TL;DR
- Across independent studies published in 2026, weekly citation churn ranges from roughly 39% to 74% depending on engine, meaning the source list in an AI answer is rarely the same two weeks running.
- ChatGPT churns fastest (around 74% weekly replacement, per SISTRIX); Perplexity is the most durable, with a citation half-life near 5.8 weeks per Scrunch/Stacker.
- Google AI Overviews behaves bimodally: roughly half of queries are fully stable, while the other ~19% rotate as fast as AI Mode — so averages hide two very different populations of prompts.
- A small core of sources (around 15% of domains in one 8-week dataset) gets cited every single week and accounts for the large majority of total citations — stability, not frequency, predicts durability.
- If you measure GEO success by a single week's citation snapshot, you are measuring noise. Track a fixed prompt set over multiple weeks before drawing conclusions about whether a change "worked."
AI citation volatility is the week-over-week rate at which large language model answer engines swap out the sources they cite for a given set of prompts. It is distinct from ranking volatility in traditional search: in AI answers, a source can be fully present one week and completely absent the next, with no gradual decline in between.
What a 90-day tracking window actually shows
The most rigorous public dataset so far comes from SISTRIX, which logged 82,619 prompts and over 1.5 million snapshots across Google AI Mode, AI Overviews, and ChatGPT Search in six countries over 17 weeks. The headline number: Google AI Mode replaces 56% of its cited sources every week, and ChatGPT replaces 74%. Neither of these decayed over the measurement window — drift held steady at 54-59% from week one to week seventeen. That consistency is the important part. If churn were an artifact of a new feature settling in, you'd expect it to taper. It didn't, across six different countries, which points to citation rotation being a structural feature of how these systems retrieve and rank sources, not a temporary bug.
A separate 8-week study from Techmagnate tracking 83,633 citations across 2,191 domains found a somewhat lower but still substantial 39% average weekly domain churn. The more useful finding in that dataset is the shape of the distribution: only 15% of domains were cited in every single week of the study, but that small group accounted for 93% of all citations recorded. Churn is high on average, but it isn't evenly distributed — a stable core exists, and it does most of the work.
A six-week, three-wave study from Digital Authority Partners reinforces this: of 1,127 cited URLs tracked across five engines, only 119 (10.6%) survived all three waves, meaning roughly 89% of citations rotated out at least once.
Which engine is most stable, and why it matters for measurement
Ranking engines by stability changes how you should interpret a win or a loss in any single week:
- ChatGPT — least stable. SISTRIX puts weekly source replacement at 74%, and Scrunch/Stacker's half-life analysis (3.4 million citation events) estimates a citation half-life of about 3.4 weeks, the shortest of the three platforms studied.
- Google AI Overviews — stable on average but bimodal. Roughly half of queries return a fixed set of sources indefinitely, while a minority (~19%) churn as fast as AI Mode. Half-life sits around 4.7 weeks per Scrunch/Stacker, but that number averages over two very different query behaviors.
- Perplexity — most durable. Half-life of roughly 5.8 weeks, and a 530,875-citation study from getmentions.ai found 11.1% of Perplexity sources were cited every day across a week-long window — far higher persistence than Gemini's 0.4% daily-retention figure in the same dataset.
- Gemini / Google AI Mode — the getmentions.ai study found AI Mode touched 40 different domains for a single query over one week and retained almost none of them, suggesting very wide, shallow sampling rather than a fixed source set.
This has a direct practical consequence: a single-week absence on ChatGPT or Gemini tells you almost nothing. A single-week absence on Perplexity, sustained for two or three consecutive checks, is a more meaningful signal, given its longer half-life.
Methodology: how to run your own fixed-prompt tracking study
If you want signal instead of noise, the study design matters more than the tool:
- Fix the prompt set before you start. Pick 20-50 buyer-intent questions your prospects actually ask, written in natural language, not keywords. Do not rotate prompts mid-study — that reintroduces the variance you're trying to isolate.
- Query the same engines on a fixed cadence. Weekly is the minimum resolution that separates real drift from day-to-day noise, based on the daily-granularity findings from getmentions.ai showing even daily snapshots are noisy at the individual-citation level.
- Log full citation lists, not just presence/absence. Track every URL cited per prompt per week so you can compute both churn rate (how much the list changed) and retention rate (which domains survived).
- Run for at least 8-12 weeks before concluding a GEO change worked or failed. Given weekly churn rates in the 40-75% range, any one-week improvement could be noise. Techmagnate's finding that only 15% of domains were cited every week is the threshold worth aiming for — consistency beats a single good week.
- Segment by engine. Averaging across ChatGPT, Perplexity, and Gemini masks the bimodal and engine-specific patterns described above. Report stability separately per engine, as in our breakdown of ChatGPT vs Perplexity vs Gemini citation behavior.
Running this manually across four engines and dozens of prompts every week is the bulk of the labor in any GEO measurement program. A free scan gives a one-time snapshot of which URLs get cited for a sample of questions, which is a reasonable starting point before committing to a recurring tracking cadence; paid tracking repeats a fixed question set weekly across ChatGPT, Perplexity, Gemini, and Claude so you can tell churn from a trend rather than guessing from memory.
What this means for measuring GEO work
The practical takeaway is that GEO success criteria need to be redefined around persistence, not presence. A content or schema change that gets you cited once is close to meaningless given baseline churn rates this high. The more defensible goal is appearing in the stable core — the roughly 10-15% of domains that survive every weekly check — because that cohort captures the overwhelming majority of citation volume in every study referenced here. Track multiple weeks, segment by engine, and treat any single data point as an observation, not a conclusion.
FAQ
What counts as "citation churn"?
Citation churn is the percentage of cited sources in an AI answer that are different from the sources cited for the same prompt in the previous measurement period, typically calculated weekly.
Is high churn the same as low quality?
Not necessarily. High churn can reflect broad, shallow sampling behavior (as seen with Google AI Mode) rather than a judgment that previously cited sources are worse — some engines simply retrieve from a wider candidate pool each time.
How many weeks of tracking are enough before trusting the data?
Based on published studies ranging from 6 to 17 weeks, at least 8 consecutive weekly snapshots are needed to distinguish a real change in citation behavior from normal week-to-week noise, given baseline churn rates of 39-74%.
Sources
Drafted with AI assistance and published by CiteFlow. Sources are linked in the text.


