How to read the report

Protocol v1 · 30 September 2026

There is no validated public 0–100 visibility score. The sample is synthetic and demonstrates counts, not live performance. Our real HTTP demonstration records a website observation and a correction to the checker. No real assistant recommendation pilot or customer result has been published.

Record the conditions

Each delivered report must include the exact 20 prompts, question IDs, branded or unbranded grouping, location and service area, baseline date and UTC timestamps. Record the service, model/version when available, API or consumer interface, search mode, account/personalization conditions and repeat count. Use independent sessions where supported; disclose any condition that cannot be controlled. One baseline repetition is included; it does not estimate variability.

API outputs do not stand in for ChatGPT or Gemini consumer search. Google AI Overviews must be recorded separately from Google searches: a search without an Overview is “not triggered,” not a recommendation failure. Failures and unavailable services are “unavailable,” not zero mentions.

Counts anyone can check

A mention is an explicit, correctly matched business name in an eligible answer. A recommendation is a positive suggestion to hire or contact that business; mere mention, criticism and mistaken identity do not qualify. Count a business at most once per response. Keep rank, sentiment, matching ambiguity and cited URLs as separate fields, with the complete answer behind each annotation. A human checks these fields against owner-verified facts.

For each service and prompt group, report recommended responses / eligible responses. Mention share is a business’s response-level mentions divided by all tracked businesses’ response-level mentions in that same group; label the competitor set and exclude unavailable/no-Overview responses. It is not market share. Branded and unbranded results are never combined for a discovery claim. The sample’s discovery count is calculated directly from its downloadable records.

Corrections and comparisons

Every correction separates observed facts, sources, a possible consequence, action, confidence, effort and a verification plan. Priority reflects factual severity, usefulness and effort. Improved recommendation rates are a hypothesis until a comparable dated retest exists; even then a change is observational rather than proof of causation. Prompt, model, interface or search-mode changes break like-for-like trends and must be flagged.

The current operator probe tool uses three API intents and a model judge. Its composite score is experimental and cannot substantiate a 20-question, five-service report. The protocol above requires a separate human-reviewed evidence packet before delivery.

Google says special AI files are not needed for its generative Search features. Our optional business export is for maintaining consistent facts, with no recommendation benefit promised. Google Search guidance.