AI Overview ranking factors are the signals Google uses to decide which pages to cite in the generated box above the organic results. There is no published list. Google has never released one, and in its official guide to optimising for generative search it rules out, in writing, five practices that are routinely sold as "factors". On Google UK the first page for this query (DataForSEO, Google UK, September 2026) is mostly generic Google ranking factor lists and ChatGPT ranking factor guides, one of which names domain authority among its top three signals; none of the snippets refers to Google's own guidance.
What does exist is a chain of filters: part of it documented by Google, part of it measured by studies of millions of SERPs. Here I rebuild it filter by filter, keeping what is confirmed apart from what is only correlation and from what has been ruled out.
Why "ranking factor" is the wrong term
In the classic SERP, ranking is a league table: ten results in order, and your page takes one row. Inside an AI Overview there is no league table. There is an answer written by a model which, to build it, splits the question into several sub-questions (query fan-out), retrieves passages from different sources for each one and then cites some of the pages it took them from. The model does not rank pages, it selects passages.
The practical consequence: your page can rank first on Google and not appear, or sit on page five and appear. That is not a hypothesis, it is what the data in the next section shows. And it changes how you work, because it moves the goal from the ranking of the page to how easily a single passage can be picked out.
Filter 1: technical eligibility
This is the only part Google describes explicitly, and the only part that depends entirely on you. For a page to end up in an AI feature in Search, four conditions apply, listed in the Google Search Central guide:
- the page must be indexed and eligible for standard Search results;
- the content must be publicly accessible and crawlable, with no paywall or crawler blocks;
- no preview controls that block snippets may be active, such as
nosnippet,max-snippet:0or thedata-nosnippetattribute on the blocks you want read; - the site must be included in generative AI features in Search Console.
The third point is the one that, in practice, shuts out more sites than people think. A tight max-snippet set years ago to protect content from featured snippets now also excludes you from AI citations, and nobody notices until they read the tag. It is worth checking before any other optimisation: if you do not pass this filter, the others do not exist.
Filter 2: organic ranking, which now counts for half as much
For a year the convenient answer was: get into the top 10 and you will be cited. The figure behind it came from Ahrefs, which in July 2025 found that 76% of citations came from the top ten positions. The same study, rerun in March 2026 on a sample more than twice the size, gives a different picture.
| Organic position of the cited page | Share of citations (March 2026) | Share of citations (July 2025) |
|---|---|---|
| Top 10 | 38.0% | 76% |
| Positions 11-100 | 31.2% | not measured |
| Outside the top 100 | 31.0% | not measured |
In six months the link between page one and citation halved. Almost a third of cited sources do not even appear in the top hundred for the original query. A large slice of that third is YouTube: 18.2% of citations coming from outside the top 100 are YouTube URLs, up 34% in six months.
Ranking is still the strongest single signal, but it has become a condition that helps rather than one that is enough. If you are still thinking only in terms of classic ranking factors, you are optimising for the wrong filter.
Query fan-out explains where the rest comes from
If a third of citations do not come from the SERP of the query typed, they come from the SERPs of the sub-questions the system builds for itself. That is why a page can be cited for "AI Overview ranking factors" while ranking for "how does query fan-out work".
Across 146 million SERPs, again from Ahrefs, the profile of queries that trigger an AI Overview is clear: 57.9% of queries phrased as a question against 15.5% of those that are not, 59.8% for questions starting with "why", 47.3% for definition requests, 46.4% for queries of seven words or more. By intent, 99.9% of AI Overviews appear on informational queries and just 1.2% on transactional ones.
In practice: the derived questions are the surface to cover, and you cover them with pages that answer one question at a time rather than a single guide that contains them all, or with an FAQ section built on the right questions. If you already work by search intent, you start ahead, because the map is the same.
The signals that correlate, and by how much
Cyrus Shepard lined up 54 published studies, patents and documented cases on citations in ChatGPT, Gemini and Perplexity, giving each factor a strength-of-evidence score. At the top are URL accessibility (9.5), search ranking (9.4), ranking for derived queries (9.3), preview controls (9.2) and the match between question and answer in the text (9.2). At the bottom, domain authority at 5.0 and the llms.txt file at 2.0.
