Methodology · Published in full

How we measure. Exactly.

Most vendors sell a proprietary "visibility score" you can't inspect. We publish our method instead — because a measurement you can't audit isn't a measurement, it's marketing. This page is the whole recipe.

The protocol

75 prompts × 5 engines × 3 runs = 1,125 observations

Step	What happens	Why it's done this way
01	Prompt construction. 75 buyer-intent prompts built from your sales calls, objection logs, and won/lost notes — phrased the way buyers actually ask, not the way keyword tools suggest.	AI queries are conversational. A prompt set derived from keyword tools measures a world that no longer exists.
02	Multi-engine coverage. Every prompt runs on ChatGPT, Google AI Overviews / AI Mode, Perplexity, Claude, and Copilot.	Engine answer sets disagree with each other. A single-engine score is a partial truth sold as a whole one.
03	Three runs per prompt per engine, spaced across the measurement window.	Engines are non-deterministic — the same prompt returns different answers on different runs. Single-run scores are noise. Three runs give a stable presence rate and an honest range.
04	Structured logging. Every observation is recorded: timestamp, engine, prompt, full answer text, brands named, sources linked, sentiment of your mention.	The raw log ships with your report. Anyone on your team can re-check any entry.
05	Identical monthly re-run. Same prompts, same engines, same protocol, every month of the program.	Changing the ruler every month makes progress unmeasurable. Holding the method constant is what turns a snapshot into a scoreboard.

The metrics

Four numbers, defined before we measure them.

Metric 1

Citation share

The percentage of observations where an engine links your domain as a source. The strictest signal — and currently the rarest, since only a small share of AI answers carry visible links at all. Reported separately, never blended.

Metric 2

Mention share

The percentage of observations where an engine names your brand in the answer, linked or not. This is most of your real exposure, and the number most vendors either ignore or quietly merge into an inflated "visibility score."

Metric 3

Mention sentiment

How you're characterized when named: recommended, listed neutrally, or flagged with caveats. Being cited as the expensive option, or the one with mixed reviews, is measurement other firms skip because it complicates the sales story.

Metric 4

Consensus set & source map

The 3–5 brands engines default to when you're absent, and the third-party sources those answers draw from. This is the strategic output: it tells you exactly whose position you're contesting and which doors the engines already trust.

From measurement to movement

What the remediation work actually is.

No secret tactics — engines reward the same things they say they reward. The craft is sequencing and execution.

Layer	The work
Technical	Crawler access for GPTBot, ClaudeBot, PerplexityBot, Google-Extended; entity-clear Organization and Service schema; consistent NAP; llms.txt where supported.
Content	Answer-extractable structure: direct claims up front, verifiable specifics, comparison and pricing pages engines can actually quote — replacing the throat-clearing they skip.
Authority	Placement in the sources your category's answers draw from: review platforms, comparison articles, communities, trade press. The slowest layer and the one that moves consensus sets.
Entity	Making your brand unambiguous across the open web — profiles, references, and third-party corroboration that let a model resolve you as one coherent thing.

Roughly 80% of this is fundamental SEO executed properly — a proportion we state upfront because vendors who don't are selling the same work twice.

Limits of the method

What this methodology cannot do

Printed in every report we deliver, and here, in public:

§1It samples; it doesn't census. 1,125 observations is a serious sample of an effectively infinite query space. Your true presence across every phrasing a buyer might use is unknowable — by us or anyone.
§2It observes outputs, not internals. No vendor has access to how these models rank or retrieve. Our inferences are disciplined, replicated, and still inferences.
§3It can't fully attribute revenue. Most AI-driven visits carry no referrer data. We instrument what's measurable and report the confidence interval, not a fantasy ROI figure.
§4It's subject to model updates. A retrain or product change can move every number in the industry overnight. Monthly re-measurement is the defense — it catches the shift and separates it from your trend.