Guide
How does LLM visibility tracking work?
LLM visibility tracking means running a fixed set of questions against AI assistants on a schedule and recording what comes back. This guide covers query-set design, handling variance, capturing citations, comparing against competitors, and the limits you should state out loud.
The short answer
Fix the questions. Repeat them on a schedule. Record mentions, citations, prominence and competitor presence per run. Report trend, not snapshots. State the variance.
That is the whole method, and the discipline is in refusing to skip the repetition — which is exactly what a screenshot-based approach does.
Designing the query set
The query set is where the value is created or lost. Five principles matter:
- Ask what buyers ask. Real phrasing from sales calls and support tickets, not keyword-tool phrasing. People talk to assistants in sentences.
- Cover the journey. Definitional questions, comparison questions, selection criteria, cost expectations and objections each surface different sources.
- Include unbranded questions. Asking “what does WindspeedSEO do?” tests whether the assistant knows you. Asking “what software manages SEO for a small business?” tests whether you get considered. The second is the commercially interesting one.
- Keep it stable. Changing the query set breaks your trend line. Add deliberately, remove rarely, and record when you did either.
- Keep it small enough to repeat. A set you run monthly beats a larger set you run once.
What to record on each run
| Field | Why it matters |
|---|---|
| Brand mentioned | Baseline presence, even without a link |
| Page cited | The actionable signal — tells you what content is being used |
| Which page | Directly instructive about what to produce more of |
| Position in answer | Primary recommendation versus passing alternative |
| Framing | Recommended, listed neutrally, or mentioned as a competitor’s rival |
| Competitors present | The comparison that gives your own number meaning |
| Run timestamp | Required for trend and for correlating with content changes |
Handling variance without hiding it
Variance is the defining property of this measurement, and the temptation is to smooth it into a single confident score. Resist that.
Report presence as a frequency — “mentioned in seven of ten runs” — rather than as a binary. Show the trend across periods rather than the latest run. And flag queries where variance is high separately, because a query that swings between full recommendation and complete absence is telling you something different from one that is consistently negative.
A change of one or two runs is noise. A query that moves from consistently absent to consistently present across two measurement periods is a result.
Correlating with content changes
You can observe that a page began being cited after it was published or substantially revised. You cannot prove the revision caused it — the retrieval behavior may have changed, or a competing source may have been removed.
The honest framing is correlational: record what changed and when, observe what followed, and accumulate enough instances that a pattern becomes credible. Treat each individual case as suggestive rather than conclusive.
Limitations to state in any report
- Sampled measurement, not complete observation.
- Cannot replicate individual users’ context or history.
- Covers supported experiences, not every AI product.
- Correlation only; causation is not established by this method.
- Vendor behavior changes without notice, which can shift results independently of anything you did.
Continue with what is AI search visibility, browse all guides, or see the WindspeedSEO capability.
Frequently asked questions
How does LLM visibility tracking work?
You define a query set of questions your buyers actually ask, run them against supported AI experiences on a schedule, and record for each run whether your brand was mentioned, whether a page of yours was cited, which page, where in the answer you appeared and which competitors appeared alongside you.
Because answers vary, the reported result is frequency and trend across runs rather than a single state.
How many queries should be in the tracked set?
Enough to cover the buying journey and few enough to run repeatedly with consistent effort — typically a few dozen. Coverage across question types matters more than raw count: definitions, comparisons, selection criteria, pricing expectations and objections.
How much do answers vary between runs?
Enough that a single run should never be treated as a measurement. Some queries produce broadly stable answers; others rotate their sources noticeably. Good practice is to report variance per query rather than averaging it away, because a high-variance query and a stable one warrant different responses.
Should I track competitors in the same query set?
Yes, because it is usually the most informative part of the exercise. Knowing that a competitor appears in eighty percent of runs where you appear in twenty tells you more than your own score does, and the pages they get cited for are directly instructive.
Can I influence what an assistant says about my brand?
Indirectly and imperfectly. Assistants draw on what is published about you across the web, so the practical levers are the accuracy and consistency of information about your business, and the quality and directness of content that answers the questions being asked.
There is no submission process, no correction mechanism and no guarantee.
What are the honest limits of this tracking?
It samples rather than observes everything, it cannot replicate every user’s context, it cannot establish causation between a content change and an appearance, and it measures supported experiences rather than every AI product in existence. Those limits should be stated in any report that uses this data.
Track AI visibility with a repeatable method
WindspeedSEO runs a fixed query set on a schedule and reports mentions, citations, prominence and competitor presence as trends.