Methodology

How Pickentra measures AI recommendations

What we ask, what we count, and what we do not claim, so you can judge the evidence yourself.

Signals

Three separate signals

Mentioned

Your business is named. A mention is not a recommendation.

Verified citation

Your verified domain is cited as a source. A citation is not a recommendation.

Method

The details

Approved neutral questions

Eight neutral, buyer-intent questions about your service and city, approved by you before any run. They do not name your business.

Frozen versions

Approved questions and the business profile are frozen as immutable versions. Changes require a new draft and approval; profile changes invalidate active approval. Old evidence keeps its original identity. Different versions are not automatically comparable.

Weekly cadence

The same questions are asked on a weekly cadence once an operator enables live runs. Live AI runs are off by default.

Official APIs and web search

The only live-capable provider is OpenAI, using the Responses API with its web_search tool. API answers can differ from what a person sees in a consumer AI app, and from other providers.

Retained evidence

Each answer is stored with its provider, model and UTC date. Every result in a report links back to the full answer text, ranked list and citations.

Conservative ranking extraction

Ranks are taken only from clearly ordered lists. Each result is yes, no, uncertain or unavailable, with a supported, uncertain or unavailable evidence label. These labels are not calibrated probabilities. Counts show their denominators; absent readable evidence is unavailable, not an absent business.

Domain citation matching

A verified citation is an inline answer citation matching the domain in your frozen approved profile. Directories, competitor sites and search metadata do not count.

Failures

Failed, missing or uncertain answers are never counted as no or as zero. If the week cannot be completed, the score is Unavailable.

Comparability

Week-to-week changes are shown only when the earlier report used comparable questions and method; otherwise the reason is shown.

Limits

What the method cannot tell you

Results cover one AI provider and eight questions per week, not every search or assistant
Other providers are counted by how many answers named them, not by market share; names come from AI text and may not match a verified business
Suggested next steps are based on recorded answers and do not guarantee rankings, leads or sales

See the exact formula on the score page.