AI mentions and AI citations are not the same metric

Ask an AI assistant “what are the best project management tools for a small agency,” and two different things happen to the brands involved. Some get named in the sentences. Some get linked in the sources underneath. The overlap between those two groups is smaller than most people expect.

Tools that report a single blended “visibility score” average those two groups together. That average is the least useful number in the report, because the gap between the two is the part that tells you what to do.

The two definitions

A mention is the brand name, or a configured alias for it, appearing in the visible answer text. It is what the buyer reads. It is the signal that maps to recommendation.

A citation is a domain the brand owns appearing in the answer’s source list. It is what the model drew on, or at least what it attributed. It is the signal that maps to source authority.

Both are observable. Neither is inferred from the other.

The four states, and what each one means

State	What it looks like	What it usually means	What to do
Mentioned and cited	Named in the answer, your domain in the sources	Working. The model knows you and uses your material	Protect it. Watch for competitors closing the gap
Mentioned, not cited	Named in the answer, sources are third parties	Your reputation travels through other people’s pages	Find which third parties are cited and get the facts right there
Cited, not mentioned	Your domain is a source, but a competitor is the recommendation	Your content is useful, your brand is not the answer	Your material answers the question without positioning you as the option
Neither	Absent from both	No presence on this question	Decide whether the question is worth competing for at all

The third row is the one that surprises people. Being a cited source while a competitor is the named recommendation is a common and specific failure. It usually means the page is written to explain a topic rather than to establish the brand as a candidate answer to it. Publishing more explanatory content makes it worse, not better.

The second row is the most common for brands with strong PR and weak technical content. The model recommends you based on what a review site, a Reddit thread, or a listicle said. That is real visibility, and it is fragile, because you do not control the source. The response is outreach to the specific domains being cited, not more content on your own site.

Why they have to be measured separately

Three reasons, in increasing order of importance.

They move independently. A brand can gain mentions across a whole prompt set while its citation rate stays flat, or the reverse. Averaging them produces a number that goes sideways while both underlying signals are moving.

They have different fixes. Mentions respond to reputation, category association, and third-party coverage. Citations respond to crawlable, substantive, well-structured pages on domains you own. Spending content budget on the wrong one is the most common waste in this category.

They fail differently. Citation detection is a URL-matching problem with well-known artifacts. Mention detection is a text-matching problem with entirely different ones. A blended score hides which half of the pipeline is wrong.

Where each measurement goes wrong

Both are harder to get right than they look, and the errors are systematic rather than random.

Mention detection

The trap is aliases. Brands want their product names, abbreviations, and common misspellings to count, which is correct. But a brand whose name is also an ordinary word turns every unrelated use of that word into a mention. The fix is case-sensitive matching for those specific aliases, applied deliberately rather than as a global setting.

The second trap is where in the answer the match occurred. Some surfaces append a sponsored unit after the organic answer. An advertiser named there is a paid placement, not an organic recommendation, and counting it inflates the number that matters most. Cutting that section requires anchoring on its actual shape and warning when the shape changes, rather than trimming a fixed number of characters and hoping.

Citation detection

The trap is what counts as a source URL. Several surfaces put a full URL inside a citation’s domain field, which produces a bare site root that duplicates a deeper link from the same origin. Counting both doubles the credit. Favicon and thumbnail service assets arrive in icon fields and are not sources at all.

Neither artifact is exotic. Both are common enough that a tool which does not filter them will report a citation rate materially higher than reality, and the inflation is not uniform across surfaces, so cross-surface comparison breaks too.

Making the distinction checkable

Whatever tool you use, the test is the same: click a mention and see the answer text with the match highlighted, click a citation and see the source URL it came from. If the interface cannot show you that, the distinction it reports is a claim rather than a measurement.

In refd, mentions and citations are stored as separate signals on every result, the highlighting in the answer view runs the same matcher the scorer runs, so a highlight appears if and only if that entity was scored as mentioned, and the raw provider response sits behind every score. The scoring contract is published, and because raw answers are retained, a corrected scorer can be replayed over historical data rather than leaving old numbers wrong.

The one-line version for a leadership deck

Mentions are whether the assistant recommends you. Citations are whether it reads you. You need both numbers, you need them separately, and you need to be able to click either one and see the sentence it came from.