Two months ago we started putting the same eight marketing questions to Google's AI Mode, logged out, on the first of the month, and writing down every site it cites. Run 1 was a snapshot. Run 2 found roughly 60% of the cited domains turned over in a month and concluded that a citation is not a position you hold.

Run 3 says that conclusion was half right, and the half it got wrong is the useful half.

The churn replicated exactly: 20 of November's 48 cited domains had also been cited in October, the same carryover count as September to October. But the churn is not spread across the eight questions. It is concentrated in one kind of question and almost absent from another — and eight independent sites have now held a slot on the same question three months running.

BuzzRiding was cited zero times out of eight. Third run, same result.

🔬 How this was run

Google AI Mode (udm=50), logged out, from Ireland, on 1 November 2026 — the same eight prompts as runs 1 and 2, unchanged, in the same order. Every prompt and every cited domain is in the November raw CSV, beside October's and September's. The run-3 design note was committed before any data was collected and carries six measurement notes appended after — including one check that failed.

Finding 1: churn is a property of the question, not of the engine

This is the finding worth the three months. Taking each prompt on its own and asking how many of October's distinct cited domains were still cited in November:

The two head-to-head comparison prompts retained the most. The broad superlative prompt retained nothing at all — and its citation count fell from 8 to 5 at the same time.

"Ask the engine to name the best tools and it reshuffles the deck every month. Ask it to compare two named products and it mostly cites the same people it cited last month."

We are not going to put a percentage on that split from eight prompts. The direction is what matters, and it is actionable: a question with named entities in it behaves like a stable slot, and a question asking for a ranked best-of behaves like a lottery.

Finding 2: eight sites have now held the same question for three months

Twelve domains appear somewhere in all three runs: youtube.com, linkedin.com, medium.com, adlibrary.com, semrush.com, coursera.org, one2create.co.uk, synthesia.io, leadde.ai, digitalmarketinginstitute.com, clickclickmedia.com.au and mersel.ai.

Eight of them have been cited on the same prompt in all three runs:

linkedin.com and youtube.com hold multiple questions across all three runs on top of that.

Last month we wrote that a citation is "closer to a coin that lands your way for a month". That was wrong for these eight. They are not being re-drawn each month; they are incumbents. Several are small sites — mersel.ai, leadde.ai, one2create.co.uk, clickclickmedia.com.au — which makes the incumbency more interesting, not less.

Finding 3: the run-level numbers barely moved

Three runs, on the two bases that are comparable:

So the pool is the same size each month, slowly citing a little more per answer, and swapping out well over half its membership. A 58% turnover with a flat total is the signature of a system with a small stable core and a large interchangeable fringe — which is exactly what findings 1 and 2 describe from the other direction.

One correction we owe on our own data: the September post reported 55 citations, but the citation_count column of September's own CSV sums to 56, and run 2's design note inherited the 55. The raw file was right; our arithmetic was wrong. September's CSV also contains one malformed entry, visbyblog, which is not a domain.

Finding 4: a third of the citation slots are on platforms, and Instagram just arrived

Of November's 80 citations, 25 point at four platforms: youtube.com (11), linkedin.com (7), reddit.com (5) and instagram.com (2).

Instagram had never appeared in either previous run. Both November citations are Reels, cited on does AI written content rank on google 2026 alongside Ahrefs and Search Atlas.

YouTube was in 6 of 8 answers, against 5 of 8 in October and 6 of 8 in September. Flat, and still the single most-cited domain in every run.

Reddit went from 2 answers in October to 3 in November. Quora, which we flagged last month as possibly the start of forum content entering the pool, did not reappear at all.

The uncomfortable read: a text-only website is competing for roughly two thirds of the slots, not all of them. The comfortable read, which is new this month, is that the other third sits on platforms anyone can publish to without owning a domain at all.

Finding 5: "GEO tracker" is now ambiguous in a third different way

This is the prompt BuzzRiding has an article targeting, and it is the third consecutive run in which the engine has not resolved what the phrase means.

