How citation share is calculated
You fix a list of prompts a customer might type, run each of them through the assistants you care about, and record which domains appear in the source list of every answer. Citation share is your appearances divided by the total appearances measured, or by the number of answers in the run — trackers differ, which is why the definition has to be written down before the first measurement.
Everything therefore rests on the query set. A list weighted towards prompts that name your brand will produce a flattering figure; a list of broad category prompts will produce a bleak one. Neither is wrong, but they are not comparable, and mixing them between runs turns the trend line into noise.
Why citation share matters
Rankings assume a page of ten blue links where position is the currency. In an assistant’s answer there is no position to win, only a short list of sources the model chose to lean on. Citation share is the closest thing to a rank tracker for that surface: it tells you whether you are in the pool of sources being drawn from, and which competitors are in it with you.
It is most valuable read as a comparison, not a score. If a rival appears across most of a category’s prompts and you appear on the few that name you, the gap is in the explanatory content, not the brand. That is a content brief, not a mystery, and it is the kind of finding that makes an AI visibility audit worth commissioning.
Where citation share goes wrong
The first trap is treating a single run as a fact. Assistants re-retrieve on every request, personalise by history and location, and change models without notice, so the same prompt can cite different domains on the same morning. A share measured once is a snapshot of one moment, not a position.
The second trap is chasing citations on prompts nobody sends. It is easy to assemble a query set that flatters your existing content, watch the number rise, and gain nothing, because the prompts were invented in a meeting rather than drawn from real customer questions. The third is quietly editing the list — adding prompts you now rank for, dropping ones you never win — which guarantees an improving chart and a meaningless one.
How to act on it
Write the prompt list from things people actually ask you: enquiry emails, sales calls, search queries in Search Console, questions your support inbox repeats. Freeze it, note the assistants, the location and the date, and re-run on a schedule rather than whenever the number is likely to look good. Track the same competitors in the same run so the comparison stays fair.
When you find prompts where you are absent, look at what the cited pages do that yours does not: answer the exact question early, use the vocabulary of the prompt, and stand alone without the rest of the site. Fixing that is ordinary content work. Pair the measurement with prompt volume estimates so you spend the effort on questions that get asked often, not merely on the ones you are losing.