How we measure, and what we refuse to claim.

1. Which questions we ask

Tracking runs against a question set, not a keyword list. Questions are the buyer-intent prompts a real customer would type into an assistant, and they are chosen by the account holder — we do not silently add or substitute them.

Branded prompts are tagged and excluded from headline figures. A question containing the client's own brand or workspace name almost always returns that brand, so counting it inflates the number an agency puts in front of a client. Those questions are still measured, still stored and still shown in the grid — they simply never set a headline percentage. Every report states how many branded answers it excluded.

2. Which assistants we ask, and how often

Every paid plan tracks ChatGPT, Google AI Overviews, and Perplexity. Claude $29, Gemini $9, Google AI Mode $9, and Microsoft Copilot $9 are available as per-client add-ons, priced per client per month because that is how they cost us to run. Pro and Agency include Microsoft Copilot as a core engine; on Solo it is a $9 per-client add-on.

Cadence. Tracking runs on an automatic daily scheduler. A full sweep of a question pool completes as often as the plan’s monthly engine-check allowance allows — a lightly-used pool cycles sooner than a full one — and every account shows its own live estimate and next scheduled window in-app. We state this precisely because "checked daily" and "every question refreshed daily" are not the same claim: the scheduler runs every day, and how much of a question pool it can sweep in that day depends on the plan's monthly engine-check allowance and the size of the pool. Each account shows its own live estimate and next scheduled window in the app.

3. What we record from each answer

Each check stores the assistant, the question, a timestamp, the answer, and two separate signals. They are separate because they mean different things, and conflating them is the most common way an AI-visibility number becomes untrue.

• Presence — the assistant named the brand affirmatively in its answer. This is what a customer actually reads.
• Source appearance — the brand's domain appeared in the source list behind the answer, whether or not the answer named it. Where the check also recorded a rank, we report the position within that specific answer's source list. It is not a position in search results.
• Citation — an attributable source link. We use this word only where the underlying record proves it, which is why our reports say "named in the answer" rather than "cited" for a presence signal.

A brand can be read as a source and still not be named. That case is reported separately and never counted as presence.

4. How the numbers are calculated

• Presence rate = unbranded answers naming the brand ÷ unbranded answers in the latest run.
• Per-assistant rate uses the same basis, restricted to that assistant, so an engine card and the headline can always be reconciled.
• Read-as-a-source rate = unbranded answers where the domain appeared in the sources ÷ unbranded answers that recorded a source list.
• Share of voice = each brand's named answers ÷ tracked answers in the same period, across the rival set configured for that client.

Denominators are declared, never implied. Every report states how many answers its percentage covers and how many were excluded. Where a plan runs multiple samples per assistant, the verdict for a question is the majority of those samples, and the report says how many were unanimous.

5. What we exclude, and why

• Sampled and failed checks. Only live results reach a client-facing report. A run that errored is dropped rather than counted as an absence, so a provider outage can never look like a lost mention.
• Branded prompts, from headline figures only — see §1.
• Competitor rows, from the client's own presence rate. They feed share of voice instead.
• Where a client's only live answers are branded, we report that no unbranded basis exists rather than printing a 0%.

6. Limits we will state plainly

Assistants are stochastic: the same question can return different answers minutes apart, which is why verdicts are sampled and why a trend matters more than any single check. Answers vary by phrasing, by locale and over time. A measurement is evidence that an answer occurred on a date — it is not a guarantee that it will recur, and no tool can promise an assistant will cite anyone. Where we publish our own baseline, it is dated and kept as a baseline rather than rewritten when it improves.

7. Change log

Material changes to this methodology are recorded here rather than applied silently.

25 August 2026	First published. Presence, source appearance and citation defined as three distinct claims; branded-prompt exclusion and denominator declaration documented; cadence stated as a daily scheduler with a budget-paced sweep rather than a fixed per-question interval.

Related: product documentation · sourced answers on GEO and AI visibility · run this method on your own domain.