About one name in five dropped out of an answer overnight
Most advice on scan cadence says weekly, and explains that daily readings are noisy. We wanted to see how noisy. Between October 5 and 13, 2026, two brand setups in Proofsource asked the same 20 buyer questions each, 40 in all, on ChatGPT, Perplexity, Claude and Google AI Overviews once a day: nine days for one setup, seven for the other. Each setup tracks its own brand and eight or nine competitors, 19 names in all.
That gave us 10,612 cases where we could compare one question, on one engine, for one name, on two consecutive days. On 2,127 of them the name was in the first day's answer. On 486 of those, 22.8%, it was gone the next day (95% interval 21.1% to 24.7%). Brands also arrived: in all, 969 of the 10,612 day pairs changed.
A wider check across 53 workspaces gave the same rate. Of 5,024 questions answered by the same engine on two consecutive days, 427 named the workspace's own brand on the first day, and 102 of those (23.9%) didn't on the second. Another 87 named it on the second day only, so the total went from 427 to 412. Plenty of churn underneath, almost no change on top.
47% of the changes undid themselves a day later
For 832 of those changes we also had the answer from the following day. In 393 of them, 47.2%, the name had gone back to where it started: in, out, in again, or the reverse. Where we could follow a change for two more days (686 cases), it held for all three days in 252, or 36.7%.
So when one answer stops naming you overnight, the odds are close to even that it names you again tomorrow. Treat a change in a single answer as a draw from a distribution, and wait for it to hold before acting on it.
A single day landed within 3.6 points of the week's average
The aggregate tells a calmer story. For the seven days both setups shared (October 5 to 11), we took every name and engine with a weekly mention rate of at least 5%, 55 of them, and compared each day's rate (20 answers) with that week's average. The median gap was 3.6 percentage points. 27 of the 385 daily readings (7.0%) were 10 points or more away, four were 15 or more, and the largest gap was 20.
The gap is smaller than coin flips would give, because the same questions come back every day and most of their answers stay put. The churn mostly cancels out across a fixed set.
But the bad weeks are real. One tracked competitor on ChatGPT was named in 75% of answers on Monday the 5th and 45% on Friday the 9th; read on those two days, that looks like a 30-point fall. Across the whole week it was named in 89 of 140 answers, 63.6% (55.3% to 71.1%), and nothing had fallen. On 20 answers, 45% carries an interval of 25.8% to 65.8%, which is the honest way to read a single day.
Choose the cadence by the decision it feeds
A daily scan collects seven times the answers a weekly one does, and costs about seven times as much to run. What that buys depends on what you'll do with the number.
- Reporting a trend to leadership on a large, fixed set of questions: weekly is usually enough. Even on our 20-question sets, most single days sat within a few points of the week, and a larger set sits closer.
- Tracking 20 questions or fewer: each reading is small, so one odd day moves it a lot. Daily scans averaged over the week give you seven readings instead of one, and an interval about 2.6 times narrower.
- Testing a change you just shipped: daily tells you which day the answers moved and whether the move held. Weekly can put an unrelated wobble in the week after your launch.
- In every case, don't act on one answer flipping. Wait for a change to hold for two or three days, or for the intervals of two readings to separate.
How Proofsource handles it
A Proofsource plan runs daily or weekly, chosen per plan. On a daily plan each question is asked once on each engine per UTC day; weekly plans run once a week. Every mention rate carries a Wilson interval, and an alert fires only when two readings' intervals don't overlap and the move clears the threshold you set, so a single day's wobble doesn't page anyone. That rule is conservative and can sit out a small real move for a few days.
Where this could be wrong
The long series come from two brand setups in two categories, 40 questions, over one week in October 2026, and only 55 names were common enough to compare with their week. Other categories, or a week with a model update, could churn more. The 53-workspace check covers only two consecutive days per question.
A "change" here means a tracked name appearing or not, as our extraction reads it. An extraction miss on one day looks like churn, so some of the 22.8% may be ours rather than the engine's. We counted each question once per engine per day, and we dropped 116 day pairs whose answers were byte-identical across days, because those were replays, not fresh answers.
The answers on the same question are related from day to day, which is why the week moved so little. A brand whose questions change daily, or that samples a fresh set each week, would see bigger swings.
