A Methodology Note Template for AI Visibility Reports
Attach a short methodology note to every AI visibility report and fill in the same ten fields each time: the prompt set and its version, the assistants and the exact surface sampled, runs per prompt, the sampling dates, the counting rule for a mention and a citation, the denominator behind every percentage, what was excluded and why, any first-party data used with its own definition, changes since the last report, and what was not measured. The note is what lets a reader re-derive a number, or decide not to trust it. A report without one asks to be believed. A report with one asks to be checked, which is the only kind worth sending.
Scientific papers have a methods section because the result means nothing without the procedure that produced it. That applies with more force here than in most marketing reports: answers vary between runs, differ between assistants and shift when a vendor ships a new model. The procedure carries a large share of the meaning.
The ten fields
| Field | Example (illustrative) | The misreading it prevents |
|---|---|---|
| Prompt set and version | v3, 25 prompts, none naming the brand | Comparing periods measured with different questions |
| Assistants and surface | Four assistants, consumer apps, web search on, signed out | Treating an API result as what app users see |
| Runs per prompt | 3 per assistant per window | Reading one lucky run as presence |
| Sampling dates | 8-10 September 2026, IST | Blending results from either side of a model change |
| Counting rule | Mention: brand named in the answer text. Citation: our URL shown as a source | Counting a competitor's page that names you as your citation |
| Denominator | 25 x 4 x 3 = 300 answers | A percentage with no 'of what' |
| Exclusions | 7 runs returned errors; kept in the denominator as no mention | Quietly shrinking the base to flatter the rate |
| First-party sources | Search Console generative AI report, impressions, 1-30 September | Adding impressions to sampled mentions |
| Changes since last report | Two prompts retired; set moved from v2 to v3 | An invisible break in the trend line |
| Not measured | Voice modes; signed-in, personalised sessions | Silence read as zero |
The template to paste under every report
Methodology
Prompt set: v__ (__ prompts; any naming our brand: yes/no)
Assistants and surfaces: ______ (app/web/API; web search on/off; signed in/out; location ___)
Runs: __ per prompt per assistant, sampled __ to __ (timezone __)
Counting rule: mention = ______; citation = ______
Denominator: __ prompts x __ assistants x __ runs = __ answers
Exclusions: ______ (kept in denominator as ___ / removed because ___)
First-party data: ______ (source, metric, date range, status e.g. preview)
Known platform changes in the window: ______
Changes since last report: ______
Not measured: ______
Raw rows stored at: ______The field people leave blank
Not measured. A report that lists four assistants and says nothing about the rest gets read as covering 'AI search'. One line fixes that, and it also carries the distinction that matters most in this channel: an assistant you didn't sample is unknown, not zero. Readers merge those two states unless you stop them.
Exclusions deserve their own line
Every run that errored, timed out or returned a refusal is a decision: keep it in the denominator as a non-mention, or remove it and say why. Either can be defended. Doing it silently can't, because removing runs almost always raises the rate. Our share-of-voice guide calls this denominator inflation; the exclusions line is where you show you didn't do it.
Keep it short enough to be read
Filled in, the note should fit in a short paragraph. Detail belongs in the stored rows it points to. If the note grows past a screen, it usually means the method changed during the period, and the honest fix is to split the report at the change rather than explain the blend.
FAQ
Does an internal report need this? — Yes, arguably more than a client report. Internal numbers get repeated in meetings without their context, and a note in the same file is the only context that travels with them.
What if the method changed mid-period? — Report the two halves separately and name the date of the break. One number spanning a method change describes neither method.
Should the note list every prompt? — Give the version and count in the note and the full list to anyone who asks. Readers who've seen the prompts trust the counts more; pasting all of them into every report makes the report too long to read.
Put this to work on your own website.
Fenn finds what your customers ask, drafts the articles and site fixes, and measures what ChatGPT, Claude, Gemini, Perplexity and Grok say about you — with every change waiting for your approval.