Research and benchmarks

Build a prompt set

Use real language from sales calls, search queries, support tickets, reviews, communities, and customer interviews. Include multiple phrasings for the same intent and follow-up questions that narrow the recommendation.

For each prompt, store:

• journey stage and topic cluster
• target audience, geography, and constraints
• engine and model or experience tested
• brands mentioned and recommended
• cited domains and exact URLs
• claims and sentiment about your brand
• date, answer, and relevant follow-ups

Run a baseline

Test under consistent conditions and repeat enough times to see variation. Separate a brand mention from a recommendation and a recommendation from a citation. They are different signals.

Inventory your evidence

List every important company claim, then locate its strongest public proof. Flag claims that are vague, old, contradictory, unsupported, or trapped in PDFs and images. This often exposes the highest-leverage content work.

Keep a frozen baseline. Without it, later improvements become difficult to distinguish from normal answer volatility.