How we measure who gets named

The prompt battery

The battery is five questions. Five is enough to cover a firm's main capabilities and few enough that every answer gets read by a person, screenshotted and explained. The rules for writing them:

1. Real buyer questions. Each prompt is something a buyer would type, taken from the firm's quote history, its engineers' description of the work, and the questions the firm's best customers asked before they bought. Never a brand-name search, never "best engineering firm UK".
2. UK phrasing. British spelling, UK product and regulatory vocabulary, and a UK location cue where a buyer would give one. An assistant answers a UK question with UK sources more often than not, and that is the market being measured.
3. Capability, not category. "Who repairs obsolete variable speed drives in the UK" rather than "drive repair company". The specific question is the one that separates a firm from a directory.
4. Fixed per firm. Once written and agreed at the baseline, the five prompts do not change. A changed prompt is a new baseline, and the report says so. This is what makes month six comparable with month one.
5. Written before the baseline, not after. The prompts are agreed with the client before the first run, so nobody can pick the questions the firm already wins.

Assistants and search surfaces

ChatGPT

The default model with web search available, logged out or on a clean account, UK location. The answer is read as a buyer would read it, and the sources it cites are recorded.

Perplexity

Default mode, UK location. Perplexity cites sources inline, which makes it the clearest view of which pages are doing the work behind an answer.

Google AI answers

AI Overviews where one appears on the query, and AI Mode. Both draw on the ordinary Google results, so those results are recorded alongside, top ten, UK location.

Microsoft Copilot

On request. Its answers come from the Bing index, which matters for procurement teams on corporate desktops, but it is not in the standard battery.

Not measured

Voice assistants, in-app assistants, and any tool that requires a paid account the buyer is unlikely to have.

What counts

Mention

The firm's name appears in the answer text. A mention in a list of six is a mention; the position is recorded too.

Citation

A page on the firm's own domain is linked or footnoted as a source for the answer, whether or not the firm is named in the text. Citations are recorded by URL.

Named first

The firm is the first named in the answer. Recorded separately because a buyer reading a shortlist often stops there.

Absent

Neither mentioned nor cited. The report also records who was named instead, with the source each one was drawn from.

Corroboration

Whether the sources behind an answer are the named firms' own pages, or third parties: directories, trade press, supplier registers, forums. Answers built only on self-description are noted as such, because they are the ones most easily displaced.

Screenshot and date policy

• Every answer is screenshotted in full, with the assistant's interface, the prompt and the date visible in the capture, and saved with the run date in the filename.
• Runs happen on a fixed working day each month, the same day for every prompt in the battery. A run interrupted by an outage is repeated the next working day and the report says so.
• Each prompt is run once per assistant per run. We do not re-run until a better answer appears.
• Screenshots are kept for the life of the engagement and handed over when it ends. The client can request the raw set at any time.
• Sample runs published on this site use the same rules, with firms described rather than named.

What cannot be attributed

AI answers vary by account, session, location and day. A run is a sample, and the report treats it as one: five prompts on one date is evidence of a pattern, not proof of what any single buyer saw.

Where an enquiry arrives, the report records how it was tracked, by a form field, a tracked number, or what the buyer said, and does not claim an AI answer caused it unless one of those routes says so.

Movement in the numbers after work is done is reported as movement, alongside the work, without asserting that one caused the other. Over several months the pattern usually becomes clear, and the report waits for it.

What we do not claim to measure

• What any individual buyer was shown.
• Why an assistant chose one source over another. We can show what the cited sources have in common; we cannot see inside the model.
• Share of voice across every possible question. Five prompts is a battery, not a census.
• Enquiries the buyer never told us about.
• Anything about a competitor beyond what is public. Every claim links to the page it came from.

How the monthly report reads

One page per prompt: the prompt, the three answers side by side as screenshots, and the mention and citation status for the client and for each firm named instead.

Under that, the sources behind the answer, whether each one is self-description or third party, and one line on what changed since last month.

A front page carries the five-prompt totals against the baseline, the enquiries that arrived and how each was tracked, and the work done that month. A closing page lists what the method could not attribute.

It is written to be forwarded to a technical director without a covering explanation.

Sample

Three sample answers from a public run on 9 September 2026 are on the homepage, with the firms described rather than named.

A full sample snapshot is published here: the redacted snapshot for a UK low voltage switchgear refurbishment and maintenance firm (PDF, eleven pages, run on 10 September 2026).

It carries the five prompts, all fifty results they returned, the record of who was named and who was cited on each one, and the closing page on what the method could not attribute.

The competing firms are described rather than named, which is the only difference between it and a client's own copy.

Changes to this method

This page is dated. If the battery rules, the assistants or the definitions change, the date changes and the old version is kept, because a client's baseline was made under the version that was current on the day.