How share of citation measurement actually works when you ask an engine
Share of citation measurement asks a fixed set of questions of an answer engine, records the source list attached to each answer, and counts how many of those sources belong to you. The result depends on the denominator chosen, so state whether it counts citations returned or questions answered.
Somebody in the commercial team forwards a screenshot. They asked an assistant which trade shows matter in industrial automation, your show came up, and the message says "we're doing well in AI". Share of citation measurement is what turns that screenshot into a number you can put in front of a board, and the method is mechanical enough that the hard part is agreeing what to divide by.
The whole thing is a loop: ask a question, keep the list of sources the answer came with, count how many of those sources are yours, repeat.
What counts as a citation
Answer engines return two different things and teams conflate them constantly.
A citation is a source the engine attaches to an answer, usually as a link, usually rendered as a numbered reference or a card underneath the text. It exists as structured data. Perplexity's chat completions API documents a citations array holding the URLs of sources used, and a search_results array of objects carrying title, url, snippet and date (Perplexity API reference, checked August 2026). That is the object you are counting.
A mention is your brand named inside the answer text with nothing linked. It is worth tracking and it belongs to a different series, which is why answer engine visibility as reach is measured separately.
For share of citation, count only the structured source list. Mixing the two produces a number that nobody can reconstruct, because one comes from a field and the other comes from a regular expression run over prose.
The loop, in the order you run it
Five steps, and each one has a decision in it that should be written down once and then left alone.
Fix the question set. Forty questions is a workable starting size for a single vertical. The questions have to stay identical between runs, and how you choose them in the first place is a sampling problem that deserves its own written frame.
Fix the engines. Query each one through its API so the source list comes back as data. Anything you cannot query stays out of the count entirely.
Run every question against every connected engine, in a fixed order, on a stated date. Store the raw response, including the full source list, before you compute anything. You will want the raw payload the first time a number looks wrong.
Normalise the URLs. Strip query strings, resolve redirects, lowercase the host, and map every host to an owner. This is where most of the manual work lives, because your portfolio probably owns eleven domains and two of them redirect through a marketing tool.
Count. Total sources returned, sources belonging to you, and questions in which at least one of your sources appeared.
The ownership map is the step people skip and then regret. An event group typically has the show site, the media brand site, an association microsite it operates on somebody else's behalf, and a legacy domain from an acquisition that still redirects. Decide once whether each of those counts as yours, write the decision in the map with a date, and stop relitigating it every quarter. A citation to a syndication partner who republished your article is the hardest case, and my rule is to exclude it from your own share and track it as a separate line, because you cannot edit that page and the credit is genuinely shared.
Log enough per run that the number can be rebuilt without rerunning anything. That means the question text, the engine, the model identifier the response carried, the timestamp, the full raw response, the normalised source list, and the owner assigned to each source. Roughly seven fields per source, which on 312 sources is a table of about 2,200 cells per run. Small enough to keep forever, and the first time somebody disputes a figure you will be glad you have it.
Which denominator are you reporting?
Here is a real shape of run. Forty questions, one engine, 312 sources returned in total, an average of 7.8 per answer. Nineteen of those 312 belong to your domains.
Nineteen divided by 312 is 6.09 per cent. That is share of citation in the strict sense: the fraction of the engine's reference list that you occupy.
Now count differently. Those nineteen sources are spread across eleven of the forty questions, because two answers cited three of your pages each and one cited two. Eleven divided by forty is 27.5 per cent. That is presence: the fraction of questions where you appear at all.
Same run, same data, 6.1 per cent and 27.5 per cent. Both are honest. Reporting either one without saying which is not, because a reader assumes the larger one when you quote the smaller and the smaller one when you quote the larger.
My preference is to report presence as the headline and share of citation as the supporting figure. Presence answers the question the commercial team is actually asking, which is whether a buyer who asks this gets told about us. Share of citation answers a question about competitive crowding, which matters later, once presence is above zero.
