What Google has actually confirmed#
Google's documentation on AI features is short, but clear on the basics (Google Search Central):
- A page is eligible if it's indexed and allowed to show a snippet. There's no submission process and no special markup.
- AI Overviews and AI Mode can use what Google calls query fan-out: the system runs several related searches behind the scenes, across sub-topics, and builds its answer from what comes back.
- Google's normal ranking and quality systems apply, with extra checks on top.
Fan-out explains something that puzzles a lot of site owners: why the page ranking first isn't always the one cited, and why a page that doesn't rank for the search sometimes is. The answer may have been assembled from searches you never see.
The data points the same way. In July 2025, Ahrefs found that 76% of pages cited in AI Overviews also ranked in the top 10 for the same query. By March 2026 the figure was 37.9% (Ahrefs). Ahrefs is careful to say the two aren't a like-for-like comparison (its tracking improved, and Google moved AI Overviews to its Gemini 3 model in January 2026), but the direction is hard to miss. Ranking well gets you considered. It no longer guarantees the citation.
AI Mode is a separate surface again. Ahrefs compared the two across hundreds of thousands of US searches in September 2025: the two gave broadly the same answer around 86% of the time, yet only 13.7% of the pages they cited were the same (Ahrefs). Being cited in one tells you very little about the other.
What Google hasn't confirmed is how the final choice is made. You'll read that citations are picked passage by passage rather than page by page. That's a reasonable theory from practitioners who've studied Google's patents, but Google hasn't said it, so treat any tactic built on it as a bet.
What you don't need, according to Google#
In July 2026 Google updated its guide to optimising for generative AI features in Search (Google Search Central). It's worth reading in full, because it rules out most of what's being sold as AI optimisation:
- llms.txt and other AI text files. Google says Search doesn't use them, so they neither help nor hurt.
- Special schema. There's no schema.org markup specific to AI features. Keep using structured data where it earns rich results, and leave it there.
- Breaking pages into small chunks. Not required. Google says its systems can find the relevant part of a longer page.
- Writing in a particular style for AI. Not needed either.
Independent testing agrees. Ahrefs looked at 1,885 pages that started using JSON-LD schema between August 2025 and March 2026 and compared them with matched pages that didn't. Citations in AI Mode and ChatGPT showed no measurable change, and AI Overview citations fell 4.6%, a drop Ahrefs couldn't pin on the schema itself (Ahrefs). Cited pages are more likely to carry schema, but that's because well-run sites tend to do everything a bit better, not because the markup earns the citation.
"Answer-first" paragraphs and FAQ blocks sit in the same place. If a direct answer near the top helps your readers, write one. Just don't treat it as a switch for AI citations, because nobody has shown that it is one.
What Search Console now shows you, and what it still hides#
For the first time, Google gives site owners some direct data on this. The Search Generative AI performance report started rolling out in June 2026 and reached all sites by 31 August (Google Search Central). It shows how often your pages appeared in AI Overviews and AI Mode, by page, country, device and date.
It's useful, and it has real gaps:
Impressions only. Clicks from AI features still sit inside the normal Web figures in the Performance report, mixed in with everything else.
One bucket. AI Overviews and AI Mode are combined, not shown separately.
No queries. You can see which page appeared, but not what the person searched or which background search found it.
UK site owners have a second development to know about. On 3 June 2026 the Competition and Markets Authority required Google to give publishers an effective way to stop their content powering AI features in Search, and to attribute it properly with clear links when it is used (GOV.UK). The result is a Search generative AI setting in Search Console that excludes a whole property from AI Overviews, AI Mode and Discover's AI features. Opting out isn't used as a ranking signal elsewhere, but you lose any impressions and traffic those features would have sent. For now the setting covers the whole site; page-level controls are due by March 2027 (Search Engine Journal).
- 3 Jun 2026CMA requires Google to offer publishers an opt-out
- Jun 2026Generative AI report starts rolling out
- 31 Aug 2026Report reaches all sites
- Mar 2027Page-level controls due
How to find the pages that are actually losing clicks#
No report tells you "AI Overviews cost you this many clicks". You have to work it out, and the honest conclusion is always "consistent with", never "proven". This is the method I'd use:
Step 5 needs repeating for a reason. Researchers who ran 11,500 queries in December 2025 found AI Overview sources were markedly less stable than ordinary results. On a scale where 1 means an identical list, running the same search twice under the same conditions scored 0.69 for the Overview's sources against 0.86 for the normal results, and switching between mobile and desktop dropped the Overview figure to 0.53 (Grossman et al., 2026). That was US data, but there's no reason to expect the UK to be steadier. A single check tells you about that moment and nothing more.
Decide what each affected page is for#
Once you know which pages are exposed, resist the urge to rewrite them all. Sort them by the job they do for the business:
| If the page... | Then... |
|---|---|
| Answers a quick question and has never led to an enquiry | Stop investing in it. Keep it accurate if other pages rely on it, and merge it if it duplicates something stronger. A rewrite won't bring back a click Google has decided to keep. |
| Answers a question with an obvious next step | Build that next step: a calculator, a comparison, a template, a price guide, a checklist. Give people something to do on the page, not just something to read. |
| Helps someone choose (comparisons, "best", "vs") | Worth competing for, both the click and the citation. First-hand testing, current prices, honest trade-offs and a stated method are what a summary can't fully carry. |
| Sells, books or brings in enquiries (service, product and local pages) | Protect it. Don't strip out proof, pricing or calls to action to make it easier for a machine to quote. Conversions come first. |
Why compete for citations at all on the middle rows? Seer Interactive's 2026 update, covering 53 brands and 2.43 billion impressions, put organic click-through rate at about 3.3% on searches with no AI Overview, 2.1% when an Overview appeared and cited the brand, and 0.9% when an Overview appeared without it (Search Engine Land). Being cited won't take you back to where you were. Not being cited is considerably worse. The study is observational, so part of that gap may simply be stronger brands winning both, but on searches that still lead to sales it's a gap worth trying to close.
Where small sites fit in#
The falling overlap between citations and top-10 rankings sounds like an opening for smaller sites, and partly it is: a page doesn't have to win the main search to be used for one of the questions behind it. But nobody has yet compared small and large sites like for like, so there's no evidence of a small-site advantage either.
What a small specialist can realistically do is be the best source for something narrow: a test you ran, a local detail, a price list you keep current, a process you know better than any national publisher. One asset like that is worth more than a dozen summary articles, and it's the same principle behind why most SEO advice doesn't fit your site until you adapt it.
Tracking it without fooling yourself#
Third-party "AI visibility" scores are tempting, but treat them as samples. Google's own guide points out that no third-party tool has access to its internal ranking or AI systems, and answers shift with device, location and day. If you want to track this yourself:
- Pick 15 to 20 questions your customers genuinely ask.
- Run each one three times, on the same type of device and in fresh sessions, once a month.
- Record whether you're cited, whether your brand is mentioned, and whether what's said about you is accurate.
- Report the share of runs you appeared in, rather than a simple in or out.
- Read it alongside the generative AI report and your actual enquiries, never on its own.
For turning any of this into a report a business owner can act on, the same principles apply as in interpreting SEO data for reporting: compare like with like, and tie every number to an outcome.
The short version: AI Overviews have changed which clicks are on offer. They haven't changed what makes a page worth citing. Find the pages they affect, decide what each one is for, and put your effort where a click or an enquiry is still there to win.