In September we put eight marketing questions to Google's AI Mode logged out and wrote down every site it cited. The point was never the single snapshot. The point was to do it again.

So on 1 October 2026 we ran the same eight prompts, in the same order, on the same engine, logged out, and recorded every citation again. Same method, one month apart.

The headline is not that the numbers went up. It is that 31 of the 51 sites cited in October were not cited in September at all — and 28 of September's sites had vanished. Roughly three in five of the cited domains turned over in a month.

BuzzRiding was cited zero times out of eight. Same as September.

🔬 How this was run

Google AI Mode (udm=50), logged out, from Ireland, on 1 October 2026 — the same eight prompts as run 1, unchanged. Every prompt and every cited domain is in the October raw CSV next to September's, with the run-2 design note recording three measurement problems we found. Perplexity and ChatGPT still require a login and are still excluded.

Finding 1: the cited set churns by about 60% in a month

Across both runs, 51 domains were cited in October and 48 in September. Only 20 domains appear in both months. Thirty-one were new in October; twenty-eight of September's dropped out entirely.

The survivors are worth naming, because they are the closest thing this data has to a stable set: youtube.com, linkedin.com, medium.com, coursera.org, semrush.com, frase.io, adlibrary.com, aiclicks.io, brandviz.ai, mersel.ai, analyticahouse.com, keomarketing.com, geoscout.pro, clickclickmedia.com.au, mambadigital.au, leadde.ai, synthesia.io, launchmind.io, one2create.co.uk and digitalmarketinginstitute.com.

Everything else was a one-month visitor.

"A citation is not a position you hold. On this evidence it is closer to a coin that lands your way for a month."

This is the single most useful thing two runs have told us, and it cuts against how GEO is usually sold. If being cited once decays this fast, a tracker that reports "you were cited" without reporting whether you stayed cited is reporting noise.

Finding 2: the volume increase is mostly a measurement artifact, and we are not going to pretend otherwise

October produced 78 citations against September's 55. That is a 42% increase and it would make a better headline than the one above.

We are not using it, because 19 of October's 78 citations were the same domain cited more than once inside a single answer, and September's CSV contains no within-answer duplicates at all. Either the answers genuinely started leaning on single sources repeatedly, or run 1's recording collapsed duplicates and we are comparing two different things. The artifacts cannot settle it.

On the basis that is comparable either way — distinct domains cited per answer — the count went from 55 to 59. A 7% increase, not a 42% one.

Seven of the eight prompts returned more citations than in September. The exception was best GEO tracker tools 2026, which fell from 11 to 7.

Finding 3: the fragmented citation graph held

September's strongest structural finding was that almost nothing gets cited across multiple answers. It held.

In October, four domains appeared in more than one of the eight answers: youtube.com (5 answers), zapier.com (3), reddit.com (2) and linkedin.com (2). In September it was three: youtube.com (6), adlibrary.com (2) and linkedin.com (2).

Two runs, and the picture is the same — there is no club of go-to sources on these topics. One strong article still will not compound across queries.

Finding 4: YouTube is still in the majority of answers, and we owe a correction

YouTube was cited in 5 of 8 answers in October.

The September post said YouTube appeared in "five of eight" answers. Checking its own CSV while building this comparison, it was six of eight. The post was wrong by one and the raw file was right; the error was ours, in the writing, not in the data. Corrected here.

Either way, the conclusion is unchanged and uncomfortable: a text-only site is structurally locked out of a meaningful share of the citation slots in most of these answers.

Finding 5: Reddit and Quora arrived

Neither appeared anywhere in September. In October, reddit.com was cited in two answers and quora.com in one — including on does AI written content rank on google 2026, where a Quora thread sits alongside Ahrefs and a ranking study.

One month is not a trend. It is worth watching, because if forum content is entering this pool, the format question is not just video.

Finding 6: the "GEO" ambiguity is resolving, slowly

September's most actionable finding was that best GEO tracker tools 2026 returned a mix of generative-engine tools and IP geolocation APIs — ipinfo.io, ip-api.com and apifreaks.com — because the older meaning of "GEO" pulls hard. BuzzRiding has an article targeting exactly that phrase.

In October, those three are gone. One geolocation citation remains: qicapp.com, a location-tracking app. So the query is resolving more cleanly toward the generative-engine meaning than it did a month ago — but it is the one prompt of eight whose citation count fell, so this is not yet good news for that article.

Finding 7: the same misattribution, two months running

In both runs, AI Mode cited an article at launchmind.io and labelled the source "Marcus Theatres" — a cinema chain. The same wrong attribution, on the same prompt, a month apart.

It is a small thing, but it is a reminder that the brand label shown next to a citation is generated, not read off the source, and can be flatly wrong twice in a row.

What we are changing because of this

💡 Key takeaway

Between September and October, 31 of 51 cited domains were new and 28 disappeared. The pool is roughly the same size and almost entirely different people. If you are measuring AI visibility, measure whether you stay in it.

The obvious caveat

n=8, one engine, one country, two days a month apart. Nothing here is a study. The prompts, both months' raw citation lists and the design notes — including the three measurement problems run 2 turned up — are all in the repository, so anyone can re-run it and disagree with us. The external work cited in these answers is worth reading directly too: Ahrefs on whether Google penalises AI content and theStacc's AI-versus-human ranking study both turned up in the October run.

Frequently Asked Questions

How much does the set of AI-cited sites change month to month?
In this two-run comparison, a lot. Of 51 domains cited by Google AI Mode in October 2026, only 20 had also been cited on the same eight prompts in September. Thirty-one were new and twenty-eight of September's were gone — roughly 60% turnover in a month.
Did Google AI Mode cite more sources in October than September?
Yes, but less than the raw numbers suggest. Raw citations went from 55 to 78, but 19 of October's were the same domain repeated inside one answer and September's file records no such repeats. On the comparable basis — distinct domains per answer — it went from 55 to 59.
Which sites get cited across multiple AI answers?
Very few. In October only youtube.com (5 of 8 answers), zapier.com (3), reddit.com (2) and linkedin.com (2) appeared in more than one answer. September showed the same pattern with three domains. The citation graph stays fragmented across both runs.
Is a single AI citation worth optimising for?
On this evidence, not on its own. With about 60% of cited domains turning over between two consecutive monthly runs, one appearance is closer to a transient event than a held position. Retention across months is the more meaningful measure.
Was BuzzRiding cited this month?
No. Zero of eight prompts in October, the same as September, on topics where the site has published articles. That result is published unchanged each month whether or not it improves.