Citation share
Citation share is a brand's share of the citations in AI assistant answers on a given topic. When people ask about choosing an agency, a product or a process, some sources get cited repeatedly and others never. Citation share says how much of that citation surface you hold. It is the share-of-voice equivalent for an environment where a model answers instead of a list of links.
In short
| What it is | Your citations as a share of all citations in AI answers on a topic |
| How it differs from a ranking | A ranking is an order in a list. A citation is presence inside the answer the user actually reads |
| Where to measure it | Bing Webmaster Tools (AI Performance), manual query testing, third-party tools with their own query samples |
| What raises it | Original data, answers up front, crawler availability, consistent brand facts |
How citation share is measured
A prompt set instead of keywords
Measurement starts with a fixed set of prompts. Not brand queries, but the purchase questions people put to an assistant: which product to choose for a given situation, how two options differ, what something costs, when a solution is worth paying for. They are written the way people actually address an assistant: full sentences with context, in the language of the market. Usually that means dozens of prompts grouped by topic or decision stage. The set stays the same between measurements: once it changes, the result changes with it and the trend stops being comparable.
Runs, recording and calculation
- Each prompt is asked in every tracked assistant (ChatGPT with search, Perplexity, Gemini, Copilot, Google AI Mode) and several times in a row, because the same prompt returns a slightly different answer with different sources each time.
- The cited domains are recorded from every answer. A domain counts once per answer even if it is linked in several places. A brand mention without a link is noted separately.
- Citation share = citations of your domain divided by all citations in the sample. Coverage is tracked alongside: the percentage of answers where the domain appeared at least once.
- The same is calculated for competitors and read as share of voice: the denominator is shared, so your share drops even when your citation count stays flat and a competitor gains.
- The result is broken down by assistant and by topic. One assistant may cite mostly you and another not at all; the total hides that.
Worked example: 40 prompts, 3 assistants, 3 runs, so 360 answers. They contain 1,800 citations, 5 per answer on average. Your domain has 90 citations, a citation share of 5%. A competitor has 180 citations, so 10%. Your domain appeared in 90 of the 360 answers, coverage 25%.
Why it is a sample, not a census
Assistants do not publish what users ask them, and an answer depends on location, language, conversation history and model version. The number is therefore an estimate over a chosen sample, not a count of what all users actually see. Three rules follow. Small differences between months sit within noise; the smaller the sample, the larger a shift must be to mean anything. Only results from the same method are comparable: the same set, the same assistants, the same number of runs. And outputs of two tools with different samples are not compared with each other, only each with itself over time. The exception is the Bing Webmaster Tools report, which counts real Copilot citations, but for one assistant and without a competitor denominator, so no share can be calculated from it.
Why it is the metric replacing rankings
Classic search was a contest for the order of links, measured in positions and clicks. In AI answers the links recede, and with them the meaning of a position. What remains is the question of whether the answer carries your brand and your numbers. A company with a high citation share is present at the moment of decision even when the user clicks nothing at all.
From our own practice: a baseline and monthly tracking
The approach that works: in Bing Webmaster Tools, the only major tool that reports AI citations directly from the operator, set a baseline and track it monthly. Complement it with a manual test of a query sample, recording whether the brand is mentioned and in what wording.
The important part is reading it right: the absolute number matters less than the trend and the ratio against competitors. Citations take months to build, and the main lever is original data the model cannot find anywhere else.
Common mistakes
- Demanding one precise number. Every tool measures on a different query sample. The trend is usable; the absolute value is not.
- Measuring without a baseline. Without one you cannot tell whether the work is paying off. Step one is measuring today's state.
- Ignoring the wording of the citation. A mention with wrong numbers or a stale description can hurt more than absence.
Related terms
See also AI visibility, GEO, AEO, zero-click search and RAG.
Frequently asked questions
How do I raise citation share?
Publish data nobody else has, put answers at the top of pages, keep the site available to crawlers and the brand facts consistent. The model picks sources by their usability.
How often should it be measured?
Monthly. The numbers are small and measuring more often only adds noise.
Can a citation be bought?
No. The model selects sources on content and availability. Citation surface cannot be paid for; ad formats inside AI answers exist, but they are labeled as ads and are not citations.
How we can help
We measure citations and build citation surface as part of our AI visibility agency service.