AI Overviews cite what Google does not rank

Two things are true of the same Google results page, and the industry mostly talks about one of them. There are ten blue links, and above them there is an AI Overview citing a handful of sources. We assumed those were largely the same sites. On 75 searches we captured by hand, they were not.

What we measured

Seventy-five commercial SEO searches, run in the United States in September 2026, one capture each, num=20 and personalisation off. For every search we recorded the organic results, the AI Overview, and every domain the AI Overview cited.

The queries are the ones a business hiring an SEO firm would actually type — answer engine optimization, ecommerce seo, b2b saas seo, shopify seo, entity seo and so on. That matters for reading the numbers: this is one vertical, not the web.

Each search was collected by opening the page and reading what was on screen. No API, no scraper, no automated searching. The raw HTML of every capture is kept, which is the only reason the counts below could be recomputed after we found two mistakes in how we were reading them.

The finding

More than half of what the AI Overview quotes is not what Google ranks underneath it. On ten of the 75 searches there was no overlap at all: every source in the overview was a site the page did not otherwise show you.

YouTube is the clearest case

YouTube is the single most-cited source across these 75 searches and appears in the organic top ten of none of them. Reddit shows the same shape at a smaller scale. Semrush and Shopify show the opposite: heavily ranked, cited less often than they rank.

We are not going to tell you why. Counting citations tells you that citation and ranking come apart; it does not tell you what causes either. Anyone explaining the mechanism from this data is adding something the data does not contain.

It is not a closed list

The obvious worry is that AI Overviews quote the same dozen large sites forever. These captures do not support that.

• 137 distinct domains were cited at least once.
• 74 of them were cited exactly once.
• The ten most-cited domains account for 37% of all citations.

Concentrated at the top, long tail underneath — which is roughly what organic search looks like too.

What Google says about getting cited

Worth putting next to the numbers, because it is the opposite of what most of this category is sold on. Google’s own documentation states:

There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.

And, specifically:

You don’t need to create new machine readable files, AI text files, or markup to appear in these features. There’s also no special schema.org structured data that you need to add.

That is Google saying the entry requirement is being indexed and eligible for a snippet. It does not say markup is useless — “not required” and “does not help” are different sentences and we are only entitled to the first one. But it does sit awkwardly against a market selling files and schema as the route in.

The academic work that named this field found something adjacent. Aggarwal et al., presented at KDD 2024, tested optimisation methods across a 10,000-query benchmark and reported visibility gains of up to 40% from adding statistics, adding quotations, and citing sources — while keyword stuffing, the oldest tactic in search, was not among the methods that worked.

Two mistakes we made reading this

Both nearly became findings. They are here because a measurement piece that hides its errors is doing the thing it is complaining about.

The placeholder. Google ships the text “An AI Overview is not available for this search. Can’t generate an AI overview right now.” hidden in the page on essentially every search — 73 of our 75, including every one that plainly did show an overview. Counted naively, that produces the headline “97% of searches cannot generate an AI Overview”. It is a placeholder element, not a statement about the search.

The empty box. Our first count said an AI Overview appeared on 75 of 75. It appeared on 74. The container Google uses stays in the page even when nothing renders inside it, and we were counting the container. One search — shopify seo help — had the box and no overview.

Both were caught by re-reading the stored HTML and finding it disagreed with what we had recorded live. That is the argument for keeping raw captures rather than just the counts.

What this does not show

• Not causation. We measured a divergence between citation and ranking, not a mechanism, and nothing here says how to get cited.
• One vertical. Seventy-five commercial SEO queries. Not news, not healthcare, not retail products.
• One moment. One capture per search, in September 2026. AI Overviews change between runs; we saw ad blocks appear and vanish within 25 minutes on the same query.
• US only. Indian captures exist in the same corpus and are not mixed in here, because the same query returns different results by market.

The version of this worth reading is the same measurement repeated in six months, when there is a before to compare against. This is the before. The same corpus read page by page — how often the #1 result is the cited one, and how often an overview cites nobody from page one — is in GEO vs SEO: the overlap, measured.

Why we ran it

Because our GEO work rests on a claim that ought to be checkable: being cited by an answer engine is a different outcome from ranking, and it needs measuring on its own terms. Fifty-four percent is the number behind that sentence.

If you want the same reading for your own category, that is what the free audit starts from. And if you want the neighbouring question — whether sites are quietly blocking the crawlers that feed these answers — we counted that too, in almost nobody has decided about AI crawlers.