Methodology and limitations

What is measured

Each cycle puts a fixed set of buying questions to several AI assistants and records which companies are named in the answers, in what position, and by how many of the assistants. A company that is never named does not appear.

This measures visibility in AI-generated answers. It is not a measure of quality, size, revenue or customer satisfaction, and it should not be read as one.

Verification

Language models routinely produce company names that do not exist. Every name returned by the measurement is checked against a live website that mentions the company before it can enter an index. Names that fail are held for human review rather than published or silently discarded.

In one measured run, 11 named companies could be confirmed and 516 could not. Verification is not a formality.

The instrument, per cycle

Scores move when companies change, and also when the instrument changes. Recording the instrument is what makes the difference visible, so every published cycle carries its full provenance.

A figure from one cycle should not be compared with one from another unless the prompt set, weight version and assistant set all match. Where they differ, the change in the instrument may account for the change in the score.

Interviews do not affect rank

The Gazette interviews companies it covers. Nothing said in an interview changes a company's position, and declining an interview carries no penalty of any kind. Interview data and ranking data are held separately and are never combined.

No suppression

The Gazette does not hold the ability to hide a score it has computed. There is no mechanism to remove a company from a published index on request, from the company or from anyone else. Errors of fact are handled through corrections.

Limitations

• Assistants are sampled, not surveyed. A fixed set of models is queried. Other assistants, and other versions of the same assistants, may answer differently.
• Model versions shift underneath the measurement. Providers update models without notice. Where a version changed between cycles, the comparison is marked rather than smoothed over.
• Absence is not evidence of absence. A company not named may simply not be described in terms the questions reach.
• Prompt wording changes results. This is why the prompt set version is published for every cycle.

Not yet settled

These are open and are published here rather than omitted.

• How the five component measures are weighted into the published composite. Until this is fixed, no composite is published.
• Whether the assistants queried should be the frontier models a buyer actually reaches, or the smaller models currently used.