What is AI answer volatility?
AI answer volatility is how much the brands an AI engine names change when the same question is asked again, or asked in different words. AI answers churn: a brand named today can be gone tomorrow and back the day after. Volatility is the size of that churn, and knowing it is what lets you tell a real change from noise before you act on one.
How much do AI answers change from day to day?
In our tracking, about one name in five dropped out of an answer overnight, and half of the changes undid themselves a day later.
Over 7 to 9 days in October 2026, two brand setups asked the same 40 buyer questions daily on four engines. Of 2,127 cases where a tracked name was in one day's answer, it was gone the next day in 486, 22.8%. Of the changes we could follow, 47.2% reversed the following day. Yet the weekly average barely moved: a single day's rate sat a median 3.6 points from the week's (the post).
Does rewording a question move answers more than asking it again?
Yes. In 13,994 answers, a brand named on one phrasing was still named when the same phrasing was asked again 77.5% of the time, but named on a sibling phrasing of the same buying decision only 55.1% of the time (the post).
Some of that is question shape. Head-to-head questions such as "A vs B" named a third brand in only 2.9% of answers, while "best" questions and questions with a price or team-size constraint named it far more often. Volatility is partly the engine and partly the words you chose.
What does volatility look like in a real example?
Tesla's grid storage topic in the public Tesla example (Oct 4, 2026, re-asked October 8 and 15) shows how far phrasing alone moves the answer. Each phrasing got 12 fresh answers across four engines:
- "What are the best grid-scale battery storage systems for a utility company?": Tesla named in 12 of 12.
- Best alternatives to a named competitor for utility-scale storage: 12 of 12.
- A peak shaving question for a large facility: 7 of 12, and none on Claude.
- A cost per MWh question and a question about comparing turnkey providers: 0 of 12 each, with no supplier named at all.
Do different engines change together?
No. When one of four engines named a brand on a question, all four did only 32.4% of the time across 185 question days in our sample, even though three engines had nearly identical average rates (the post).
Each engine agreed with itself a day later more than it agreed with any other engine on the same day. So volatility has to be read per engine: a blended number can stay flat while one engine drops you and another picks you up.
How should you act on a change in AI answers?
Treat one answer flipping as a draw, not a result. Act when a change holds for two or three days or when the intervals of two readings separate, and compare the same questions on the same engines.
Proofsource alerts follow that rule: an alert fires only when the old and new Wilson intervals do not overlap and the move clears your threshold. See Alerts and daily AI visibility monitoring.
Common questions about AI answer volatility
Why did ChatGPT stop recommending us overnight?
Most likely noise. In our tracking about 23% of names in an answer were gone the next day, and nearly half of those changes reversed a day later. Wait for the change to hold before you rewrite anything.
Is volatility the same on every engine?
No. Engines differ in how often they search, how many sources they read and how much their answers repeat. Measure volatility per engine on your own questions.
How do I reduce volatility in my numbers?
You cannot change the engines, but you can change the measurement: a fixed question set, several phrasings per decision, repeats on separate days and an interval on every figure.
Does high volatility mean AI visibility cannot be measured?
No. Single answers churn, but a fixed set of questions moves far less, because most answers stay put and the changes cancel out. Measure the set, not the answer.
Related terms
- AI answer sampling
AI answer sampling is asking the same questions on each AI engine more than once and reading every visibility figure as an estimate from that sample, not a fixed fact.
- Wilson interval
A Wilson interval is the range a measured percentage could plausibly sit in, given how many answers it is based on.
- Prompt
In AI visibility, a prompt is a question tracked across AI engines, written the way a buyer would ask it.
- Mention rate
Mention rate is the share of AI answers to a question that name your brand.
Worked example: Tesla, public AI answers, Oct 4, 2026: Proofsource's first scan of Tesla, 25 buyer questions and 5 brand questions, one answer per question on each of four engines. A small sample from one brand, shown as an example, not a benchmark.
See where AI leaves you out, and what to fix.
See which buyer questions name your brand, who gets named instead, and what to fix. 200 answers. 25 Prompts/Question. Top AI engines. 2 days. Free. No card.
Google AI OverviewsGoogle AI Mode
