Mentions vs citations in AI search, and why the difference matters
Most AI visibility tools report one number. There are two, they mean different things, and conflating them inflates every metric on the dashboard.
A mention is your name appearing in an AI answer. A citation is that answer pointing at your URL. They are not the same thing, they are not worth the same, and a tool that reports both as "citations" is telling you a flattering story rather than a true one.
The distinction, concretely
Ask an assistant "best waterproof trail running shoes" and you might get:
For wet, technical trails, the Summit Trail Runner GTX is a common recommendation — it uses a Gore-Tex membrane and includes a rock plate. Reviewers at outdoorgearlab.com rate it well for grip.
Sources: outdoorgearlab.com, runnersworld.com
If you are Summit, this answer contains a mention — your product is named, in a favourable context — and no citation. The reader learns your product exists. They then click through to a review site, which may well link to a competitor's store.
Now the same answer with your domain in the provider's source list. That is a citation: a path back to you and evidence that the grounded response associated the URL with its answer. OpenAI returns URL-citation annotations for web-search responses, Gemini returns grounding citations, and Perplexity exposes citations or search_results in its API responses; their exact formats differ. See the provider documentation linked below.
Why conflating them is a problem
Three reasons, in increasing order of seriousness:
- The number is ambiguous. Reporting mentions and citations together makes the headline figure impossible to interpret.
- It hides the actionable gap. "Mentioned but never cited" is a specific, fixable diagnosis: the model knows your product but does not consider your page a source worth pointing at. That usually means thin content, missing structured data, or crawl access problems. Merged into one number, that diagnosis disappears.
- It cannot be audited. A citation is verifiable — the URL is in the response. A mention is a string match, and string matching is where false positives live.
Where the false positives come from
Naive mention detection is genuinely hard to get right:
- Substring matching. A brand called "Ace" matches "surface", "placement" and "space". Whole-word matching with Unicode boundaries is the minimum bar.
- Short names. Two-character brands should not be matched at all; the noise floor is higher than the signal.
- Negative context. "Unlike the Trail Runner GTX, which runs narrow…" is a mention, but not the one you would want counted as a win without seeing it.
The last one is why evidence matters more than the count. Any mention worth reporting should come with the sentence that produced it, so you can judge it yourself.
Domain matching is its own trap
Deciding whether a cited URL is yours is less obvious than it sounds. All of these should resolve to the same identity:
https://www.example.com/product/x?ref=ai
example.com
shop.example.com
https://example.com/
And these should not:
example.com.evil.net
notexample.com
Getting this wrong in the permissive direction inflates your numbers. Getting it wrong in the strict direction hides real wins — a citation of shop.example.com scored as a miss because the store domain is example.com.
Multi-part public suffixes make it worse: example.co.uk has three labels but only two of them are yours. Domain comparison should use the maintained Public Suffix List rather than splitting on dots.
How we score it
We store mentions and citations separately, each with its evidence, and weight them differently:
visibility = 100 × ( 0.7 × citedPrompts / trackedPrompts
+ 0.3 × appearingPrompts / trackedPrompts )
Citations carry 70% of the weight because they are verifiable and clickable. Mentions carry 30% because presence in the answer still has real value — a shopper who reads your product name will search for it.
The important property is that both terms are derived from stored responses. Every point in that score traces back to a specific answer from a named provider at a known time, which you can open and read.
What to do with the distinction
Once the two are separated, the diagnosis writes itself:
- Neither mentioned nor cited → the model does not associate your product with this question at all. Check crawl access first, then whether the product data is complete enough to be retrieved.
- Mentioned, not cited → it knows the product but does not rate your page as a source. Usually thin content or missing structured data.
- Cited, not mentioned → your page was used as evidence for a general answer. Often a category or guide page doing the work rather than the product page.
- Both → what you are aiming for.
Each of those needs a different fix, which is the entire argument for measuring them apart.
Primary sources
See what AI systems say about your products
Connect WooCommerce and get your AI readiness audit, buyer prompts and first recommendations on day one.
Start free AI audit