The Babel42 study ran 601 Google searches every day for 15 days and recorded which sites each AI Overview cited. None of the 418 tracked AI Overviews cited the same set of sites every day. Each one cited about 7 sites per day, but 18 different sites over the 15-day window. Only 11% of AI Overviews cited exactly the same sites as the day before, and the first-listed source changed in 42% of searches from one day to the next.
If you ran an audit on Monday and checked your AI visibility on Tuesday, you would be looking at a different answer.
The data
Babel42 ran 9,675 searches between September 26 and October 12, 2026. They recorded 660 searches on day one, then re-ran 601 of them every day at 13:00 UTC for 15 days. They captured the AI Overview, its citations, and the organic top 10 for each search.
The volatility numbers:
- 0 of 418 AI Overviews cited the same sites every day over 15 days
- Each AI Overview cited about 7 sites per day but 18 different sites over 15 days
- Only 11% cited exactly the same sites as the day before
- The first-listed source changed in 42% of searches from one day to the next
- AI Overviews and AI Mode had only 33% of cited sites in common for the same search on the same day
What this means
Google's AI Overview source selection is not a ranking. It is a sampling process. The model retrieves a pool of candidate sources and selects a subset for each answer. The subset changes every day, even for the same query, because the retrieval and selection process has stochastic elements.
This has three implications for anyone doing AI visibility work:
1. A single citation check is noise. If you ask Google a question today and check whether your site is cited, you get one sample from a distribution. Tomorrow you might be cited. Today you might not. The 11% day-to-day stability means that 89% of the time, the answer changes. A one-shot check tells you almost nothing about your actual citation rate.
2. Citation tracking requires repeated sampling. To know your real citation rate for a query, you need to run it multiple times over multiple days and compute the percentage. One run is a coin flip. Ten runs over two weeks gives you a rate. This is what the paid citation tracking tools (Otterly, Profound, Peec AI) do: they run prompts on a schedule and aggregate results. The open-source tool does not do this. The GeoReady platform is building toward it.
3. The calibration study's null result is more understandable in this context. If citation is this volatile, any single-day test of whether a score predicts citation is testing against a moving target. The 341-page calibration study ran on one day (October 17). On a different day, with different Perplexity citations, the result might be different. The pre-registration committed to one day. The honest reading is: one day is not enough, and the null result reflects that limitation as much as it reflects the score's predictive power.
The practical consequence
If you are a site owner and you check your AI visibility once, you are reading one frame of a film. The frame might show you cited. The next frame might not. The film has 15 days of frames, and you saw one.
This is the argument for continuous monitoring. An audit tells you if your page is readable by AI crawlers. It is a static check. Citation tracking tells you if you are actually being cited, over time, with the volatility accounted for. The audit is a photograph. Citation tracking is a video.
The Babel42 data also found that 46% of sites cited by AI Overviews do not rank in Google's top 10 for the same search. So even the relationship between organic ranking and AI citation is unstable. You can rank #1 and not be cited. You can rank #50 and be cited. And which one you are might change tomorrow.
What to do
For audit: run a free audit on geoready.dev to check whether AI crawlers can reach your site. This is the static layer. It does not change day to day. If your score is low, fix it. If it is high, move to tracking.
For tracking: if you can afford it, use a paid citation tracker (Otterly from $29/mo, Peec AI from $95/mo) to run your target prompts daily and track citation rate over time. If you cannot, run your prompts manually on ChatGPT, Perplexity, and Google 5 times over 2 weeks and count how often you appear. Five samples is better than one. Ten is better than five.
For monitoring: the GeoReady platform at app.geoready.dev runs scheduled audits and tracks score history. Citation tracking across engines is on the roadmap. The CLI is on GitHub, open source, MIT licensed.
The audit is the photograph. The monitoring is the video. Citation is the film that changes every day, and you need to watch more than one frame to know the story.
Top comments (0)