To track AI search rankings over time, lock a fixed panel of buyer prompts, run it against the same providers on a weekly schedule, and record mentions, positions, and cited sources the same way every run. Consistency is the entire method: change the prompts, the providers, or the counting rules midstream and your trend line stops meaning anything.
This page is the trend methodology. If you have not built a first measurement yet, start with how to measure AI visibility and come back.
Build the panel once, then freeze it
Write 20 to 50 prompts that cover how buyers actually ask: best-tool recommendations, comparisons against named competitors, alternatives to the category leader, and pricing or fit questions, with a few phrasing variants of each. Then freeze it. The temptation to "improve" prompts each week is the most common way teams destroy their own data, because a reworded prompt is a different prompt, and its results are not comparable with last week's. Your panel is an instrument; recalibrating it every week means never measuring anything twice.
Why weekly is the right cadence
Weekly beats daily because AI answers vary run to run; day-to-day movement is mostly that variance, not real change, and chasing it burns attention. Weekly beats monthly during active work because content, entity fixes, and retrieval surfaces like Perplexity move on a timescale of weeks; a monthly reading arrives too late to steer the next sprint. At steady state, monthly is fine. The fuller cadence logic, including when to run event-triggered extra scans, is in how often to run AI visibility scans.
Counting rules, decided before run one
- A mention is your brand named in the response text; decide upfront whether product names and misspellings count.
- Position is your rank among all brands named in that response; record it every time, not just mention yes or no.
- Run each prompt several times per measurement and record the mention rate across runs, not a single coin-flip result.
- Log which providers had search or retrieval active, since ChatGPT with and without search mode are effectively different surfaces.
- Keep sessions clean: no logged-in history or memory coloring the answers.
Changing the panel without breaking the trend
Markets shift, and eventually your panel needs new prompts. Add them as a versioned expansion: keep the original panel running untouched, tag the new prompts as panel v2, and report the two series separately until v2 has its own history. Never edit or delete old prompts to make room. Annotate the timeline, too: note when you shipped content, fixed crawler access, or got a PR hit, so movements in the line can be traced to causes. If maintaining all this by hand sounds like a part-time job, that is because it is; a free scan repeats the same methodology on schedule without the spreadsheet.
Start the trend line this week
Every week without a baseline is a week of progress you cannot prove later. Run a free scan now and repeat it on schedule; the trend takes care of itself.
Frequently asked questions
How do I track how often AI mentions my brand?
Run a fixed set of buyer prompts against each AI engine on a schedule, several runs per prompt, and record the share of responses that name your brand. That mention rate, tracked weekly with an unchanged prompt panel, is the AI equivalent of rank tracking. No provider offers logs or analytics, so sampling is the only method.
How do I monitor AI visibility weekly without it eating my week?
Automate the repetitive part. Manually rechecking 30 prompts across four providers with multiple runs each is hours of work that a scheduled scanning tool does identically every time, which also removes human counting drift. Reserve your manual time for reading the deltas and deciding what to do about them.
Why do my AI visibility numbers bounce around week to week?
AI responses are probabilistic, so some movement is natural variance rather than real change. Reduce the noise by running each prompt multiple times and watching multi-week trends instead of single readings. A sustained shift across several weeks and multiple prompts is signal; a one-week spike usually is not.
Related answers