In one sentence: brand share in AI answers is the proportion of a fixed set of buyer questions where a generative engine names your brand, measured against the same competitor set, on the same questions, run the same way each time.
The definition is easy. Getting a number you would defend in front of a client is not. Every dashboard in this category reports something called share, and most of those numbers are higher than the truth. Below are the four failure modes we know best, because our own platform shipped all four and we had to fix them.
Failure 1: questions that name your brand
If the question is “is Acme any good?”, the answer will say Acme. That is not visibility; that is an echo. A prompt set assembled by an LLM will drift toward these on its own — ours once produced fifteen questions of which thirteen contained the brand name, and the resulting score read 13% across five engines when the true value was zero.
A self-referential question measures nothing except whether the engine can read the question. Exclude it from the numerator and the denominator, and check the exclusion at generation time, not only at scoring time.
The fix has to live in three places: the generation prompt must be told not to write them, a deterministic filter must catch the ones it writes anyway, and the fallback template set must be clean. A filter alone still burns the quota on questions you will throw away.
Failure 2: counting yourself differently from competitors
The classic version: your brand matches on any substring, competitors match on exact name. Or your brand is detected by an LLM pass while competitors are detected by a word-list. Or a normalizer folds “Acme” and “Acme Systems” into one entity for you and leaves the competitor split across two rows.
Any asymmetry moves in one direction, because you built the detector while looking at your own name. The rule is that the same code path, in the same pass, decides presence for every brand in the set — and ties resolve against you, not for you.
Failure 3: negative mentions counted as presence
An answer that says “Acme is often criticised for X, so most buyers choose Y” contains your brand. Whether that counts depends on what the number is for. Mention rate and recommendation rate answer different questions, and collapsing them makes the metric useless exactly when it matters — the moment a competitor comparison starts going against you.
Report them as two columns. A brand can be mentioned in 40% of answers and recommended in 4%.
Failure 4: unmeasured rendered as zero
A question that failed to run, an engine that timed out, a week with no scheduled run — none of those are zero. Rendered as 0% they pull every average down and make a flat trend look like a decline. Rendered as blank they tell the truth. This sounds pedantic until a client asks why their score dropped in a week when nothing was collected.
How to compute a number you can defend
| Decision | Rule |
|---|---|
| Question set | Fixed and versioned. No brand names in the question. Change it and the series starts over. |
| Competitor set | Declared in advance, identical detection for every member including you. |
| Engines | Fixed list, recorded per run. Adding an engine is a methodology change. |
| Citations | Only sources actually marked in the answer body. A retrieval pool is not a citation — counting it inflates the figure several times over. |
| Missing data | Null, never zero. Show the gap. |
| Cadence | Same interval, same time of day, results anchored to the run timestamp rather than to when the job finished. |
None of this makes the number look better. It makes it survive being checked, which is the only property that matters once someone is paying for it.