How accurate are AI visibility numbers?

This guide measures how much of an AI visibility number is chance. On 29 October 2026, one rival in our own project fell 11.5 points in a day and recovered the next day. The guide gives the margin of error of a day, a week and 28 days, measured in the 28 days to 30 October 2026. It also shows what each window hides. The deliberate choices behind a number, the counting unit, the platforms, the prompts and the rivals, are in our guide on why the numbers differ. The formula itself is in our guide on the visibility score.

Why do AI visibility scores differ between tools?

AI visibility scores differ between tools in part by chance, because each score is a sample estimate with a margin of error. The margin of error is the half-width of a 95% confidence interval around a measured share. An interval built this way contains the true share in 95 of 100 samples. In the 28 days to 30 October 2026, one day of our project held 329 answers. Two tools that each read 329 answers for a brand near 41% can differ by 7.5 points by chance alone. Over a week of about 2,300 answers each, chance accounts for up to 2.8 points. A tool that reports a number without its answer count hides the margin.

How much does a daily number move by chance?

In the 28 days to 30 October 2026, a rival's median daily step over all 7 platforms was 1.2 to 2.1 points. Our own project tracks the category we sell in, AI visibility tools. It asks each prompt on 7 AI platforms every day, and it monitors five rivals. In the 28 days to 30 October 2026, it stored 8,953 answers, about 320 a day. The table gives each brand's daily visibility over those 28 days. A step is the change from one day to the next, in points. A week is one of the four 7-day blocks of the window. The daily series holds the four rivals named most, so AthenaHQ has no row here.

The lowest and the highest day of a brand sit 6 to 12 points apart. Its lowest and highest week sit 1 to 4 points apart. Otterly AI's lowest day, 29.5% on 29 October, followed 41.0% and preceded 41.6%.

How many answers does a stable number need?

A number with a margin of 2 points needs about 2,300 answers at a visibility near 40%. We computed each margin as the 95% binomial interval for a share of 41%, Profound's level, and again for 14%, Scrunch's level. The table gives the margin for the four sample sizes our project produces.

Read these margins as a floor. The formula treats every answer as an independent draw. Our guide on why the numbers differ found that only 8% of one brand's same-prompt pairs on consecutive days changed. So a run of days holds fewer independent draws than answers.

The measured spread agrees with the floor. Profound's daily figure over all platforms stayed within 5 points of its 28-day figure on 27 of the 28 days. Each of its four weeks stayed within 2 points of it.

More prompts give more answers a day, so the margin shrinks without more days. Our guide on how many prompts to track sizes the set. Our guide on a prompt set from Search Console shows where the prompts come from.

What does one platform's daily figure hide?

One platform on one day holds about 47 answers in our project. Its figure can move by 10 points or more in a day and return the next day. On Gemini, Otterly AI appeared in 22 of 47 answers on 28 October. It appeared in 7 of 47 on 29 October, and in 16 of 47 on 30 October. On all platforms together, the same brand read 41.0%, 29.5% and 41.6% on those three days. Over the 28 days, Profound's median daily step on Gemini was 4.3 points, twice its step over all platforms.

On Perplexity, Profound appeared in 4 of 44 answers on 5 October and in 14 of 47 on 6 October. It stayed between 23% and 32% on each of the 24 days after that. That jump held for more than three weeks, so it is a change. The dip on Gemini reverted the next day, so it is noise. Gemini gave 22 answers on 5 October, half its usual 44.

What do a 7-day and a 28-day window each hide?

A window shows one total, so it hides the events inside it, such as a missing platform or a changed setting. A dashboard tile compares the picked range with the range of the same length before it.

Home›Overview

Example

Visibility, 7 days▲5.8 pts vs previous 7 days

41.1%

210 of 511 answers

Visibility, 30 days▲3.2 pts vs previous 30 days

35.8%

784 of 2,190 answers

Every 7-day change in the table is smaller than 2.8 points, the margin of a change between two weeks. None of them passes that margin. The 28-day figures agree with the 7-day figures within 1 point. The 28 days also held three events that the totals do not show:

• Copilot returned no answer on four days, 6 to 9 October. The project's daily answer count fell from 329 to 282. Profound's figure over all platforms stayed between 39% and 42% on those days. Copilot is its strongest platform, at 55% over the 28 days.
• Each platform that answered gave 44 answers a day until 5 October and 47 a day from 6 October. The one exception is Gemini on 5 October, with 22. The prompt set grew by three.
• The project collected its first answers on 25 September, so the 28 days before the window hold 8 days of data. In those days the prompts ran in Ireland. They run worldwide now. Profound's 28-day change reads +13 points. It measures the setting, not the market.

A failed attempt counts in no denominator, and the platforms fail at different rates. In the 28 days, Copilot needed 786 failed attempts for its 1,119 answers, Perplexity 226 for 1,303, and Gemini 47 for 1,307. A tool that does not ask again after a failure has fewer answers from Copilot than from Gemini, and its total leans toward Gemini.

How do you read a change against its margin?

Read a change against the margin of a change, and then against the answer count behind it. A change compares two samples, so its margin is about 1.4 times the margin of one sample. Three reads follow from our numbers in the 28 days to 30 October 2026:

• The margin of a change between two days is about 7.5 points at 41%, over all platforms. Profound's median step was 2.1 points. Its largest was 7.6 points, from 45.9% on 28 October to 38.3% on 29 October. It read 44.4% the next day. A 95% margin lets about one step in 20 pass by chance, and Profound had 27 steps.
• The margin of a change between two weeks is about 2.8 points. The largest week-to-week change was 3.3 points, for Otterly AI, between the first and the second week. The first week lacked Copilot on four days, so the two weeks did not hold the same platforms.
• A number over fewer answers than usual has a wider margin. 282 answers instead of 329 means one platform is missing.

A change that passes these reads still needs a cause. Our guide on why the numbers differ lists the causes to check.

What to do next

1. Write the answer count beside every number you report, with its window and its platform count.
2. Test each change against the margin of a change at your answer count, not against the margin of one number.
3. Pick the metric for the decision before you track it. Our guide on AI visibility metrics says which metric supports which decision.

For how FullReach AI collects the answers, read the methodology.