Often it is just noise. AI engines rarely give the same answer twice, so a visibility score is an estimate built from a sample of answers. A weekly move of several points can come from chance alone. Treat a move as real only when it passes four checks:
- It is bigger than the margin of error.
- It lasts for several weeks.
- It shows up across several prompts.
- Nothing else explains it: no engine-wide change and no edit to your own setup.
Below, we explain each check and what three GEO (Generative Engine Optimization) tools, Genezio, Peec AI and Profound, document to help. We did not test the tools hands-on; the comparison uses their public documentation, checked on 27 October 2026.
Why does a visibility score move when nothing changed?
Because the same prompt gets a different answer each time. A visibility score, the share of answers that mention your brand, is a sample, not a fixed reading.
A preprint from the University of St. Gallen re-ran the same prompts on ChatGPT, Gemini, Google AI Mode and Perplexity within a single day (Schulte et al., 2026). Two answers to the same prompt shared only about a third to a half of the brands they named. The authors concluded that this same-day randomness "alone accounts for most of the observed instability" from day to day. The study is not peer-reviewed, and its first author is affiliated with Aurora Intelligence, the GEO data company that supplied the data.
Vendors agree. Peec AI's help center says "daily and weekly fluctuations are completely normal and expected" (Peec AI).
Check 1: Is the move bigger than the margin of error?
It depends mostly on how many answers sit behind the score. With a few dozen answers, chance alone can move it by more than 10 points.
The margin of error is the ± figure around a score. The score plus or minus that margin is a 95% confidence interval. If you repeated the measurement many times, about 95 of every 100 such ranges would contain the true rate.
The standard textbook formula for a percentage shows how the margin shrinks with more answers. Suppose your brand appears in 40% of 50 answers. The margin is about ±14 points, so the true rate could be anywhere from about 26% to 54%.
| Answers behind a 40% score | Approximate 95% margin of error |
|---|---|
| 50 | ±14 points |
| 200 | ±7 points |
| 800 | ±3 points |
Halving the margin takes four times as many answers. Treat these figures as a floor: the formula assumes independent answers, but answers to the same prompt are related, so the real margin is usually wider.
What this means: a change between two weeks must clear a wider bar, because both weeks are noisy. The margin for the change is about 1.4 times the single-week margin. With 200 answers a week, that is about ±10 points, not ±7.
Checks 2 and 3: Does the move last, and is it broad?
A move that clears the margin once can still be a fluke.
Lasting matters because each extra week adds answers and narrows the margin. The St. Gallen authors recommend rolling averages over two to four weeks for per-brand estimates. That is one study's advice for one market, not a universal standard.
Breadth matters because prompts differ in stability. In the same study, some prompts named nearly the same brands every time, while others changed almost entirely. A move driven by one or two volatile prompts is likely churn; a move across many prompts is harder to explain by chance.
Check 4: Could an engine change or your own setup explain it?
Even a large, lasting, broad move may not be about your brand. The engine may have changed for everyone, or you changed what the score is calculated from.
A Genezio research recorded two engine-wide shifts in October 2026, each on a matched set of prompts:
- Google AI Mode listed about a third fewer sources per conversation across the same 1,083 questions. A re-check of the raw pages put most of the drop on Google's side.
- Claude listed about a third more sources per answer after some of its web searches began returning up to 20 results instead of 10. That was measured on about 1,100 matched prompts in late October, only two days after the change.
Both concern sources, not brand mentions, but show how fast an engine can change. If competitors and your whole prompt set moved on the same day, the engine probably changed. A few control prompts, unrelated to your recent work, make this easier to spot. Changes in how a tracking tool reads answers can also create look-alike trends.
Your own edits can move the number too. Peec AI notes that "visibility often drops after adding many new prompts," because teams add harder prompts over time (Peec AI). Profound applies your current topics, tags and brands to all stored answers, so excluding a brand or moving a prompt can change past numbers. It advises exporting figures at reporting time (Profound docs). The safest comparison uses a fixed set of prompts and competitors.
What do GEO tools show to help you check?
Genezio documents an uncertainty range; Peec AI and Profound document markers for changes.
| Genezio | Peec AI | Profound | |
|---|---|---|---|
| Uncertainty shown with the score | "Accuracy range" at a 90% level by default, not 95%; method not documented | None documented | None documented |
| Data behind the score | Recent conversations until the range is narrow enough | Responses in the selected period; frequency not documented on the metric page | Prompts run daily |
| Markers for your own changes | Not documented | Dotted lines when prompts are added | Manual chart annotations |
| Markers for engine changes | Not documented | Not documented | Model changes added automatically |
| Past numbers after setup edits | Not documented | Not documented beyond the prompt note | Recalculated from current setup |
Sources: Genezio, Peec AI metric, Peec AI help, Profound changelog (markers since 9 October 2026), Profound docs. None of these ranges or markers has been independently checked.
So when is a move worth acting on?
When it passes all four checks: it is bigger than the margin of error, it lasts for several weeks, it shows up across several prompts, and nothing else explains it.
A displayed uncertainty range helps with the first check; change markers help with the last. No tool documented here covers all four, and none replaces reading the answers. Even a move that passes every check shows that something changed, not why: linking it to your content work needs more than correlation. For the metrics themselves, see our guide to GEO metrics.