How often should I track AI visibility?

Monthly is a sensible rhythm for most businesses. Daily checks mostly capture noise, and AI recommendations tend to move over weeks and months.

Recommendations, mentions and citations as separate rates, plus the sources that recur, across a fixed question set and each assistant.

How do I know a change is real?

Compare the same questions across periods and check whether the change is larger than the normal run-to-run variation. If not, call it no clear change.

Not at first. A spreadsheet and saved answers work for a small set. Tools help once you have many questions, assistants and runs.

How to track AI visibility over time: Citewake

To track AI visibility over time, fix a set of questions, ask them on a regular schedule in a clean session, keep every answer in full, and compare the same questions period to period. Report rates for recommendations, mentions and citations separately. Judge a change real only when it is larger than the ordinary variation between runs. Monthly is a sensible rhythm, because recommendations move over weeks, not hours.

What does tracking need to get right?

The foundations are the same ones that make a one-off test fair, applied with more discipline.

A frozen question list. If the questions change, you cannot tell whether results moved because of you or because of the wording.
A fixed method. Same assistants, same search setting, same session type, same number of repeats.
A fixed definition of each measure. Recommendation, neutral mention, negative mention, citation. Written down.
Saved answers. Store the full text and sources, so that you can recount or reread when your thinking changes.

Build a benchmark set of questions and test AI answers properly cover the first two in detail.

What should you measure?

Track several things separately rather than blending them.

Measure	Question it answers
Recommendation rate	In what share of answers are we put forward?
Mention rate	How often are we named at all, and in what tone?
Own-site citation rate	How often do answers cite our domain?
Third-party presence	Do the pages cited list us?
Share of voice	What share of all recommendations do we hold?

Keep a view per assistant and per type of question. A flat overall figure can hide that you gained on one assistant and lost on another. Share of voice in AI answers explains the relative measure.

Also track the sources. Which pages are cited for your category, and has that list changed? New sources appearing, or old ones dropping away, often explain why your results moved.

How often should you check?

Monthly suits most teams, for three reasons. Answers vary run to run, so a single day is noisy. Recommendations tend to shift over weeks as pages change and models update. And a regular monthly pass keeps all the answers in a period close together, so they describe one moment.

Daily checks have a role in a narrow case, such as watching one question after a specific change. But as a general programme they produce a lot of noise and invite false alarms. Why AI visibility should be measured monthly, not daily makes the case in full.

Citewake's method works on monthly editions for the same reason: every question is asked of every assistant in a single pass, so that an edition describes one moment.

How do you tell a real change from noise?

Compare like with like: the same question, the same assistant, this period against last. Then ask whether the change is bigger than what you would see from run-to-run variation alone.

A practical approach:

Run your repeats, so each question has a rate rather than a single result.
Compute the change per question, then summarise across questions.
Be cautious with small changes. If a rate moves from 30 percent to 33 percent on a modest number of runs, that is probably not a change at all.
Say "no clear change" when the data does not support more.

Mark known disruptions on your timeline, such as a major update to an assistant, because they can shift answers independently of anything you did.

What do you do when a result drops?

First, check whether it is a real drop. Compare the change with ordinary run-to-run variation, and look at how many runs and questions sit behind it. Many apparent drops are noise.

If the drop looks real, look at the details. Did it occur on one assistant or all? On certain kinds of question only? Did the cited sources change, for example a roundup that used to list you no longer appearing? Did a competitor gain, which would reduce your share without any change in your own rate?

Then check what changed on your side: pages updated, redirects altered, access blocked by a site change. Finally, check whether an assistant was updated around that time, since model changes can shift answers independently of anything you did.

Resist the urge to react to a single month. Note the change, look for cause, and see whether the next period confirms it before committing resources.

How do you connect results to what you did?

Keep a log of your own moves: pages published, comparison content added, listings secured, corrections requested. Put those dates beside your results. You are looking for movement that follows an action, and for sources that appear once an action is complete.

Be modest about cause. A date match is a clue, not proof. Other things change at the same time. If a cited page is the very article you published, that is about as close to a direct link as the evidence gets.

How should you set it up in practice?

Start simple. A spreadsheet with one row per question per run, columns for assistant, date, brands named and roles, and a link to the saved answer, is enough for a small programme. Move to tooling when manual runs become the bottleneck.

Whatever you use, insist on traceability: every figure should lead back to the words an assistant used. For a quick baseline before you set up the routine, the Citewake checker offers a first look.

What is the short version?

Fix the questions and the method, ask monthly, keep every answer and report separate rates for recommendations, mentions and citations. Call a change real only when it is larger than ordinary variation, log your own actions beside the results, and avoid reacting to a single noisy period.