How to Track AI Citations
Fix a list of twenty to thirty buyer questions before looking at any results, ask them on a schedule across the engines your buyers use, and record whether you appear, who else does, and the date. Citation share is appearances divided by questions asked, and it only means something with the denominator attached. Watch your access log for AI crawler hits at the same time. Nobody can tell you how often agents asked about your category, and position is the one measurement that cannot be reconstructed later.
Tracking AI citations means two different things, and one of them cannot be done by anybody, including the vendor charging you for it. Separating them is most of the work of building reporting that survives a sceptical question.
Two different questions, and only one is answerable
"Track AI citations" usually means one of two things, and it is worth separating them before choosing a tool, because one of them cannot be done by anybody.
| The question | Answerable? |
|---|---|
| How often did people ask AI about my category? | No. No engine publishes query logs. Any number here is a vendor's own traffic or a model's guess. |
| When somebody does ask, am I in the answer? | Yes. Ask the question yourself, on a schedule, and record what comes back. |
The method, in five steps
1. Write the question list first
Twenty to thirty questions a real buyer would ask, written before you look at a single result. Mix the obvious ones with the ones you would rather not know about, and include your category's comparison questions, because those are where competitors get named.
The discipline is the whole method: the list is fixed and you do not edit it because a question is unflattering. A question set curated after seeing results measures your curation, not your visibility.
2. Decide which engines count for you
They behave differently and there is no reason to track all of them. A business selling to developers cares about different surfaces than one selling to procurement teams. Pick the two or three where your buyers actually are and track those properly rather than five badly.
3. Ask on a schedule, and record the whole answer
Record four things per question per run: whether you appeared, which competitors appeared, what was cited as the source, and the date. That last field is the one people skip and the one that turns a spreadsheet into a trend.
Answer engines are non-deterministic, so a single run is a sample. Weekly is enough for citation presence. Which is a different cadence from position tracking, for a reason worth knowing.
4. Watch your access log at the same time
The cheapest signal in this entire discipline and the most neglected. Your server already records every request from GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and Claude-SearchBot. Two things it tells you that nothing else can:
- Whether you are reachable at all. If those user agents never appear, nothing downstream can work, and the most common cause is still a robots.txt written in 2023 to keep scrapers out.
- Which pages they actually fetch. A crawler returning repeatedly to three pages is telling you which of your content it finds useful, which is more informative than most dashboards.
5. Compute one number, and state its denominator
Citation share is appearances divided by questions asked. Eleven of thirty is a number you can act on. "Strong AI visibility" is not.
Always carry the denominator. "You appear in eleven answers" is meaningless without "of thirty questions asked, on 5 September". A rate without a denominator is decoration.
The one thing you cannot buy back
Citation presence can be measured retroactively in a rough way, because you can ask the questions today and get today's answer. Position cannot.
Where you ranked last month, across engines and discovery registries, is gone unless somebody recorded it at the time. No vendor can sell it to you, no archive holds it, and no amount of budget reconstructs it. This is the single strongest argument for starting a measurement programme before you have a strategy: the cheap part is the part that expires.
What a workable weekly view contains
- Citation share over the fixed question set, with the denominator.
- The questions you lost, named, with whoever was cited instead.
- Crawler hits by user agent from the access log, with the pages they fetched.
- Position across the surfaces you care about, with the movement since last week.
What is deliberately absent: any volume figure, any visibility score out of a hundred, any opportunity estimate. Every line is a count or a position and every line names its source.
Four ways this goes wrong
| Mistake | Why it ruins the data |
|---|---|
| Editing the question list | You end up measuring the questions you chose to keep |
| Recording a failed request as a zero | An engine timing out becomes "we lost the citation", and you go and fix a page that was never broken |
| Measuring once | Non-deterministic systems need a series. One run is an anecdote |
| Blending everything into one score | The weights are an opinion, and a falling score tells you nothing about which part moved |
Then act on it, with the levers that are measured
Tracking is only worth the effort if the losing questions change something. The techniques with real evidence behind them come from a controlled study presented at KDD 2024, which tested method families and found three that beat baseline: cite sources, add direct quotations, and include named statistics. The same study found keyword stuffing performed worse than the baseline.
So the loop is: find the questions where a competitor is cited and you are not, read what the cited page does that yours does not, and edit for specificity and attribution (worked examples) rather than for repetition.
One structural note. Off-site brand mentions correlate roughly three times more strongly with being recommended than backlinks do, and generative answers prefer third-party sources over a brand's own pages. Some of what your tracking reveals will not be fixable by editing your own site, and recognising which is which saves a lot of wasted effort.
How do you track AI citations?
Fix a list of buyer questions before you look at any results, ask them on a schedule across the engines that matter to you, and record for each one whether your brand appears and who else does. The metric is the proportion of that fixed list where you are cited, tracked over time. A list edited after seeing results measures your editing.
Can you see how many people asked ChatGPT about your brand?
No. Answer engines do not publish query logs and there is no equivalent of a keyword planner, so any monthly figure for how often you were asked about is either a vendor's own test traffic or a model estimate. What is measurable is the answer, not the demand.
What is citation share?
The proportion of a fixed question set where your brand appears in the generated answer. It only means something if the question list is defined in advance and kept stable, and if the same questions are asked on the same schedule.
Why do two AI visibility tools report different numbers?
Because answer engines are non-deterministic and the tools ask different questions at different times from different locations. A single run is a sample rather than a measurement, which is why only a repeated fixed question set produces a usable trend.
Can you track AI citations for free?
Yes, and doing it by hand for a month is a good way to learn which parts are worth automating. You need a question list, a spreadsheet, a schedule, and your server access log for the crawler side. The paid tools mostly buy you frequency and history.
Want to see where you are cited today?
The free audit checks crawler access, structured data and answer-shaped content, and shows what an AI system can currently learn about your brand.
Run the free audit