Sentiment Analysis

Sentiment analysis is the automatic classification of a piece of text by the attitude it expresses: positive, negative, neutral, or a finer scale between them. In measuring AI visibility it is applied to the sentence in which an assistant names a business, because being named and being recommended are different outcomes.

How a mention is read

For each answer in a question panel this site records whether the business was named, whether it was named first, and who else was named; it does not score tone. A name in a list of options, a name followed by a caveat, and a name given as the first recommendation are three readings of the same bare fact of being mentioned, and telling them apart is the sentiment step. This site’s score counts all three as named and reports being named first beside it; the scoring is on the methodology page.

Where the classifier is not enough

Hedged answers. In this site’s own one-question probe of 9,240 small businesses, 5,928 answers gave a clearly differentiated list and the rest hedged, and tone inside a hedge is hard to read by machine; the probe is described in what 9,483 small business websites serve to an AI crawler. Comparisons, a positive sentence about a competitor in the same paragraph, and praise with a condition attached all defeat a naive label. That is why every answer is kept verbatim: a tone classification that cannot be checked against the sentence it came from is not a measurement.

A common question

Why keep every answer verbatim if a classifier can already label the tone?

Because a tone label nobody can check against the sentence it came from is not a measurement. Hedged answers are the hard case: in this site's one-question probe of 9,240 small businesses, 5,928 answers gave a clearly differentiated list and the rest hedged, and tone inside a hedge resists machine reading. Comparisons, a warm sentence about a rival in the same paragraph, and praise with a condition attached all break a naive label, so the wording of each answer is kept.

Sources

The probe counts are this site’s own, from one assistant (Claude with retrieval), one question per business, one day; directional, not a rate. Scoring rules from the methodology page.