Methodology
The procedure, published in full
Every report we produce is a set of claims about what search engines and assistants currently say. Below is exactly how those claims are produced, so a client can hold a finished report against the method that made it and see whether the two agree.
Checked across
- ChatGPT
- Google AI Overviews
- Google AI Mode
- Perplexity
- Gemini
- Copilot
- Claude
Google AI Overviews and Google AI Mode surface inside search results — tracked as part of the search channel, not as separate assistants.
Why this is public
A report that says an assistant does not mention you is asking for trust. Published alongside the procedure that produced it, the same report becomes something you can check.
Every measurement here is made by a person. Nothing is generated automatically, and the procedure below is the one that person follows.
The questions
Each engagement gets its own set of questions, written in the words a buyer would actually use rather than the words the client uses about themselves. A question phrased in a firm's own vocabulary only tests whether that firm can find itself.
Four kinds go into every set:
- Your name, in quotes. This establishes whether an entity for you exists at all, and surfaces anyone whose name sits close enough to yours to be returned instead.
- Your name, unquoted. This shows whether the name gets corrected toward a better-established one.
- The category question, with nobody named. This is who gets recommended to a buyer who does not yet know you exist, and it is usually the commercially important answer.
- The problem question, in a buyer's own sentence. It frequently returns a completely different set of names from the category question, so both are asked.
The questions do not change
Once a set is written it is frozen. New questions get added with new numbers; existing ones are never edited.
This matters more than it sounds. Editing a question rewrites the history it was measuring, and there is no way to detect that afterwards. A year of one question is worth more than a year of several nearly identical ones.
Where they are asked
Both channels, in a fixed order, every run:
- Search results, read as a buyer sees them.
- Google AI Overviews and Google AI Mode, recording whether they appeared at all, since both are conditional.
- Each assistant in turn, one question per conversation.
Unnamed questions come first
The category and problem questions are always asked before any question containing your name.
Naming you changes every answer that follows it in the same session. Asked in the other order, the results would describe a conversation we had rather than the position you hold.
How each session is controlled
An assistant answers a logged-in account differently from a stranger. A prospect looking you up is a stranger, so the measurement is taken as one.
Every session runs signed out, with stored personalisation switched off, one question per conversation, and the country recorded. Answers move between countries and between weeks on their own, so the date, time and locale go into the record alongside the answer.
What gets written down
For every question, on every surface: the answer word for word, every party it names, and every source it cites.
Answers are kept verbatim rather than summarised, because a change in wording is the earliest sign that something has moved, and a summary destroys it.
Absence is written down as a result. A surface that returned nothing about you is recorded as returning nothing, never left blank, because a blank cannot be told apart from a question nobody asked.
What we hold ourselves to
These rules decide whether a report is worth paying for:
- A surface we did not open is reported as not measured. It is never reported as finding nothing.
- Where our own measurement turns out to be at fault, the mistake stays in the report with the correction beside it.
- We separate what we could not see from what was not there. Treating the first as the second is the likeliest way to be confidently wrong about a client.
- Every figure in a report was read off a screen. We publish no percentage, score or ranking that we did not observe.
- An answer we quote is quoted exactly, including where it is unflattering to us.
How often
A baseline is taken once, before anything is changed. It cannot be reconstructed later, which is why it comes first.
After that the same questions run weekly, and the predictions made in the baseline are scored at four weeks and at twelve. Being wrong on our own record, in writing, is what makes the next report worth reading.
The named method these runs belong to is on the method page.