How to report AI visibility honestly: six rules for the numbers
In short
An AI visibility figure means nothing until it says what it divides by. Keep questions that contain your name apart from the ones that do not, because the first kind almost always names you. Call a change only when it is bigger than the usual swing, and let every figure lead back to an answer someone can read.
Why this needs rules
The same answers can give very different figures, depending on what you count. The example below is made up to show how. It is not a measurement of any real company.
Say a brand asks 40 questions on four assistants, so it gets 160 answers. Ten of the questions contain its name, which gives 40 branded answers. The other 30 questions give 120 answers without its name.
| Counted over | Answers that named the brand | Share |
|---|---|---|
| Every answer | 40 of 160 | 25% |
| Answers to questions with its name | 34 of 40 | 85% |
| Answers to questions without its name | 6 of 120 | 5% |
Nothing is wrong with the arithmetic in the first row. It counts questions like "How much does Brand X cost?", which name the brand in the question and get its name back. Those tell you nothing about whether a buyer who has never heard of you will find you. The honest figure is the last row.
The six rules
Every number under each rule is a made-up example, chosen to show the rule. None is a measurement.
1. Say what you divide by
Write every figure with both counts beside it. Write "named in 6 of 120 answers", not "5 per cent". A percentage on its own hides whether it came from 5 answers or 500, and whether it counts answers, questions or mentions.
2. Keep questions with your name apart
A question that contains your name is answered with it almost every time. In the example above, that is 34 of 40 answers. Mixing those in turned 5 per cent into 25 per cent without a single buyer finding the brand.
3. Never compare against a figure you did not measure
If you only searched the answers for your own name, a rival shown at 0 of 120 was never looked for. It does not mean they were absent. Compare only brands you actually counted, on the same questions.
4. Call a change only when it beats the swing
Assistants answer the same question differently on different days. A 2026 preprint, CITECHOICE, found that about one in seven regenerations flipped whether a source was cited. It has not been peer reviewed. So measure how much your answers move on their own before calling anything a trend.
Say the answers swing by up to 16 percentage points on their own. A move from 2 of 35 answers to 9 of 35 is a rise of about 20 points, which beats the swing. A move from 2 of 35 to 6 of 35 is about 11 points, which does not.
5. Two readings before a trend
One scan is a reading. A trend needs at least two scans that agree, asked the same questions. Say a brand goes from 3 of 120 to 6 of 120 in one month. That is one step, not a trend. If the next month also shows about 6 of 120, the step is more likely real. Changing the questions between scans resets the comparison.
6. Let every figure lead back to an answer
Anyone reading your report should be able to open the answer behind a number and read it. If "6 of 120" cannot be opened to the six answers that named you, nobody can check it, including you.
A report you can copy
Here is the shape of a monthly report that follows the rules. The values are made up to show the format. Replace them with your own counts.
| Line | Example value |
|---|---|
| What we asked | 30 buyer questions without our name, to 4 assistants |
| Answers that named our brand | 6 of 120 |
| Answers that named the leading rival | Rival A in 41 of 120 |
| Questions with our name, for contrast | 34 of 40 answers named our brand |
| Change since last scan | Up from 3 of 120. Too small to tell against normal variation |
| Most-cited site | A review site, cited 22 times |
| One answer to read | Linked, word for word |
A report that says "too small to tell" is more useful than one that draws a line through noise.
If this is your first scan, write "no earlier scan to compare" on the change line. That is more honest than any trend drawn from a single point.
How Agentcites follows these rules
- The headline figure is the share of answers that named you, counted only on questions without your name, with both counts beside it.
- Questions with your name are labelled and kept apart in the question lists.
- A change is marked "Too small to tell" when it is within the measured swing, once swing sampling is on.
- Every figure opens to the answers behind it, word for word.
The method page explains how every figure is counted.
Questions
How do I report AI search performance to my marketing team?
Report the share of answers that named you on questions without your name, with both counts. Add the leading rival on the same questions, one answer to read, and whether the change since the last scan beats the usual swing.
What is a good share of voice in AI search?
There is no fixed benchmark. Compare yourself with the rivals you counted on the same questions, and make sure the figure leaves out questions that contain your own name.
Why is my share of voice so much higher in one tool than another?
Often because one counts questions that contain your name and the other does not. In the made-up example on this page, that difference alone turns 5 per cent into 25 per cent.
How often should I measure?
Often enough to see a change, and always on the same questions. Weekly is common. Whatever the gap, measure the usual swing before you call a change.