Here is a pattern that confuses a lot of teams. You test a category prompt — "best [your category] for [your buyer]" — in ChatGPT or Perplexity, and your brand actually appears. Then you look at the sources the answer links to, and none of them are yours. The assistant learned about you from a review site, a directory, or a Reddit thread. Meanwhile a competitor is mentioned and cited: their comparison page, their profile on the category's top review platform, an editorial roundup that features them.
Both of you were "visible." Only one of you controls the material the answer was built from. That difference is what citation share measures — and it is why two brands with similar mention counts can be in completely different competitive positions.
Most marketers already know "share of voice." This article defines the citation-side metrics precisely, shows the arithmetic with a worked example, and explains why, when you need to fix AI visibility rather than just report on it, citation share is usually the number to start from.
The five metrics, defined
All five metrics come from the same raw material: a fixed set of buyer prompts, run across one or more AI providers, ideally more than once. Each run produces answers, and each answer contains (a) text that may mention or recommend brands, and (b) citations — the source URLs the answer links to or was grounded in.
For the worked examples, assume one benchmark run: 20 prompts × 4 providers × 3 repeats = 240 answers, which together surface 620 citation URLs.
Mention rate
The percentage of answers that name your brand at all. If your brand appears in 72 of the 240 answers, your mention rate is 30%. It answers the most basic question — does AI know we exist in this context? — and nothing more. A mention can be positive, negative, or a passing aside in a list of eight.
Recommendation rate
The percentage of answers that actively recommend your brand — a shortlist placement, a "best for X" designation, a direct "I'd suggest…". If 36 of the 240 answers recommend you, your recommendation rate is 15%. This is stricter than mention rate and closer to commercial reality: buyers act on the shortlist, not the honorable mentions.
Citation rate
The percentage of answers that cite at least one page you own. If 24 of the 240 answers link to your website, your citation rate is 10%. This tells you how often your own content is good enough — clear enough, specific enough, crawlable enough — for an AI system to use it as evidence.
Citation share
Of all the citations across the whole run, the percentage that point to sources you own. If 50 of the 620 cited URLs are yours, your citation share is 8%. Where citation rate is per-answer ("did they cite us?"), citation share is per-citation ("how much of the total evidence pool is ours?"). It is the market-share view of trust: the same calculation works for each competitor and for each third-party domain, which is what makes it a comparative, diagnostic metric rather than just a brand health number.
Source type share
The distribution of all citations by type of source — owned sites, review platforms, directories, community threads, editorial coverage, marketplaces. In our example run: review sites 30%, directories 19%, community 15%, editorial 12%, owned sites (all brands combined) 24%. Source type share describes the terrain: it tells you where the trust in your category actually lives, before you ask how much of it you hold.
Why citation share matters more than it sounds
AI answers are not assembled from brand reputations floating in the ether. When an assistant answers a commercial question, it retrieves and weighs sources — specific pages on specific domains — and composes its recommendation from what those sources say. Google is explicit that its AI features build on the same Search systems that crawl, index, and evaluate pages (see Google's guidance on AI features and your website), and answer engines like Perplexity expose their source lists directly. The answer chooses sources first; brands get mentioned as a consequence.
That has a blunt competitive implication: a competitor does not outrank you in AI search — their sources outnumber yours. When ChatGPT consistently recommends a rival, it is usually because third parties validate that rival: they hold the top profile on the review platform the answers keep citing, they appear in the comparison articles, they are named in the community threads. The assistant is not playing favorites. It is reflecting the evidence pool — and the evidence pool is measurable.
This is also why citation share is more actionable than mention share. A low mention rate tells you that you have a problem. A citation breakdown tells you what the problem is made of: which domains dominate the answers, which of them you are absent from, and which competitor owns each one. Mentions are the scoreboard; citations are the game film.
- Mention rate 30% — the name comes up
- Citation share 2% — almost no owned sources in the pool
- Answers describe the brand from third-party summaries
- No control over accuracy, pricing, or positioning in answers
- Visibility depends entirely on sources others control
- Mention rate 28% — roughly the same headline number
- Citation share 14% — owned pages appear in the evidence pool
- Answers quote the brand's own descriptions and data
- Strong presence on the review sites the answers also cite
- Errors are fixable: the cited pages are theirs to update
Owned vs third-party citations: the six source types
Every citation in the pool belongs to someone. The first tagging pass is ownership: owned (your domains), competitor-owned, or neutral third party. The second pass is type — because the fix for a missing directory listing is nothing like the fix for absent editorial coverage.
The shape of the split matters as much as the total. Near-zero owned citations means your site is not citable — the content is vague, unstructured, or invisible to crawlers — and every answer about you is secondhand. Near-zero third-party citations is the opposite failure: you are the only one talking about yourself, and assistants weight independent corroboration precisely because you cannot manufacture it. And when one competitor dominates a single source type — say, 60% of the review-site citations — you have found the specific battleground, not just the score.
