How competitor sets get built
Share of voice in AI answers is a ratio: how often you are named, against how often anyone in your category is named. The denominator is usually assembled automatically, by extracting company names from the answers themselves.
That extraction has no way of knowing that two strings are the same company.
What that does to the number
A bank appearing sometimes as 'Alior' and sometimes as 'Alior Bank' enters the set as two separate competitors. Every such pair inflates the denominator - the category looks more crowded than it is - and your share comes out lower than it is.
On the project where we caught it, merging the variants moved the measured share from 2.12% to 3.73%. Nothing about the client's visibility changed. The previous number had simply been wrong.
Why it survives so long unnoticed
The error is invisible in the direction that would make anyone check. A deflated share looks like a problem to solve, not like a bug to investigate, and it produces exactly the kind of report that leads to more budget rather than to a second look.
It also compounds quietly: the more spellings a category uses, the worse the distortion, and categories with long formal company names are the worst affected.
What we do about it
The competitor set gets reviewed by a person every cycle, and name variants get merged before anything is counted. It is a dull ten minutes that decides whether the rest of the report means anything.
It is also the reason we are cautious about any AI visibility figure produced without that step - including figures that flatter us.