How to report a client's visibility inside AI answers.

Being cited and being visited are separate events. Somebody can read an answer that named the client, get what they came for and never click, and that is the outcome the format is designed to produce rather than a failure of it.

So a visit count measures the tail. Report it as your AI performance and you are reporting the narrow part while calling it the whole, and the figure will move for reasons that have nothing to do with whether you are being cited more often.

Five things, and only the last one is in the client's analytics.

• Whether the client appears for a chosen prompt

  Ask an assistant the question a buyer would ask and record whether the client was named. This is the closest thing to the metric everyone wants, and it is a sample rather than a count.

• Which page gets cited

  When a citation carries a link, the URL tells you what the client is being cited for. It is usually an explanatory page rather than a product page, which is useful for deciding what to write next.

• Who appears alongside the client

  The other names in the same answer are the competitive set as the assistant understands it, which is often not the set the client would have listed.

• Assistant crawler activity

  Server logs show which assistants fetched which pages. This is separate from visits and it is the one signal here that is a full count rather than a sample.

• The visits that do arrive

  Worth keeping as a floor and a trend, on the understanding that a referrer is not always passed, so the figure is never a total.

An assistant's answer varies between runs, between accounts and between regions. One check is an anecdote, and a single anecdote presented as a metric is worse than nothing because somebody will act on it.

What makes it a measurement is holding the method still. Fix the list of prompts, fix the cadence, run the same set every time, and record the full answer rather than a yes or no, because the wording is where you find out what the client is understood to be. When you change the prompt list, the series starts again, and that has to be written down beside the number.

This is the step that gets skipped and it is the one that decides whether any of the above was worth doing. Results pasted into a sheet that is overwritten each month produce a current figure and no history.

Keep one row per prompt, per assistant, per run date, with the answer text. Add to it rather than replacing it. Once the rows accumulate, the interesting question becomes answerable: not whether the client appears today, but when they started appearing, and what changed in the same week.

State the method in the report. This is a fixed list of prompts, asked on this date, and here is the list. A client who can see the method reads a wobble as a wobble; a client shown a bare percentage reads any drop as a collapse and asks what you did wrong.

Keep the sampled visibility and the referred visits as two separate lines that are never added together. They measure different events, and combining them produces a number that cannot be explained when somebody asks what it counts.

Sources without an API are ingested too, including spreadsheets and emailed files, so a prompt log you collect by hand lands in the same warehouse as the ad platforms and the analytics rather than living in a document beside them.

Syncs add to the tables rather than replacing them, which is the property this measurement depends on more than most. Threshold and change checks run on those tables, so a prompt set that stops returning a client can raise an alert instead of being noticed a quarter later.