AI Mention

How is an AI mention verified rather than guessed?

By checking the answer text, not by asking a model. The usual approach is to have a second model read the response and list the brands it saw, and that approach fails in two directions at once: it invents brands from a saved competitor list that are not in the text, and it drops unfamiliar names that are.

RankX AI bounds it with two deterministic passes. The first drops any mention whose brand string does not actually appear in the response text, so a fabrication cannot survive. The second captures every tracked competitor that is visibly present in the text regardless of what the model said, so an unfamiliar name cannot be dropped. Neither pass invents anything: the first only removes, the second only adds text-present brands. Neither pass is clever, and that is the design rather than a shortcut. The parts of this pipeline that can invent things are the parts a model runs, so the parts that decide what actually gets stored are ordinary string matching against text a person can read back.

Why is a missing verdict not a zero?

Because “we looked and you were never named” and “we have no reading” are different facts, and only one of them is bad news. In the product, a check with no determination is stored as null and is never folded into the not-mentioned count. The denominator for a rate is the checks with a verdict either way, never the checks that ran.

This is not a fussy distinction. On live data at one point, 1,023 of 1,210 tracked answers carried no verdict. Dividing by the larger number would have understated a real mention rate by around 85%, and reporting zero where the truth was “unmeasured” would have told a customer they were invisible when nobody had looked. Both were one careless line away, in a report going out over an agency’s letterhead.

What can an AI mention count never tell you?

Why. An assistant does not report which page, which sentence or which third-party source put your name in an answer, and where the name came from training rather than retrieval there is nothing to point at anyway. A mention count tells you the outcome moved, and the attribution is inference.

The recommendation that follows, and it is a position: treat mention counts as a trend line rather than as a scoreboard, and pair them with the one thing you can attribute, which is whether a page of yours was cited. A rising mention count with no citations behind it usually means the answer learned about you somewhere other than your own site, which is a public-relations finding rather than a content one.

Related terms

• AI CitationAn AI citation is a link an AI assistant attaches to its answer, naming a page as a source.
• AI Share of VoiceAI share of voice is the proportion of AI answers naming your brand, measured across a panel of tracked prompts and compared against the rivals named in the same answers.
• AI Brand SentimentAI brand sentiment is how favourably an AI assistant describes a brand when it names one: recommended, listed neutrally, or named as the option to avoid.
• Prompt VolumePrompt volume is how often people put a particular question to an AI assistant, the assistant-side equivalent of search volume.