Skip to main content
Every metric counts what the analysis finds in the stored answers. The analysis runs after each collection. It does two separate things: it finds the brands, and it labels the tone of the answer about each one.

Finding a brand

A brand counts in an answer when the answer writes its name or one of its other spellings. That match is exact. One kind of match is judged instead, as the section on spellings below says.
  • Capitals do not matter, except in a short spelling. A spelling of 3 letters and digits or fewer matches only in its own case. Otherwise a brand named “Ace” would match every “ace”.
  • Spaces, dots, hyphens and underscores inside a name are one character. So “Peec AI”, “Peec.ai” and “peec-ai” are one name.
  • A capital inside a name can take a space before it. So “OtterlyAI” also matches “Otterly AI” and “Otterly-AI”. After the space, the capital must stay a capital. So “LinkedIn” never matches “linked in”.
  • A name never matches inside a longer word.
  • The analysis reads the first 9,000 characters of each answer. A brand that the answer names only after that point does not count. Most answers are shorter.
  • A name in a citation chip only does not count. Some platforms print a source’s name beside a sentence. One mention in the answer’s own text is enough.
  • A brand counts once per answer, however often the answer names it.
The answer view highlights every brand that the match found. The match also records where the answer first names each brand. That place gives the brand its position in the answer. No model decides the place.

Tone

For each brand that the match found, a language model reads the answer. It gives the brand one sentiment label: positive, neutral or negative, the tone of the answer about the brand. The model must quote the words behind the label, exactly as the answer writes them. When the quote is not in the answer, the analysis discards the label. The mention then stays unclassified. An unclassified mention still counts for visibility, Share of Voice and position. It stays out of the sentiment shares. The model never decides whether a brand is present. The match decides that.

Brands that nobody monitors

The model also lists the other brands that an answer names. They become detected brands on the Competitors page. A detected brand counts in no figure until somebody monitors it. It still takes a place in each answer that names it, so it can move a monitored brand down in position.

Spellings the project does not hold

The model can also find a monitored brand under a spelling that the project does not hold, such as “PeecAI” for “Peec AI”. A check then compares the spelling with the brand’s name. It passes when one holds the other, when the spelling is the name’s initials, or when the two differ by a small typo.
  • A spelling that passes the check counts. The mention is a judged match.
  • A spelling that fails the check does not count.
The Competitors page lists both kinds under the brand, with the answers that used them. Adding a spelling makes its match exact from then on.

When a brand or a spelling is added later

  • A new competitor or a new spelling counts in each new answer. When you add one, the answers already collected are counted too, and the figures update when that count ends. Visibility, Share of Voice and position change for past days too.
  • Sentiment starts from the next collection. The answers collected before stay unclassified for that brand. So for some time a new competitor’s sentiment covers fewer answers than its visibility does.
  • Removing a spelling does not remove past matches. An answer does not record which spelling it matched.
  • A brand that nobody monitors any more leaves every figure. Its mentions stay stored, and they return when somebody monitors it again.