Why Report AI Visibility as “Recommended in 3 of 5 Scans” Instead of a Percentage?

Why does a small denominator change everything?

AI answers vary from run to run. If a question is asked five times and you are recommended three times, the true rate could plausibly be anywhere from about a quarter to well over three quarters. Writing "60%" invites the reader to compare it with last month's "40%" as if the difference meant something, when both are consistent with the same underlying rate.

A count keeps the uncertainty visible. "3 of 5" and "2 of 5" read as close, because they are. "60%" and "40%" read as a large drop.

How uncertain are small counts?

The ranges below are 95 per cent intervals for the underlying rate, using a standard method for proportions. They show why a percentage from a handful of runs says less than it seems to:

These ranges assume each run is independent; runs of the same question tend to resemble each other, so the real uncertainty is, if anything, wider. Two things stand out. Even a perfect "5 of 5" is consistent with a brand that appears only a little more than half the time. And the range only becomes tight once the base reaches the hundreds, which is a total across many questions, not a single question.

What might this look like in a client report?

Imagine a report that says a client's recommendation rate on one question fell from 67% to 33%. Written as counts, it reads "2 of 3 runs last month, 1 of 3 this month", and the table above shows both results overlap almost entirely. The client hears "no clear change on this question" instead of "we halved". The same report can still say that across all 300 answers the client appeared in about 30 per cent. Because each question's three runs tend to agree, treat that as accurate to roughly ten points either way, not the five a simple formula gives; it is still far firmer than any single question's result.

When is a count better?

• Per question. Each question usually has only a handful of runs per engine, so counts are the honest unit.
• Per engine on one question. "Named in 2 of 3 runs on Perplexity, 0 of 3 on Gemini" is clear and points to a source difference.
• For non-specialist readers. Clients and executives understand "3 of 5 times" without explanation.
• When the denominator varies. If some questions had fewer valid runs, for example because one engine failed or Google showed no AI Overview, counts show it; percentages hide it.

When is a percentage better?

• For totals across many answers. Across 20 questions, five engines and three runs, a percentage summarises hundreds of answers sensibly.
• For comparisons across sets of different sizes. Comparing two brands measured on the same large set is easier in percentages.
• For trends over time, once the base is large and stable.

Even then, state the base: "appeared in 31% of 300 answers" tells the reader far more than "31%".

What goes wrong with each?

Percentages from small bases create false precision and dramatic-looking swings. They also tempt people to average per-question percentages with different denominators, which gives questions with few runs as much weight as questions with many.

Counts without context can be hard to compare across questions with different numbers of runs, and they make a portfolio total awkward to read.

Both go wrong when "mentioned" and "recommended" are merged. Being listed among "other options" and being put forward as a choice are different outcomes and should be counted separately.

How should a report combine them?

1. Use counts at the question level: named, recommended and cited, out of the runs actually completed, per engine.
2. Use percentages with their base for totals, and only across the same question set.
3. Keep brand-name questions out of the totals, because they inflate visibility.
4. Show which runs did not happen or produced no AI answer, rather than counting them as losses.
5. Mark changes that are too small to mean much, so readers do not act on noise.

For more on the arithmetic, see how to calculate AI visibility as an appearance rate across repeated runs and whether 20 monthly queries are enough to detect trends.

What does a reader need to see next to the number?

Whatever the format, a visibility figure is only as useful as the context printed beside it. At minimum a reader should be able to find:

• Which questions produced it, and whether any name the brand.
• Which engines and settings were used.
• How many runs each question had, and how many completed.
• What counted as a mention, a recommendation and a citation.
• When the answers were collected.

Without these, "60%" and "3 of 5" are equally hard to trust. With them, either format can be read correctly.

What does Kovmere measure?

Kovmere asks up to 20 questions for one market, up to 18 buyer questions that do not name you and up to 2 brand checks kept separate, three times on each of OpenAI, Perplexity, Google Gemini, Anthropic Claude and Google AI Overviews in a baseline. It measures three outcomes separately: mentioned, where your name appears as a whole word; recommended, judged only after a mention; and cited, where your own domain is among the sources. When Google returns results without an AI Overview, that is recorded as its own state rather than scored as zero. On Pro, single-run re-tests follow in weeks two to four. We do not describe here how its report presents the recommendation figure.

What do people ask about counts and percentages?

Is "3 of 5" more accurate than "60%"?

It is the same result. It is more honest, because it shows how little evidence the number rests on.

How many runs make a percentage meaningful?

There is no fixed threshold, but percentages become steadier as the base grows into the hundreds. For a single question with a few runs, counts are better.

Should failed runs count as zero?

No. A run that did not complete, or a search that showed no AI answer, is missing data, not a loss. Count it separately.

Can I compare my count with a competitor's percentage?

Only if both come from the same questions, engines, runs and settings. Otherwise they measure different things, however similar the numbers look.