AI citation volatility by industry: one-shot checks miss the signal

Two repeat ChatGPT answers to the same prompt shared only 21.2% of cited domains on average. The other 78.8% churned. Google AI Overviews was steadier, but still shared only 31.5% of sources.

That makes a screenshot a weak way to measure. If a source appears once or disappears once, the right question is not whether the engine changed its mind. It is how often that source appears across repeated runs.

The same number reads more clearly as churn: 78.8% for ChatGPT and 68.5% for Google AI Overviews. Even the steadier surface replaces more than two thirds of its cited source set across repeated answers.

A typical ChatGPT prompt drew from 81.3 distinct domains across about 22 runs. Only 1.43 domains appeared in at least 80% of answers. Those anchor domains carried 22.1% of all citations. Google had a larger core share at 33.2%, which is one reason it looks steadier.

The spread across industries is narrower than expected. ChatGPT source churn ranges from 75.4% in Travel and Tourism to 81.1% in Commerce and Shopping. The noisiest and steadiest industries differ, but both are still too volatile for one-off checks.

The highest-churn industries are Commerce and Shopping, Software, Artificial Intelligence, and Information Technology. These are exactly the categories where source pools are broad, vendor pages change frequently, and the model can choose among many plausible evidence paths.

On Google AI Overviews, AI, Software, Sales and Marketing, Data and Analytics, and IT are still among the most volatile sectors. Consumer Electronics and Sports are steadier, but their churn is still above 63%.

Narrowing to one week raises overlap, but not enough to rescue one-shot measurement. ChatGPT overlap rises from 21.2% to 26.7%; Google AI Overviews rises from 31.5% to 36.8%. The short-window result says the same thing: the noise comes from the answers themselves, not only from sources changing over a month.

If your citation share moves by a few points, this dataset says to treat it as a hypothesis, not a verdict. In categories where roughly 75-81% of the cited source set changes across repeat ChatGPT answers, alerting should require repeated absence or presence, not a single missing source.

The small durable core is where source strategy has the cleanest signal. A domain that appears in one answer may be noise; a domain that appears in most repeated answers is one the engine returns reliably for that category.

Run the same buyer question several times before treating citation movement as meaningful. Track how often each source appears, whether anchor domains keep appearing, and churn for the category, alongside the raw list of cited domains.

The practical gain is restraint: avoid reacting to a single missing citation, but move fast when the same source disappears across repeated runs and across the prompts it used to anchor.