Tracking Brand Visibility on Claude

Question (from a frozen set), timestamp, region, language, and device or client type. Any of these can change the answer, so all belong in the record.

Keep the full answer text for every sample. Storing only 'mentioned / not mentioned' makes re-auditing impossible once the judgement rule changes.

On this platform sources are listed per answer, so citation share can be counted directly, which determines what can be measured here and what can only be estimated.

Fix the denominator first. Adding or dropping questions mid-flight, or sampling unevenly, means two months of data are not the same measurement — a bigger problem than sampling noise.

Judging a mention needs both string matching and semantic review: string matching alone pulls in unrelated things with the same name, semantic review alone misses names run together inside answer tables.

Never compare a cumulative figure with a monthly one. Stretch the window and almost every brand approaches 100%, destroying the signal. Report daily-average or monthly figures and label the column.

Citation-source behaviour differs sharply by platform, so a rising blended figure may be one platform moving. Record per platform, and report English and Chinese separately.

Whether visibility produced visits belongs on first-party analytics and server logs. Requests that do not execute JavaScript never reach front-end analytics.