How to calculate citation share, step by step
The calculation is simple; the discipline is in the sampling. Here is the full procedure.
- 1. Fix a prompt set. 15–25 prompts in real buyer language: category prompts ("best X for Y"), problem prompts, comparison prompts, local prompts if relevant. The set stays constant between runs so the numbers are comparable. (Building this set is step one of the broader AI visibility audit checklist.)
- 2. Run across providers, repeatedly. ChatGPT, Perplexity, Gemini, Google AI as relevant — providers cite differently, and answers vary between runs, so a single scan is a screenshot, not a measurement. Two or three repeats per prompt per provider is a workable floor.
- 3. Extract every citation URL. Every source link from every answer goes into one table: prompt, provider, run, answer, cited URL, domain.
- 4. Tag each domain twice. Ownership — yours, a named competitor's, or neutral — and source type from the six above. Discard citations irrelevant to the category (generic dictionary or news-of-the-day links) so the denominator stays honest.
- 5. Compute the shares. Your citation share = your owned citations ÷ total relevant citations. Then compute the same number for each competitor, and the source type share for the whole pool.
- 6. Build the competitor dominance table. For each major source type, which brand's pages — or profiles, or reviews — hold the citations? This is where the fix list comes from.
Concretely, with our 620-citation example: you hold 50 citations (8% citation share), Competitor A holds 96 (15%), and Competitor A alone accounts for over half of all review-site citations. That last line is the actionable one. Your problem is not "AI visibility" in the abstract; it is one competitor's dominance of one source type that the answers in your category keep leaning on.
Steps 3 through 6 are easier to see than to describe. Roll the same 620 citations up to the domain level and the run stops being a score and becomes a ledger — every row is a domain the answers actually leaned on, with your 50 and Competitor A's 96 sitting in plain sight among the third parties:
| Domain | Source type | Citations | Share of 620 |
|---|---|---|---|
| yelp.com | Reviews | 118 | 19.0% |
| competitor.com | Owned — Competitor A | 96 | 15.5% |
| reddit.com | Community | 64 | 10.3% |
| trustpilot.com | Reviews | 57 | 9.2% |
| yellowpages.com | Directories | 55 | 8.9% |
| yourbrand.com | Owned — you | 50 | 8.1% |
| industryweekly.com | Editorial | 41 | 6.6% |
| Everything else (long tail) | Mixed — remaining reviews, directories, community, editorial | 139 | 22.4% |
| Total relevant citations | 620 | 100% |
How to improve your citation share
Because the metric decomposes by source type, so does the work. The breakdown tells you which of these to do first.
One boundary worth stating plainly: none of this works as manipulation. Google's guidance on optimizing for generative AI features warns against special "AI hacks" and inauthentic mentions, and every review platform and community polices fakes. Citation share is a trust metric; faking the inputs targets the one thing the metric exists to measure. The durable path is the boring one — be genuinely present, accurate, and useful on the sources AI already trusts.
No prompt set, provider, or level of source work guarantees a mention or a "ranking" in ChatGPT or Google's AI features — anyone promising that is selling something. What citation share gives you is the next best thing: a measurable, comparable, fixable account of where the trust behind your category's answers actually comes from.
FAQ
What is AI citation share?
AI citation share is the percentage of all citations in a set of AI answers that point to sources you own or control. You measure it by running a fixed set of buyer prompts across AI providers, collecting every cited URL, tagging each domain by owner and source type, and dividing your brand-owned citations by the total relevant citations. It shows how much of the evidence behind AI answers in your category comes from you versus from third parties or competitors.
How is citation share different from mention rate or AI share of voice?
Mention rate and share of voice count how often your brand name appears in AI answers — they measure the outcome. Citation share counts which source URLs the answers are grounded in — it measures the cause. Two brands can have similar mention rates while one holds most of the citations that shape how the category is described. When you need to fix visibility, citation share tells you which sources to work on; mention rate only tells you that something is wrong.
How many prompts and runs do you need to measure citation share reliably?
Enough that individual answer variation stops dominating the number. A practical baseline is 15 to 25 real buyer prompts, run across at least two or three providers, repeated two or three times — a few hundred answers producing several hundred citation URLs. A single run of a single prompt is a screenshot, not a measurement: AI answers vary between runs, so citation share is only meaningful as a rate over a repeatable sample, tracked on the same prompt set over time.
Can I increase my citation share without spamming?
Yes — and spamming is the one approach that reliably backfires. Citation share grows when the sources AI already trusts have accurate, substantive material about you: clear answer-shaped pages on your own site, complete and correct profiles on the review sites and directories your category's answers cite, genuine participation in community threads, and earned editorial coverage. Fake reviews, planted forum mentions, and mass-generated pages violate platform rules and Google's guidance, and they target trust signals precisely because those signals are hard to fake.