search results, although they can still appear as browsing links. That crawler is configured independently of GPTBot, which serves model training. OpenAI publishes no criteria explaining why one page is cited rather than another. For many organisations, four lines of robots.txt settle the question: stay out of model training, stay visible in assistant answers.

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

That is a business choice about your content, not a technical optimisation, and it deserves a deliberate answer.

What Google documents: no special optimisation

Google added an optimisation guide for its generative AI features on 15 May 2026, updated on 10 July 2026. Its position has been stable since 2025: ordinary SEO good practice still applies, and there is no additional requirement for appearing in AI Overviews or AI Mode. Google also confirmed in May 2026 that its spam policies apply to generative responses in Search.

The positive recommendations fit into a few lines. Produce content with a point of view of its own rather than a rewrite of what exists. Meet the technical requirements of Search and stay indexable. Look after page experience on every device. Add images and video where they serve the subject. Feed Merchant Center and your business profile if you sell products or serve a local area. None of that is new, which is the point.

The myths Google names

The same guide has a section on the received ideas circulating under the labels GEO and AEO. Google states there that you need no new machine-readable files, text files for AI, particular markup or Markdown versions of your pages. A note dated 15 June 2026 covers llms.txt: it is not used for Google Search, and neither helps nor harms.

  • Splitting content into small blocks "for the AI" is not required.
  • Writing in a particular style aimed at language models is not required.
  • Structured data is not required for generative AI features, although it keeps its role for ordinary rich results.
  • Chasing artificial mentions on other sites is discouraged.

None of this means that no work is worth doing. It means the useful work is search engine optimisation, applied to the substance of the page with more discipline than most sites manage.

Priorities for a smaller organisation

1. Check access and indexing first

A page that is not in Google's index cannot appear in an AI Overview. Start with the indexing reports in Search Console and a careful reading of your robots.txt. An old block, inherited from a previous agency or added by a plugin, can be excluding OAI-SearchBot without anyone remembering it. Our page on getting pages indexed by Google covers that check.

Snippet controls sit in the same family. To keep a page indexed while limiting how much can be quoted, the directive lives in the page head, not in a new file:

<meta name="robots" content="max-snippet:50">

2. Write pages that answer one precise question

Because the system works through sub-queries, it favours pages that settle a bounded subject: what a price includes, how long something takes, the conditions under which you take a job on, an honest comparison between two options. A service page that stays vague gives nothing to quote. This is not an argument for chopping text into fragments, but for checking that each page carries verifiable facts and a position that is yours. That is the ordinary discipline of SEO content, not a new speciality.

3. Make your business information consistent

Name, address, service area, opening hours and contact details should read identically on your site, on your business profile and in any directory you maintain. An inconsistency costs you an enquiry with or without AI, and assistants compound it by reading several of those sources at once. If you serve a defined area, treat this as local SEO work rather than an AI question.

4. Keep structured data faithful to the visible page

Markup is not required for AI features, but whatever you already have must describe what the page shows. A FAQPage block listing questions absent from the page, or reviews that cannot be verified, falls under practices Google acts on, generative responses included. Removing inaccurate markup is safer than adding more.

Measuring without over-reading

Since 3 June 2026, Search Console has offered a generative AI performance report, rolled out to all sites by 31 August 2026. It shows how often links to your site were displayed in AI Overviews and AI Mode, by page, country, device and date. Two links to the same site in one answer count as a single impression. There are no clicks, no click-through rate and no queries; Google says further metrics will come "over time", without a date.

The same tool holds the Search generative AI setting. Opting out removes your links and content from AI Overviews, AI Mode and the generative features of Discover, usually within one to two days. Google states that this choice is not a ranking signal for the rest of Search, and does not govern model training, which is handled through Google-Extended.

What a website can actually measure, September 2026
Metric Available Limit
Impressions in AI Overviews and AI Mode Yes, in Search Console No clicks and no associated queries
Visits arriving from ChatGPT or Perplexity Partly, through the referrer in your analytics tool Native apps may send no referrer, so those visits are counted as direct
Citation rate per query No official data Third-party tools rebuild samples of answers without access to platform data
Enquiries or sales attributed to AI No A mention, a visit and a customer remain three separate observations

The caveat about referrers comes from Cloudflare, whose June 2025 analysis notes that traffic from the Claude native app carries no Referer header, and that the same is likely of other apps. Your "AI traffic" figures are therefore more likely understated than inflated. Google advises scepticism towards tools claiming access to internal metrics. Connecting this data to commercial outcomes is the job of ordinary traffic management, not of a visibility score.

What is still speculation

Several widely repeated claims have no official source. The percentage of queries said to be "migrating to AI" varies from one consultancy to the next and rests on no data published by Google or OpenAI. The click studies that do exist, including the Pew Research Center work from 2025, use United States browsing data and describe an association, not a loss you can apply to your site. Published lists of "ChatGPT citation factors" come from third-party observation of answer samples, unconfirmed by OpenAI.

So two realities sit side by side. On one side, documented facts: conditions of access, the controls you hold, an impressions report, the absence of any special requirement. On the other, a reasonable hypothesis: a precise, sourced page stands a better chance of being quoted than a generic one. Treating the second as if it were the first is how organisations buy services that promise what nobody measures, which is worth remembering when choosing an SEO provider. Treating the first as negligible is how a forgotten block or an empty page quietly keeps you out of a channel your market already uses.

The practical conclusion is unglamorous. Check what crawlers can reach. Write pages that say something specific. Keep your business details straight. Measure what is reported, and label the rest an estimate.

Common questions

Do I need an llms.txt file to appear in Google's AI answers?

No. Google has stated since June 2026 that llms.txt is not used for Google Search and has no positive or negative effect on visibility. Keep the file if other services read it, but it replaces no work on the pages themselves.

How do I know whether my site appears in AI Overviews or AI Mode?

Search Console has offered a generative AI report since June 2026, counting impressions of your links in those features by page, country, device and date. It gives no clicks and no queries, and has been available for all sites since 31 August 2026.

Does blocking GPTBot remove me from ChatGPT search?

No. OpenAI documents separate crawlers: GPTBot for model training, OAI-SearchBot for display in ChatGPT search. Disallowing GPTBot while allowing OAI-SearchBot is supported, and a robots.txt change is picked up within roughly 24 hours.

Is structured data mandatory for AI search?

No. Google states that no particular schema.org markup is required for AI Overviews or AI Mode. Structured data keeps its value for ordinary rich results, provided it describes the visible content of the page accurately.

Can I remove my site from Google's AI answers without leaving Search?

Yes. The Search generative AI setting in Search Console excludes a site from AI Overviews, AI Mode and the generative features of Discover. Google says this choice does not affect ranking in the rest of Search and does not govern model training.

Is traffic from ChatGPT visible in my analytics?

Partly. Visits from the ChatGPT web interface generally carry an identifiable referrer. Native apps may not send one, in which case those sessions are recorded as direct. Whatever you see is therefore likely to understate the channel.

This link opens in a new tab.