How we assess AI visibility.

We use realistic customer questions to examine how a business appears in AI answers. We save the underlying observations and distinguish what happened from our hypothesis about why it happened.

Initial assessment

The free call uses a preliminary review of public information and selected AI answers. It is a starting point for discussion, not a comprehensive audit or a stable visibility score.

Scoped audits and research panels

For deeper work, we agree the buyer situations, platforms, geography, and repeat checks. We record the prompts, answers, dates, visible sources, and how the results were classified.

Platform differences

Consumer interfaces and APIs can produce different answers. We state which surface was tested and do not present an API result as a consumer-interface observation.

Limits

AI answers vary. A cited source does not prove why a brand was recommended, and improvement after a change does not by itself prove that the change caused it. We track repeated observations and explain other plausible factors.

Independent brand research

Our published teardowns state what we would investigate or change if hired. They do not imply a client relationship or an achieved result.

Published research panels

A panel generally contains four to five unbranded buyer questions. Each question is repeated five times in fresh ChatGPT Temporary Chats, with web search enabled and personalisation off. Articles state their question groups, observation dates, geography and valid run counts.

We record brand inclusion, recommendation role, first appearance where the format permits comparison, and visible source domains. Interrupted runs are replaced. Counts describe the tested questions and period; they do not establish why a recommendation happened.

Historical public API snapshot — retired

This describes the former automated tool for interpreting older records. It is no longer offered as a public website audit.

The free audit researches a brand with cited web search, then proposes six unbranded discovery questions and two branded validation questions. You can select questions and add your own, up to ten in total.

Each question receives one independent OpenAI API request. The configured model must complete a web search and return an answer with usable citation URLs. The report identifies the model and observation time, includes the answer, and links to its cited sources. This API result may differ from the consumer ChatGPT interface.

Searches have a deadline. Failed, incomplete or uncited responses are labelled as failed checks. They never count as evidence that a brand is absent. Failed questions can be retried without rerunning the verified answers. An incomplete panel receives no overall score and cannot be submitted as a complete email report.

The historical discovery score

An explicitly recommended brand receives two points for that discovery answer; a mention without a recommendation receives one; absence receives zero. The discovery score is the points earned divided by twice the number of valid discovery answers, multiplied by 100 and rounded to the nearest whole number.

Branded questions appear separately and contribute no discovery points. A panel containing only branded questions has no discovery score. Classification uses a second model reading the observed answer, so inspect the answer when a classification looks questionable.

Corrections and replies

If a research article names your brand and you have a correction or useful context, email [email protected]. See our prompt-panel guide for how questions are selected and interpreted.