Mentions and sentiment

A mention is the moment your brand shows up by name in an LLM response. Sentiment is what tone the LLM used when naming you. Both feel obvious until you try to count them at scale.

What counts as a mention

A mention is recorded when:

Your brand name (or one of its aliases) appears in the response text
Surrounded by word boundaries (so "tic-tac-Apple" doesn't match "Apple")
Case-insensitive by default (case-sensitive on demand for brands like "RAG" that need it)

The match is deterministic: same response, same brand config, same mention count every time. No ML hand-waving on the count.

Aliases matter

For most brands, the canonical name catches ~70% of mentions. Aliases catch the rest.

Examples of aliases that materially change mention counts:

Legal entity name — "Cognigy" vs "Cognigy GmbH"
Common abbreviations — "GitHub" vs "GH"
Product line vs company — "Workspace" vs "Google Workspace"
Misspellings the LLM produces — "GetVocal" vs "Get Vocal"
Domain references — "getvocal.com" when no company name is used

The AI alias suggester (the "Suggest with AI" button on brand-edit forms) uses Tavily web-search grounded against your brand's website, Wikipedia entry, news, and Crunchbase to surface real aliases — not LLM hallucinations. Niche B2B brands especially benefit because they're not in the LLM's training distribution.

Exclusions matter too

Some words look like brand names but aren't. The exclusions list catches them — for example, "Apple" as a brand should not match when the response is about apple as a fruit.

You'd add to exclusions when:

A specific phrase keeps producing false positives ("Get Vocal" as a generic instruction in copywriting articles)
A competitor name overlaps with a common word (the real competitor "Slate" vs the writing surface "slate")

What counts as a citation

A citation is a URL the LLM references when answering — either inline ([1] https://example.com), in a sources block at the bottom, or in Perplexity's structured citations payload.

We parse every URL out, normalize the hostname (e.g. www.example.com and example.com become the same), and store it linked to the mention's response. That gives you:

Top cited sources — the domains LLMs trust when answering questions in your category
Gap sources — the domains that cite your competitors but not you (the Tier 1 source recommender on /insights surfaces these as "publish here next")

Sentiment, position, and context type

The second-pass LLM extractor (BYOK only) classifies each mention along three axes:

Sentiment

Value	Meaning
`positive`	LLM described your brand favorably (best-in-class, recommended, market leader, etc.)
`neutral`	LLM listed your brand without endorsement (mentioned alongside others, factual statement)
`negative`	LLM described your brand unfavorably (criticized, listed as a cautionary example, deprecated)

Sentiment is per-mention, not per-response. If you're mentioned twice in one response — once positively and once neutrally — that's two separate sentiment data points.

Position

Value	Meaning
`first`	First brand mentioned in the response — usually the LLM's top recommendation
`later`	Mentioned but not first

First-position mentions matter disproportionately. LLMs tend to anchor on the first option they name; downstream user behavior follows.

Context type

Value	Meaning
`recommendation`	LLM is recommending your brand for a specific use case
`informational`	LLM is describing what your brand does, factually
`cautionary`	LLM is warning the user about something related to your brand
`dismissive`	LLM is excluding your brand from the relevant set

Context type is what makes share of voice numbers actionable — being mentioned 50% of the time as "the leader" is different from being mentioned 50% of the time as "outdated alternative."

What we don't classify

Intent of the user — we don't know if the human asking the question was a prospect or a competitor researching you. We only see the LLM's response.
Truthfulness of the LLM — if Claude tells someone your brand "supports SAML SSO" and you don't, we report the mention. We don't fact-check the LLM against your product reality. That's your job to act on.
Multi-turn behavior — we measure single-prompt responses. Whether your brand survives a follow-up question is a future feature (see roadmap for Conversation depth analysis).