Two figures from Ahrefs complete the picture, and they should be read as correlations, not causes. Across 75,000 brands, brand mentions on the web correlate with AI visibility at 0.664, against 0.218 for backlinks: three times as much. On freshness, cited content is on average 25.7% more recent than uncited content, 1,064 days against 1,432. The first figure explains why classic link building pays off less than expected here; the second why updating old articles often pays off more than writing new ones.
The five things Google says are not a factor
This is the part almost nobody on the first page reports, and these are verbatim quotes from Google's documentation.
| Commonly recommended practice | What Google says | Verdict |
|---|---|---|
| Publishing an llms.txt file | "Google Search ignores them" | Ruled out |
| Chopping text into tiny chunks for AI | "There's no requirement to break your content into tiny pieces for AI to better understand it" | Ruled out |
| Adding structured data specifically for AI | "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add" | Ruled out |
| Hitting an ideal page length | "There's no ideal page length" | Ruled out |
| Writing in a particular way so models understand you | "You don't need to write in a specific way just for generative AI search" | Ruled out |
| Indexing, crawlability, preview controls | Explicitly listed requirements | Confirmed |
| Organic ranking | Not mentioned by Google, measured by studies | Strong correlation, falling |
| Brand mentions on the web | Not mentioned by Google, measured by studies | Correlation (0.664) |
On the same page Google says what does count: "Creating content that people find unique, compelling, and useful will likely influence your website's presence in generative AI search more than any of the other suggestions". It sounds generic, but it is a precise instruction, because it comes straight after five technical denials. Structured data is still useful for featured snippets and rich results, and it makes sense to keep maintaining your schema markup: it is just not what gets you cited in an AI Overview.
The bottleneck is narrower than coverage suggests
Coverage figures tell you how often AI Overviews appear; they do not tell you how selective they are. In the UK, SISTRIX found AI Overviews on more than 18% of the 51 million keywords it tracks in March 2025, up from 4% a few months earlier, while featured snippets halved to around 5%. Informational keyword sets, with health at the top, are the most affected; purely commercial ones much less.
We have not found a UK study that measures how many domains actually get cited. The clearest national figure we know of comes from Italy, where we work: SEOZoom analysed more than 30 million Italian keywords and found an AI Overview in 54% of SERPs, but the domains cited as a source number around 175,000, 0.1% of the 168 million indexed.
| Sector | AI Overview coverage in Italian SERPs |
|---|---|
| Health | 40.3% |
| Finance | 20.2% |
| Travel and tourism | 6.4% |
| Retail and grocery | 4.3% |
The ranking of the most cited sources in that Italian study says more than coverage: YouTube with 875,914 mentions, Wikipedia with 864,517 and in the primary source position 45% of the time, then my-personaltrainer.it with 403,934 mentions earned with just 21,953 URLs. The third name is the interesting one, because it is neither an encyclopaedia nor a video platform: it is a niche site on a single subject. Specialisation pays more than size, and for a business that will never be Wikipedia it is the only practical route.
For comparison, Semrush, on more than 10 million keywords and a mainly US sample, puts queries with an AI Overview at 15.69% in November 2025, after a peak of 24.61% in July. SISTRIX UK at 18%, Semrush at 15.69% and SEOZoom Italy at 54% cannot be compared: sample, market, date and keyword selection all differ. Read them as different measurements of the same phenomenon, which is why a figure from another market quoted without context tells you nothing useful about the SERPs you actually work on.
AI Overviews and AI Mode do not draw on the same sources
Ahrefs compared 730,000 pairs of answers generated by AI Overviews and AI Mode for the same queries. Citations overlap in 13.7% of cases. Nine times out of ten the two systems say the same thing, but they take it from different pages: AI Mode answers are about four times longer and name 3.3 entities on average against 1.3.
Being cited in one does not get you cited in the other, so they need to be tracked separately. Anyone working on GEO and AEO as if they were one channel is measuring two different channels badly.
What a citation is actually worth
Seer Interactive followed 53 brands, 5.47 million queries and 2.43 billion organic impressions between January 2025 and February 2026. The most useful figure is expressed per million impressions on informational queries.
| Scenario | Organic clicks per million impressions |
|---|---|
| No AI Overview in the SERP | about 33,500 |
| AI Overview present, brand cited | about 20,743 |
| AI Overview present, brand not cited | about 9,445 |
Being cited is worth more than twice the clicks of not being cited, and still less than a SERP with no AI Overview. The same study reports a movement nobody expected: organic CTR on queries with an AI Overview climbed back from 1.3% in December 2025 to 2.4% in February 2026, after eighteen months of decline. A three-month rebound is not a trend, but it is enough to stop assuming the curve only goes one way. For the full numbers, see our SEO and AI statistics and the analysis of the impact of AI Overviews; on how to spot the drop in your own data, I have put together the figures on AI Overviews' impact and a method for measuring it.
