Define each event
A mention means the answer names the brand. A citation means the answer links to your source. A referral is a visit attributed to the service. A conversion is an action with your own explicit definition. Keep these columns separate; an answer can mention a company while citing someone else.
Use one observation row per prompt, engine, locale, date and run. Store the answer and URLs so another reviewer can reproduce your classification. Normalize your own domain variants deliberately and preserve the original URL. Do not count a competitor’s review of your brand as a citation to your website.
Keep a fixed question panel
Build questions from genuine customer needs: category learning, alternatives, implementation and evaluation. Freeze the panel for a measurement wave. If you add easy branded prompts midway, report the new panel separately; otherwise the apparent improvement may be a denominator change.
Record failed or unavailable observations instead of silently dropping them. The worksheet separates planned from completed observations and computes citation frequency only over completed observations. It does not imply that missing results would resemble observed results.
Read the metrics carefully
If 60 of 80 planned observations complete and 15 cite your domain, coverage is 75% and observed citation frequency is 25%. Calling the result “25% market share” would be unjustified: it describes this panel under these conditions. Repeat runs and report variation rather than treating one answer as a permanent rank.
Google reports AI-feature appearances within overall Web search performance. Do not infer a separately measured AI channel from an aggregate report unless the reporting tool actually supplies that distinction.
Connect visibility to useful visits
Measure landing pages, qualified actions and downstream outcomes with consistent attribution windows. Referral data can be missing or incomplete. A self-reported discovery field can add context, but it is not precise click attribution. Reconcile the systems rather than forcing all observations into one number.
Before buying a monitoring platform, ask whether it exports raw answers, linked sources, timestamps, locale and the exact prompt panel. Test repeatability and failure reporting. A polished score with no inspectable observations is difficult to audit.
Use the audit to decide the next improvement
Compare cited and uncited pages for missing answers, unclear scope, weak evidence or access problems. Prioritize changes that also help a real reader. Keep a change log so later observations can be interpreted, while acknowledging other causes such as engine changes and competing pages.
The calculator accepts counts you collected. It does not run an AI search or validate your classifications. Percentages are descriptive and are not causal estimates or forecasts.