Both numbers are more useful once you know how crowded the list is. Sort the 312 sources by owner. If the top three domains account for 71, 44 and 38 sources, that is 153 between them, or 49 per cent of every source returned across the whole run. Your 19 sit in the long tail with everybody else. That single line reframes the target, because moving from 6.1 to 9 per cent means taking share from three specific sites you can name, and two of them are probably an association and a directory that will still be there next year.
There is a third denominator worth computing and worth keeping out of the headline. Weight each appearance by its position in the source list, on the theory that the first source is read more often than the eighth. The weights are arbitrary, nobody has published a click distribution over answer engine source lists that I can point at, and an arbitrary weighting inside a headline metric is how a measure loses its audience.
How much does the number move between runs?
More than a monthly report can absorb, and this is the single most useful thing to establish before anybody sets a target.
Run the same forty questions three times in one afternoon, changing nothing. A plausible result is nineteen of your sources on the first pass, twenty-one on the second, sixteen on the third. Mean 18.7, range of five, which is 27 per cent of the mean. In share of citation terms that is 6.1, 6.7 and 5.2 per cent.
If your reporting cadence is monthly and your run-to-run spread is a point and a half, then a month-on-month move of a point is noise and you will spend the meeting explaining it anyway. The fix is to run the set three times per reporting period and report the mean with the range beside it. Three runs of forty questions is 120 API calls, which costs almost nothing and removes the largest source of false narrative in the whole exercise.
Aggarwal and colleagues introduced generative engine optimisation as a formal framework at KDD in 2024, along with GEO-bench and a set of visibility metrics, and demonstrated visibility gains of up to 40 per cent from their strategies. Their own framing treats visibility as something measured over a benchmark of many queries, which is the right instinct. A single question tells you nothing, and forty questions run once tells you less than you think.
What the metric will not tell you
Three things, and it is worth being blunt about each because all three get read into the number anyway.
It does not tell you whether the answer was right. A citation to your exhibitor list page attached to an answer that gives the wrong hall numbers still counts as a citation, and scoring whether the citation was correct is a separate series with a separate method.
It does not tell you anything about engines you did not query, and quietly implying otherwise is the most common defect in published visibility numbers. Declaring engine coverage beside the figure is what makes the denominator meaningful.
It does not transfer between verticals. Aggarwal and colleagues found the effectiveness of their optimisation strategies varied across domains, which means a share of citation of 6 per cent in industrial automation and 6 per cent in packaging are not comparable achievements, and a portfolio benchmark across eight shows in eight verticals is a table of numbers with no common scale.
Where this stops
The measurement is a description of what one engine returned to one set of questions on one day. It contains no causal claim at all. If the number goes up after you publish twelve pages, the number went up and you published twelve pages, and untangling those requires either a very long series or an experiment.
The deeper limit is that the underlying systems change without telling you. A model version ships, a retrieval index refreshes, a partnership adds a new source, and your series has a step change in it that has nothing to do with your corpus. Record the model identifier returned with every response, and when the identifier changes, mark the break in the chart. A series with unmarked breaks is worse than no series, because somebody will draw a trend line through it.
Pick ten questions a buyer in your vertical would genuinely type, run them once against one engine through its API, and count how many of the returned sources are yours. Ten questions in an hour will tell whoever owns the editorial operation whether the answer is near zero or near a third, and that alone decides whether this is worth building properly.
Questions people ask about share of citation measurement
- What is share of citation?
- It is the proportion of sources returned by an answer engine, across a defined set of questions, that point at your own properties. Ask forty questions, record every source the engine attaches to each answer, count the total and count yours. Nineteen of 312 returned sources is a share of citation of 6.1 per cent.
- Should share of citation be counted per source or per question?
- Both are legitimate and they give very different numbers from the same run. Counting sources gives the share of the citation list you occupy. Counting questions gives the share of answers you appear in at all. On one run these came out at 6.1 per cent and 27.5 per cent, so the report has to name which one it is using.
- How do you get the source list out of an answer engine?
- Through an API that returns it as structured data. Perplexity's chat completions response documents a citations array holding source URLs and a search_results array whose objects carry title, url, snippet and date. Reading source lists off a screenshot works for one question and stops being practical somewhere around the tenth.