Set up LLM citation tracking: prompt set design, sampling method, metrics that survive non-determinism, and how to report AI visibility without overstating precision.
LLM citation tracking measures how often assistants reference your brand when answering questions in your category, and how accurately they describe you. It is the reporting layer underneath any AI search program, and it is easy to do badly.
The central difficulty is non-determinism. The same prompt produces different answers across sessions, so naive measurement produces numbers that move for no reason. A defensible method handles that directly.
Design the prompt set first
The prompt set is the instrument. Get it wrong and every subsequent number is meaningless.
Write the questions in buyer language, conversational and specific. Group them by intent: category definition, comparison, vendor selection, and direct brand questions. Include several phrasings of each question, because small wording changes shift results.
Freeze the set. Adding and removing prompts between runs makes period-over-period comparison invalid. Version it instead, and note the version on every report.
Thirty to sixty prompts is a workable size for most businesses. Coverage of the real buying questions matters far more than volume.
Sample rather than check
Run every prompt multiple times, on different days, in sessions without personalization or conversation history. Three runs is a practical minimum; five gives noticeably steadier numbers.
Record for each run: the assistant and model version if exposed, the date, the full answer text, every cited URL, whether your brand appears, and how it is described.
Your primary metric is citation rate: the share of runs in which you appear, per prompt and aggregated. A binary appeared-or-not from a single run should never be reported as a result.
Metrics worth reporting
Citation rate by prompt group, trended over time. This is the headline.
Share of voice against named competitors across the same runs, which shows whether you are gaining or the whole category is shifting.
Description accuracy, scored simply: correct, partially correct, or wrong. Track wrong descriptions as defects with owners, because they are usually fixable at the source.
Source overlap, meaning which domains are cited when you are not. This is the most directly actionable output of the whole exercise, since it names the places corroboration lives in your category.
Referral traffic from assistant domains, understood as a floor rather than a full count, since many assistants pass no referrer.
Reporting without overstating precision
State the method on the report: prompt set version, number of runs, dates, assistants covered. State the limits plainly: sampled not exhaustive, non-deterministic outputs, incomplete referral attribution, and model updates outside your control.
Compare trends across at least three measurement periods before drawing conclusions. Single-period movements are usually noise, and treating them as signal leads to reactive work that wastes budget.
When a metric jumps sharply, check for a model update or a change in the assistant's search behavior before attributing it to your own work.
Building versus buying
Manual tracking in a spreadsheet is entirely viable at the scale most businesses need, and it forces familiarity with the actual answers, which has its own value.
Tools help once you need frequent runs across many prompts and assistants. Evaluate them on whether they disclose sampling method and run counts. A tool reporting a single number without describing how it was produced is not measuring anything you can defend.
Whichever route, keep the raw answers. Aggregates hide the misdescriptions and competitor patterns that drive the actual work.
Frequently asked questions
How often should we run the measurement?
Monthly suits most programs. Weekly is justified during an active remediation push, but expect more noise at that cadence.
Does personalization affect results?
Yes. Run in clean sessions without history so results reflect the general case rather than one profile.
What citation rate is good?
There is no universal benchmark. Judge against your own baseline and against named competitors in the same runs.
Can we track this in analytics alone?
No. Analytics captures only referred visits, which undercounts substantially and tells you nothing about how you were described.
Explore our software and web capabilities or contact Voquarn Code to set up measurement for AI search visibility.
Written by
Moueen Togarvi
Founder & CEO at Voquarn Code, focused on product engineering, search growth, and practical AI systems.