Original measurement

Half the sources Google’s AI Overview cited were still there eleven days later. The other half were gone.

We recorded which pages Google’s AI Overview cited across 25 queries in our category, then ran the identical measurement again eleven days later. Of the citation slots that existed the first time, 38% no longer existed the second. This is the surface that carries almost all of our category’s measured demand, and it is the first thing anyone selling AI visibility should be able to tell you about it: a citation there is neither a position you win nor a lottery you re-enter every time. It is roughly half durable.

38%

of the AI Overview citation slots we recorded were gone eleven days later

Key findings

Across 16 queries measured twice eleven days apart, the median overlap between the two sets of sources Google’s AI Overview cited was 52%.
Of the 119 citation slots recorded in the first reading, 74 held, 45 were gone eleven days later and 31 slots were new - a loss of 38% of the original slots.
The turnover is not the citation set shrinking: the same queries carried 119 citations in the first reading and 105 in the second, while 31 entirely new sources appeared.
Durability is not the preserve of the large publishers: of the 74 slots that held, 27 belonged to YouTube, Reddit, Google and Semrush and the other 47 to smaller sites.
The share of AI Overview citations pointing at domains outside the organic top ten was 49% in the first reading and 47% in the second - the same finding, on data collected eleven days later.

How this was measured

When: Two readings of the same map: 29 August 2026 and 9 September 2026.
Instrument: One script, run twice, with every setting fixed in the file rather than passed in: the same 25 queries, United States, English, desktop, ten organic results deep. The second reading was taken deliberately and the first was kept unchanged beside it.
What counts as a citation: A source Google’s AI Overview block referenced for that query, read from the reference list rather than from the visible text. Duplicates within one block count once.
Two measures, not one: A single overlap figure can be produced by two different worlds - a set that rotates, and a set that merely shrank. So the sets are compared symmetrically (52%) and again by asking only what share of the first reading survived (61%). They agree, which is the reason to believe either.
What is excluded: Two queries carried a local pack instead of an AI Overview. Six more returned an AI Overview block with no reference list at all, the same six on both days - the known shape of a block Google loads separately. Those eight are excluded rather than counted as zeros, which leaves 16 queries cited on both days.

What we measured

Everything else we publish measures an assistant answering a question about a local business. This measures the other surface, and it is the larger one: of the demand we have measured in our own category, almost all of it sits on Google, where the answer arrives as an AI Overview with a list of sources beneath it.

The question we set out to answer is the one a buyer asks second, right after “can you get me cited”. It is: if I am cited, how long does that last? Nobody in this category publishes an answer, ourselves included - so we took the same reading twice and compared it.

The result

Of the 119 citation slots in the first reading, 74 were still there eleven days later, 45 were gone, and 31 slots were held by sources that had not been cited at all the first time. The median query kept about half of its cited set.

The second measure agrees with the first. Comparing the two sets symmetrically gives a median overlap of 52%; asking only what share of the first reading survived gives 61%. Neither number is near zero and neither is near a hundred, and that is the finding.

It is not that the citation list got shorter

The obvious alternative reading is that Google simply cited fewer sources the second time, and that everything else follows from that. It did cite slightly fewer - 119 across these queries, then 105. But a shrinking list cannot produce 31 sources that were not there before, and 31 is what we counted. Sources left and other sources arrived.

Who held, and who rotated

The most durable single source is YouTube, which held its slot on 14 of the 16 queries. That is the result anyone would guess, and it is worth publishing because the next one is not.

Of the 74 slots that held, only 27 belonged to YouTube, Reddit, Google’s own pages and Semrush together. The other 47 belonged to smaller sites - the marketing shops and small software companies that already hold most of the citation slots in this category. Durability is not something only the large publishers get.

And one figure that did not move

We already publish that about half of the AI Overview citations in our category point at domains that are not in the organic top ten for the same query. That was measured at 49% of 127 citations. Eleven days later, on a fresh reading, it is 47% of 105.

We did not set out to replicate it - the second reading was taken to measure turnover - which is the most convincing way for a figure to be replicated.

What this means if you are buying

A citation in Google’s AI Overview is worth having and is not worth treating as won. About half of what is cited today is cited eleven days from now, so the honest unit is a rate over repeated readings, not a screenshot of one good day.
Anyone who shows you a single screenshot of your business inside an AI answer has shown you something with a roughly even chance of still being true in a week and a half. That includes us: it is the reason our own reading is sold as a repeated measurement rather than a certificate.
The 31 new slots are the encouraging half. Sources that had never been cited on a query were cited on it eleven days later, and most of the durable slots belong to sites the size of the companies we sell to. The surface is not closed.

Honest limits of this measurement

Two readings, eleven days apart. That measures whether the set moves; it does not measure the shape of the movement, and two points cannot tell a steady drift from a single reshuffle.
16 queries, all in our own category. What holds for “generative engine optimization” need not hold for a query about a local trade, and we would not quote this figure at a plumber.
One country, one language, one device, ten results deep. AI Overview content varies by all four, and a reading taken any other way is not comparable with these.
Six queries returned an AI Overview with no reference list on both days. We read that as the block being loaded separately rather than as a genuine absence of sources, and we exclude those queries instead of scoring them zero. If that reading is wrong, the turnover figure is too optimistic rather than too pessimistic.
Whether a source “held its slot” is measured at the level of the domain, not the individual page. A site cited for one page and then another counts as having held.

Questions about this research

So is being cited in AI Overview worth anything?

Yes, and less than a ranking. Half of it persists over eleven days, which is long enough to matter and short enough that a single check tells you little. That is an argument for measuring repeatedly, which is what we sell, and against anyone selling a one-off certificate.

Does this mean the same thing as your finding that AI answers turn over completely?

No, and the difference is the point. On the assistant surface, two readings of the same question ten minutes apart shared none of the businesses named. Here, on Google, about half the cited sources persist over eleven days. Two surfaces, two different behaviours, both measured by us.

Can I do anything to be one of the sources that holds?

We do not know, and we will not pretend otherwise. We have tested five explanations for who gets cited - page structure, sentence-level text matching, a page per trade, the size and shape of the site, and off-page authority - and all five failed. What we can tell you is whether you are in the set today and whether you were last month.

Why only eleven days?

Because that is the gap between the two readings we actually took, and we are not going to publish a monthly or annual figure we have not measured. The next reading extends the series; when it exists, this page changes.

We run this measurement for a living. If you want it run on your own category and your own city, the free scan is the same instrument on one question, and the report is the full set. Nothing here is behind a form.

More measurements from the same instrument

Every category we measure has its own page, with the questions customers put to an assistant in that category and what a reading costs. See all industries we measure →

Published 31 August 2026. Every figure on this page is recomputed from raw observation files by a single script, so any of them can be traced back to the readings behind it. If you are named here and believe a reading is wrong, tell us: we will run the measurement again and publish what it returns, whichever way it goes.