How to measure it, from June 2026
Until recently the answer was: you cannot measure it, you estimate it with third-party tools. Since June 2026 Search Console includes reports for generative AI features, showing the impressions a site gets in AI Overviews and AI Mode, segmentable by page, country, device and date.
What the report does not give you: queries, clicks, average position, or where in the answer the citation appears. And generative impressions do not add up with classic Search impressions, because they are counted differently: at property level, two URLs from the same domain in the same answer count as one impression, while at page level each URL is counted separately.
One detail worth knowing before you build a dashboard on it: the data lives in a separate report, not in the API's "search appearance" dimension. When we queried Search Analytics on the searchAppearance dimension for one of our properties for 1 May - 20 August 2026, the only values returned were the classic types. If you have automated extraction via the API on that dimension, you will never find AI Overviews there. For how the tool works in general, see our guide to Google Search Console.
Where I would start, in order
- Check the preview tags across the whole site. It is half an hour's work and it unblocks or blocks everything else.
- List the derived questions your topic opens up, with one page per question rather than a guide that holds them all.
- Rewrite the first 150 characters of each section as a straight answer to the question in the heading. That is the passage a model extracts, not the paragraph that leads into it.
- Update content older than two years on the cluster's informational queries before writing new pieces.
- Brand mentions off the site: trade fairs, trade associations, niche directories, industry publications. They count for more than links.
On an industrial project we run for Macropix, a Milan-based LED screen manufacturer, the Italian keyword "monitor pubblicitario" (advertising display) moved from position 88 to 2, and on the LED wall cluster the domain now has a 25.55% share of voice, ahead of Amazon at 18.08% (Semrush, July 2026). That result came from work on site structure and installation case pages, not from optimisations designed for AI: the full story is in the article on SEO for manufacturing, and similar projects are among our case studies. If you want help setting up work on AI citations, that is what we do with our AI SEO service.
One thing the industry rarely says: the fact that 31% of citations come from pages outside the top 100 is presented as good news for those who do not rank. It is not. It means the selection criterion has moved to a surface you cannot control with the classic levers, where it is much easier to lose a citation than to win one, and where nobody tells you when it happens. The predictable part of the work is still the old one: be indexed, be readable, answer the question in the first line. The rest, for now, you observe.
Frequently asked questions
Not in the sense of a list published by Google. There are documented technical requirements (indexing, crawlability, preview controls) and signals that studies of millions of SERPs show to be correlated with citations, such as organic ranking, brand mentions and content freshness. Much of what else circulates as a factor is ruled out by Google's own documentation.
It helps, but it is neither enough nor essential. According to the March 2026 Ahrefs study of 863,000 keywords and 4 million cited URLs, 38% of cited pages are in the top 10, 31.2% between positions 11 and 100 and 31% outside the top hundred. In July 2025 the top 10 share was 76%.
Google states that structured data is not required for generative search and that there is no special schema.org markup to add. It is still useful for rich results and featured snippets in classic Search, so it makes sense to keep it, but it is not the lever that gets you a citation in an AI Overview.
Since June 2026 Search Console includes reports on generative AI features, with the impressions earned in AI Overviews and AI Mode broken down by page, country, device and date. It does not show queries, clicks, average position or where in the answer the citation appears, and generative impressions should not be added to classic Search impressions.
SISTRIX found AI Overviews on more than 18% of the 51 million UK keywords it tracks in March 2025, up from 4% a few months earlier, with informational and health queries the most affected. Figures vary a lot by market and sample: in Italy, SEOZoom measured 54% of SERPs across more than 30 million keywords, yet only about 0.1% of indexed domains were cited as a source.
More than not being cited, less than having no AI Overview at all. In Seer Interactive's 2026 data, for every million organic impressions on informational queries there are about 33,500 clicks with no AI Overview, about 20,743 when the brand is cited and about 9,445 when there is an AI Overview but the brand is not among the sources. International sample of 53 brands.
