Prompt Intents

Two prompts can look similar and measure completely different things:

• “best cordless vacuum for pet hair” — you either come up or you don’t. A mention is earned.
• “Dyson vs Dreame for pet hair” — you come up because the question says so. A mention proves nothing.

Scoring both with one rubric would make the second kind free points. So every prompt carries an intent, and the intent decides the rubric, the bucket, and which averages the probe counts toward.

The four intents

Intent	Names your brand?	Example	Bucket
Category	No	”best massage gun 2026”	Discovery
Scenario	No	”quietest massage gun for a 10-person startup”	Discovery
Comparison	Yes	”Dyson vs Dreame hair dryer”	Branded
Trust	Yes	”is Kospet legit in 2026”	Branded

Discovery — category and scenario

Neither names any brand. These measure unprompted visibility: out of everything the assistant could have recommended, did it reach for you?

• Category prompts are the head terms — broad, high-volume, hardest to win, usually dominated by incumbents.
• Scenario prompts are the long tail — a persona, a use case, a budget, a constraint. This is where niche brands surface most, and where the buyer is closest to a decision.

Winning one and not the other is a common and actionable pattern, which is why the panel splits them out rather than reporting “discovery” as a lump.

Branded — comparison and trust

Both name your brand in the prompt text. These measure how you’re judged once you’re already in the conversation.

• Comparison puts you against a named rival: who does the assistant pick? Reported as a win rate over win / tie / lose verdicts.
• Trust asks whether you’re legitimate — “is X legit”, “does X honor its warranty”. Reported as a reputation verdict over legit / mixed / scam. There is no win rate here: it isn’t a ranking question.

What the bucket changes

	Discovery	Branded
Position scored	Yes (0–85)	No — set to 0
Frequency bonus	Yes (0–8)	No — set to 0
Sentiment scored	Yes (0–15)	Yes (0–15)
AI Score from	Position + sentiment + frequency	Verdict band
Counts in Mention Rate	Yes	No — reported separately
Counts in Share of Voice	Yes	No
Counts in AI Score average	Yes	Yes

Why Mention Rate excludes branded prompts

A comparison prompt guarantees a mention. If branded prompts counted, adding more of them would raise Mention Rate without anything changing in the market — the metric would measure your prompt library instead of your visibility. They are reported underneath as a separate x/y line so the information isn’t lost.

How prompts get an intent

Generated prompts

Prompts → Generate with AI asks for a count per intent and produces exactly that many. The generator is told your brand and industry, and works under hard rules:

• Category and scenario questions must not name any brand.
• Comparison and trust questions must name your brand; competitors may only appear alongside it.
• It may never produce a ranking of competitor brands that omits you.
• Questions must be geography-neutral — no “in the US”, no “in Germany”. The region is applied at run time as a retrieval parameter, so a country inside the prompt text would measure the same market twice and break cross-region comparison.

Because the counts are requested per intent and returned in labelled sections, a generated prompt’s intent is known by construction rather than guessed afterwards.

Imported and manually written prompts

Anything you paste or type is classified automatically. Classification runs on Gemini 2.5 Flash and never blocks the import: if the model is unavailable or returns the wrong number of labels, each prompt falls back to a deterministic rule instead.

The fallback rules:

1. Does the text name the target brand? Matching uses the brand name plus aliases derived from your website domain. If it does not name the brand, the prompt can only be category or scenario — the branded intents are defined by naming the brand.
2. If it doesn't: scenario or category? Markers like "for", "under $", "budget", "team", "beginner", "use case", or a "best … for …" shape indicate a scenario. Everything else is a category head term.
3. If it does: comparison or trust? Comparison markers — "vs", "versus", "compare", "alternative" — mean comparison. Everything else that names the brand is trust, which is the better home for "is X good for Y" phrasing.

The brand-consistency correction

After classification, one correction is applied in one direction only: if a prompt names your brand but was labelled category or scenario, it is moved to a branded intent.

That direction is safe and necessary — a prompt containing your brand name cannot measure unprompted discovery, and leaving it in the discovery bucket inflates Mention Rate.

The reverse correction is deliberately not applied. Customers frequently write real comparisons using product model numbers rather than the brand name (“MW06 vs Baby Brezza”), and product names have no reliable source to check against. Demoting those to discovery would manufacture the exact error the correction exists to prevent.

Choosing a mix

There is no enforced ratio, but a library that answers useful questions usually looks roughly like:

Bucket	Share	What it tells you
Category	~25%	Whether you exist in the head terms at all
Scenario	~35%	Where you can realistically win today
Comparison	~25%	Whether you survive a direct matchup
Trust	~15%	Whether reputation is a blocker before purchase

Ten prompts is enough to see a signal; a few dozen makes the per-engine and per-region splits meaningful.

AI Score → The two rubrics in full detail. Prompts guide → Building and running the library in the app.