Authority Measurement

A dashboard is useful only when its numbers can change a real decision. We begin by naming that decision: whether to repair technical eligibility, strengthen a service page, invest in a market, improve source evidence, protect an established query family or stop work that is producing activity without qualified demand. The measurement plan then identifies the smallest set of observations capable of informing that decision. A ranking report, an AI answer screenshot and a sales enquiry may all matter, but they describe different stages and cannot be added together as if they were the same unit.

Every engagement receives a measurement dictionary. It defines the property, canonical route set, language, country, device, date window, brand exclusions, query families, prompt families, model or platform, sample conditions, conversion event and commercial qualification rule. It also records what the source cannot show. Without those definitions, two people can read the same chart and reach opposite conclusions; with them, the next observation can be compared with the first rather than reconstructed from memory.

The conventional search layer uses first-party platform data where access is authorised. In Google Search Console, clicks, impressions, click-through rate and average position are segmented by query, canonical page, country, device, date and relevant search appearance. Brand and non-brand demand are separated. A founder-name query at position three does not compensate for a service query at position eighty, and an average across six languages can conceal that one edition is gaining exposure while another is invisible.

Search Console is valuable, but it is not a complete record of all searches. Anonymised queries are omitted from tables, data can be truncated, chart and table totals can differ because of aggregation, and recent rows can still be preliminary. We retain the date range, filters, export method and aggregation choice with every baseline. When an API export is used, row limits and pagination are recorded. A change is never described as growth merely because two differently filtered reports produced different totals.

Google’s generative AI performance report can show impressions from supported generative Search features and group them by page, country, date or device. That is a useful first-party exposure signal. It is not a full prompt log, and it does not show every generative answer outside Google Search. We therefore keep it distinct from ordinary Search performance and from controlled observations made in other answer systems. If the report is unavailable because the property has insufficient impressions, the absence of a report is recorded as unavailable data—not as zero visibility.

A controlled AI-answer panel begins with fixed prompts that represent real buyer questions in each target language. For every observation we record the exact prompt, edition, platform, model or surface where visible, date, access state, response, brand mention, source link, recommendation state and portrayal accuracy. Repeated runs reveal variability; they do not create a census of what all users saw. A brand may be named without a citation, cited without endorsement, or recommended with no measurable visit. Those events remain separate so a favourable screenshot cannot inflate the entire report.

Visibility matters commercially when it contributes to an action by the right organisation. We inspect landing paths, referral sources, consented analytics events, enquiry forms, calls or messaging routes, CRM stages and the prospect’s own account of how they found the company where those records exist. OpenAI states that ChatGPT referral links carry a utm_source=chatgpt.com parameter, which can make some inbound visits identifiable. That still does not expose impressions inside ChatGPT, offline consideration, copied links, untracked devices or every assisted touchpoint.

Attribution is therefore expressed in levels. A directly observed referral and completed enquiry can be associated with its source under the agreed rules. A prospect who reports finding the company in an AI answer provides self-reported evidence. A rise in direct traffic after a publication is only a possible association unless stronger evidence exists. Qualified demand is defined in advance—such as market, need, authority, budget range or delivery fit—so a surge of irrelevant form submissions cannot be presented as business success.

The executive view begins with what changed, what did not and whether the evidence is strong enough to act. It separates verified implementation from platform outcomes: a page can be deployed and crawler-accessible without being indexed; indexed without gaining rank; visible in a sampled answer without being cited; cited without producing qualified demand. Each statement carries its source and comparison window. Where sample sizes are small, we show the raw count and avoid percentages that exaggerate movement.

Recommendations are tied to a diagnosed constraint and an acceptance test. The report may call for a canonical repair, a better commercial destination, a stronger public method, clearer author attribution, new first-party evidence, a native-language rewrite, a narrower prompt panel or no change at all. The next cycle does not begin with a fixed article quota. It begins with the unresolved decision that has the greatest credible commercial value and the lowest avoidable risk.