This is the canonical specification behind every visibility number rawmktg publishes, in articles, teardowns and the free tools. It is versioned. When it changes, the version and the changelog at the foot of this page change with it, and any figure that depends on the change is restamped. If a teardown and this page ever disagree, this page is correct.
Most AI-visibility numbers you will see anywhere, including screenshots in sales decks, come from a single query typed once into one engine. That is not a measurement, it is an anecdote with a timestamp. AI answers are non-deterministic: ask the same question ten times and the set of brands named changes, sometimes by fifteen points, with nothing altered on any website. A method that ignores that variance reports noise as fact. This page defines the method that does not.
01What has to be in a prompt portfolio?
Every prompt is a real question a buyer in the category would ask an assistant, phrased the way a person phrases it, not a keyword. Prompts are split into three buckets, commercial (best, top, alternatives), research (how, what, comparison), and brand (named-entity checks). Each bucket carries a weight, because an appearance in a high-intent commercial answer is worth more than one in a definitional aside. The portfolio is stored with a version tag such as v2026.08 and frozen for the life of that version, so every cycle measures the same instrument.
| Tier | Prompts | What it answers | Confidence |
|---|---|---|---|
| Baseline / diagnostic | 50 to 150 | A fast read: are we present at all, and where are the obvious holes | Directional, not for board slides |
| Decision-grade / standard | 250 to 500 | The default. Share and trend you can act on and defend | Roughly a plus or minus 2-point margin |
| Enterprise / category | 500+ | Full category coverage, per-segment and per-region breakouts | Tight intervals on sub-segments |
Different rawmktg articles run different tiers on purpose. The prompt-to-citation baseline starts at 50 to 150 because it is a 30-day diagnostic; the Share of Model spec runs 250 to 500 because it is decision-grade. Those are not contradictions, they are named tiers of the same method. What never changes between them is the run count.
02How many times is each prompt run?
Ten runs is the working default. Below eight, cycle-to-cycle swings are dominated by sampling noise rather than real change, which is how a brand appears to gain or lose fifteen points in a month with no work done. This is the single parameter that must be identical everywhere, and it is the one earlier drafts were loosest about. A baseline diagnostic may use a narrow portfolio, but it still runs each prompt the full eight-to-twelve times, otherwise it is not a baseline, it is a guess.
If you take one number from this page, take this: 8 to 12 runs per prompt per engine, default 10, at every tier and in every article. A wide portfolio run once is worse than a narrow portfolio run ten times.
03Which engines count, and how are they weighted?
The engines diverge enough that optimising for one does little for another, which is why cross-engine source overlap is low and why the report keeps them apart. Weights reflect reach and how much each engine actually influences a buyer, and they live in the versioned config so a weight change is a version change. The composite is only ever presented alongside the per-engine rows it is built from.
04How is a number bounded and guarded?
SoM(b) = Σ_e Σ_p w_e · w_p · score(b,e,p) ÷ Σ over every brand in the field b brand score 0.30 mention + 0.30 recommendation e engine + 0.20 position + 0.15 sentiment + 0.05 prominence p prompt w_e, w_p engine and prompt-bucket weights Share of Model is a share of the whole field, not a raw inclusion rate.
n = z^2 · p(1-p) / e^2 # observations per brand per engine z = 1.96 (95%) p = expected inclusion rate e = target margin For p = 0.30 and e = 0.02, n ~ 2,000 observations. At 10 runs per prompt that is ~200 prompts per engine cell.
The sample maths is why decision-grade lands at 250 to 500 prompts: at a 30% base rate and a plus or minus 2-point target, you need roughly 2,000 observations per engine, which at ten runs is about 200 prompts per engine cell. You can size any target yourself with the sample-size and confidence planner. The 200-observation floor is enforced in the query, not left to judgement.
-- No cell is reported until it clears the sample floor. SELECT brand_id, engine, AVG(present) AS inclusion FROM observations WHERE portfolio_version = 'v2026.08' GROUP BY brand_id, engine HAVING COUNT(*) >= 200; -- under-sampled cells are withheld, not shown
05How often is it re-run, and what makes a figure decision-grade?
The distinction is a labelling rule, not a soft preference. Directional numbers, from a baseline portfolio, are allowed and useful, but they are never dressed up as decision-grade, and they never appear without the word. Declared inputs that cannot be measured, whether a brand publishes original research, has named expert quotes, runs an aged and authentic community account, are kept in a visibly separate declared section with an asserted, not measured label, and are never folded silently into a measured composite.
06What changes, and how do you know it changed?
| Version | Date | Change |
|---|---|---|
| v1.0 | 26 Aug 2026 | Initial published standard. Fixes the run count at 8 to 12 (default 10) across all tiers and articles; defines the baseline / decision-grade / enterprise portfolio tiers; sets the 200-observation per-cell reporting floor. |
Frequently asked questions
How many prompts should an AI-visibility measurement use?
It depends on what the number is for, and rawmktg uses three named tiers. A baseline diagnostic uses 50 to 150 prompts and is directional. A decision-grade programme, the default, uses 250 to 500 prompts, which at ten runs per prompt gives roughly a plus or minus 2-point margin. Enterprise or full-category work uses 500 or more. The tier sets how wide the question set is; it does not change the run count, which is fixed.
How many times should you run each prompt?
Eight to twelve times per prompt per engine, with ten as the default, at every tier. AI answers are non-deterministic, so a single response is a coin flip. Averaging 8 to 12 runs is what converts anecdotes into an inclusion rate with a usable confidence interval. Below eight runs, month-to-month swings are dominated by sampling noise rather than real change.
Why not just average all the engines into one score?
Because ChatGPT Search, Google AI Overviews, Perplexity, Copilot and Gemini cite different sources for the same question, so a blended number hides where you are winning and losing. rawmktg measures and reports each engine separately, then combines them with evidence-based weights held in a versioned config, and always shows the per-engine breakdown behind any composite.
What makes a number decision-grade rather than directional?
Three things together: a portfolio of at least 250 prompts, 8 to 12 runs per prompt per engine, and every reported brand-engine cell clearing 200 observations. If any of the three is missing, the figure is published as directional and labelled that way. Declared inputs that cannot be measured are kept in a separate section and never folded into a measured score.
Why is this methodology versioned?
So it cannot be edited invisibly. A method that can change without a record is not a standard. Every change to a weight, threshold, run default or portfolio version is logged in the changelog on this page with its effective date, and any published figure that depends on the change is restamped.
Read the full spec and worked example in Share of Model, measured properly. Size your own sample with the sample-size and confidence planner. See the baseline version in prompt-to-citation tracking, and the metric taxonomy in citation vs mention vs recommendation.
rawmktg. publishes data-driven teardowns and technical playbooks on GEO, agentic commerce and B2B AI-search visibility. Method: same data, same lens, every time. Contact: vinayak@rawmktg.com