Why is measuring AI citations hard?
Measuring AI citations is hard because no AI engine publishes a report of when it cited your site. Search engines expose impressions and clicks through a console; AI answer engines expose nothing equivalent. The same prompt can produce different answers and different citations in different sessions, locations, and days.
That means measurement is sampling, not reading a dashboard. The method below is what a Search Visibility Digital FTE on ftes.ai runs as its measurement sweep, and it is the honest version: dated samples, recorded citations, and explicit gaps.
What do you sample?
Sample a fixed set of prompts that represent the questions your buyers ask: definitional questions, comparison questions, how-to questions, and vendor-selection questions in your category. Run each across the engines that matter, such as ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews or AI Mode, on a fixed schedule such as weekly.
For each prompt and engine, record whether your domain was cited, in what position among the citations, which competing sources were cited, and what the answer said about you if anything. Store the answer text with the date, because engines change and the history is the evidence.
- Prompt set: fixed, representative, revised quarterly
- Engines: each one your buyers use, run separately
- Recorded: cited or not, position, competing sources, answer text, date
- Cadence: weekly, with repeated samples to see variance
What corroborating signals exist?
Two signals corroborate the samples. Referral traffic from AI surfaces appears in analytics when a user clicks a citation; it is small and undercounts, because many answers are read without a click, but its trend is informative. Branded search lift, meaning more people searching your name, often follows being cited in answers, because users read the name and then search for it.
Neither is a citation count. Both move in the same direction as citation presence, and together with the samples they make a case that a single metric cannot.
How do you report honestly?
Report citation presence as a share of sampled prompts per engine, with the sample size and date, and with ranges where repeated samples varied. Report the competing sources, because being cited alongside the right ones matters. Report referral and branded trends as corroboration, labelled as such.
Report what is unmeasured as unmeasured. If an engine is not sampled, say so. If a metric cannot be computed from the data you have, do not estimate it. The ftes.ai performance view is built on this rule: if the platform cannot measure it, the UI does not show it.
What do you do with the results?
Use the results to decide what to write. Prompts where you are not cited and a competitor is show you which questions your content does not answer well, or does not answer at all. Prompts where citations shift week to week show you where the engine has not settled and a better answer can win.
A Search Visibility FTE turns this into a loop: measure, propose the content that would earn the missing citation with a worth-it verdict, pass the fact gate, publish on approval, and measure again. The citation samples are the targets, and the performance view is where the loop is reported.
What does a labelled example of the citation record look like?
As a labelled example with your own results to fill in: the record is a table with one row per prompt per engine per sample date. A row for the prompt what is an AI employee, run on a named engine on a stated date, holds whether ftes.ai was cited, its position among the citations if so, the other domains cited, and the first sentence of the answer's description of the term, stored verbatim. The same prompt on the same engine a week later is a new row, so variance across weeks is visible rather than averaged away.
Beside the table sit the corroborating series from analytics: sessions referred from AI surfaces per week and branded search impressions per week from the console, each with the date. The report reads across them: share of prompts cited per engine this month against last, with sample size, the competitors cited most often on the prompts where you are not, and the referral and branded trends as corroboration. Engines not sampled and metrics not computable are listed under unmeasured. The Search Visibility FTE keeps exactly this record and reports from it in the Performance tab.