September mixed generative-engine tools with IP geolocation APIs. October mostly dropped those, keeping one location-tracking app. November mixed in vehicle fleet tracking: 4 of the 15 citations were samsara.com, gomotive.com, bouncie.com and a Business News Daily guide to GPS fleet tracking systems.

It was also the highest-citation answer of the run — 15 citations across 14 distinct domains, against 7 in October. The engine is not narrowing on this phrase. It is casting wider and pulling in a new wrong meaning each month.

Three runs is enough for us to stop calling this a quirk. For any phrase where an older, denser meaning exists, the AI answer may never settle, and an article targeting it is competing against a dictionary problem rather than against other articles.

Finding 6: on tool questions, the vendors take the slots

On best GEO tracker tools 2026, 7 of 15 citations were product homepages rather than coverage of those products: otterly.ai, peec.ai, profound.com, radarkit.com, samsara.com, gomotive.com and bouncie.com. On HeyGen vs Synthesia it was 3 of 7 — heygen.com twice and synthesia.io. The editorial citations that remained went to Zapier's AI-visibility-tool roundup, margen.net's GEO tracking comparison and a handful of others.

Runs 1 and 2 counted vendor pages the same way, so this is not a measurement change. It does mean the citation count on a tool question is not a count of publishers, and that a publisher is competing for fewer slots than the headline number suggests.

What went wrong in this run

Two things, both recorded in the design note rather than smoothed over.

The check we promised did not happen. Every run is supposed to re-test whether Perplexity and ChatGPT have opened up logged-out access. This month the browser refused to navigate to perplexity.ai at all, so the engine was never reached. That is not evidence that it still requires a login — it is a check that failed. Both engines stay excluded on run 1's reasoning and the re-check is still owed.

Enumeration missed links we could see. On two prompts the citation-enumeration step skipped links the page clearly contained — two YouTube links on the GEO tracker answer, a fourth Reddit link on the model-comparison answer. Following run 2's rule, the CSV records what was enumerated, so YouTube and Reddit are recorded at or below their true presence in those two answers, never above.

One thing that went right by disappearing: the launchmind.io article that AI Mode labelled "Marcus Theatres" in both previous runs was not cited at all this month.

What we are changing because of this

💡 Key takeaway

Monthly churn in Google AI Mode's cited set held at ~58% for a second month — but it is concentrated almost entirely in broad "best X" questions. Eight independent sites have now been cited on the same specific question three months running. If you want to be cited, pick a question, not a topic.

The obvious caveat

n=8, one engine, one country, three days a month apart. Nothing here is a study, and the per-prompt retention counts above are single observations, not rates. All three months' raw citation lists and all three design notes — including the failed check in this one — are in the repository, so anyone can re-run it and disagree with us.

Frequently Asked Questions

Do AI search citations last, or do they turn over every month?
Both, depending on the question. Across three monthly runs of the same eight prompts, about 58% of cited domains turned over each month — but the churn was concentrated in broad "best tools" questions, while head-to-head comparison questions retained most of their sources and eight sites have now been cited on the same question three months in a row.
Which sites does Google AI Mode cite most consistently?
In this three-run index, twelve domains appeared in all three months: youtube.com, linkedin.com, medium.com, adlibrary.com, semrush.com, coursera.org, one2create.co.uk, synthesia.io, leadde.ai, digitalmarketinginstitute.com, clickclickmedia.com.au and mersel.ai. Eight of those were cited on the same prompt each time.
How much of an AI answer's citations go to social platforms?
In the November 2026 run, 25 of 80 citations pointed at YouTube, LinkedIn, Reddit or Instagram. YouTube alone was cited in 6 of the 8 answers. Instagram appeared for the first time in three runs, with two Reels cited on a question about AI content ranking.
Why do some keywords never get a clean AI answer?
Because an older meaning of the phrase competes with the one you intend. Across three runs, "best GEO tracker tools 2026" returned IP geolocation APIs, then a location-tracking app, then GPS vehicle fleet tracking — a different wrong meaning each month. Check for that collision before targeting a phrase.
Was BuzzRiding cited this month?
No. Zero of eight prompts in November, the same as September and October, on topics where the site has published articles. The result is published unchanged every month whether or not it improves.