CiteSprout

1. Read the site. We fetch your homepage and claude-sonnet-4-5 infers the brand name, product category, and three real competitors. That inference is shown above the score so you can see if it got it wrong.
2. Build a prompt matrix. 15-20 buyer questions in three categories from GEO-bench (Aggarwal et al., KDD 2024): informational (broad research), transactional (budget, team size, compliance), navigational (names a specific product). Six are hand-written templates; the rest are generated from them by claude-sonnet-4-5.
3. Validate every question. Each prompt is re-classified by sentence embeddings (all-MiniLM-L6-v2) against labelled examples. The generator's own label is never trusted. Anything ambiguous is rejected rather than counted in the wrong bucket, which is why roughly half get discarded.
4. Ask the models for real. Every question goes to claude-sonnet-4-5 and gpt-4o with web search switched on, at default settings. Plain calls without search return no citations and invent tool names, so grounding is required, not optional.
5. Read the answers. We record whether your brand was named, whether the tone was positive, neutral, or negative (twitter-roberta-base-sentiment), which competitors appeared, and every URL the model cited. The raw response is stored before parsing, so a parse failure never loses it.
6. Score it. A counted mention is your brand named favourably or neutrally. Each category's rate is the share of its answers that counted, weighted 80/10/10 per GEO-bench. A category with fewer than 3 successful answers reports "insufficient data" instead of a number, and failed calls are excluded entirely — a provider error is missing evidence, not a gap.

Every number is one click from the transcript that produced it. If a figure here cannot be traced to a stored answer, treat it as a bug.