Tracking AI visibility: a routine for the week, the month and the quarter, and what to ask of a tool

Why a routine, rather than a number

Answers change for reasons on every side: something on your site, something a competitor did, something inside the assistant. A visibility number on its own records that something moved and explains none of it. A routine keeps three records side by side, what the answers said, what your site and listings said, and what you changed, so that a movement can be traced to a cause, or honestly left unexplained.

It also stops the two habits that make tracking useless: checking an assistant every day, which records the run-to-run variation and nothing else, and changing the questions whenever a result disappoints, which ends the series. How to measure AI visibility covers the arithmetic of telling a change from noise.

Each week: the things that break without warning

Some failures arrive overnight and take every answer with them. A security plugin updates, a CDN setting changes, a redesign ships with a noindex tag still on the new pages. In our audit of 9,483 small business websites, 7.2% served our crawler a challenge page instead of the site, a block nobody on those businesses had chosen. Ten minutes a week catches it early.

• Run the crawler access check and compare it with last week's result.
• Fetch the homepage and one service page from outside your own network, and confirm a page came back rather than a challenge.
• View the source of the pages that changed this week and search for noindex.
• Check that the hours, prices and phone number on the site still match your map listing, especially after any change.

Each month: the panel

The same ten to twelve buyer questions, put to ChatGPT, Claude, Gemini and Perplexity three times each, with every answer read and kept word for word: whether you were named, recommended or first, who else was named, what was said about you, and which pages the answer cited. Setting up prompt monitoring gives the panel file and the record format, and our methodology is a narrower procedure run for clients: twelve questions put to one assistant, Claude, before the fixes, again once they are live and every month with monitoring, with a record of whether the business was named, whether it came first, and who else was.

Run it on the same days each month, from the same place, in the same way. Then write one sentence about what moved and what did not, before looking at anything else. The sentence is the part of the month worth keeping; GEO reporting for management builds a one-page report around it.

Each quarter: the competitors, and the panel itself

• Reread the pages of the businesses named most often across the quarter, and note what changed on them. Competitor analysis in AI answers has the fields to record.
• Review the panel. Retire a question no buyer asks, with a note and a date, and add new questions as new lines with their own start date. Never reword an old question in place; its history belongs to the old wording.
• Recheck your name, address and phone across your pages and the listings that carry them, since listings drift one edit at a time.
• Lay the three monthly sentences side by side and decide what the next quarter's work is.

The change log

Keep one log beside the panel for everything that could move an answer, whoever made the change. It is what lets you say, three months later, that the named rate on repair questions rose after the repair page went up, or that it fell the week the host changed its bot settings.

What to ask of a tracking tool

A tool can take the monthly runs off your hands. Before paying for one, ask it the same questions you would ask a person doing the work.

1. Does it run each question several times on each assistant, or once?
2. Does it keep every answer word for word, and let me read them?
3. Can I fix the questions, and the place they are asked from, so months compare?
4. Does it report each assistant separately rather than blending them into one number?
5. Does it record the sources an answer cited?
6. Does it say how it asks: through the assistant's own product with search on, or through a programming interface that may answer differently from what a buyer sees?
7. Can I export everything, so the series survives if I stop paying?

A tool that fails the second question is giving you a number without its evidence. One that fails the sixth may be tracking a different thing from the answers your buyers get. Either way, someone still has to read the answers; a count of mentions says nothing about whether what was said was true.

What tracking cannot do

• Change the answers. Tracking tells you where you stand; the work on your pages and listings is what moves it.
• Count customers. Add an AI assistant to the question you ask new customers about how they found you, and keep that count beside the panel.
• See one buyer's conversation. Assistants may take a person's earlier questions and settings into account; a panel asks each question fresh, the same way every month, and that is what makes it a series.