AI Mentions vs. Citations: One Gets You Recommended. The Other Gets a Footnote.
A firm can be named without its website being cited. A website can be cited without the firm being recommended. Collapse those events into one “visibility” score and the dashboard becomes impressively precise about the wrong thing.
The name and the footnote do different jobs
Suppose an assistant answers a buyer with three firms, explains why each might fit, and links to two directory pages plus a bar association article. Your firm can appear in that answer even when your domain appears nowhere in the sources. That is a mention without an owned citation.
The reverse also happens. Your article about a filing deadline may support one sentence in the explanation while the assistant recommends somebody else. That is an owned citation without a recommendation. It may send a visit and build informational authority. It did not win the shortlist.
Neither outcome is fake. They are simply different stages of influence, and a useful measurement system refuses to make them roommates in one vague percentage.
Track the answer as a small chain of events
For each prompt and run, record at least four things: whether the brand was named, how prominently it was named, which URLs were cited, and which claim each citation appears to support. Add sentiment or recommendation strength when the distinction matters. “Consider Firm A” is different from “Firm A exists,” and both differ from “Firm A has faced complaints.”
A practical funnel is retrieval, citation, mention, recommendation, action. Most public tools can observe only the middle three, and even there the plumbing is partly hidden. A cited URL proves that the system displayed the URL; it does not prove the page alone caused the sentence. A mention proves output presence; it does not prove a click, inquiry, or retained client.
This is why “AI traffic” and “AI visibility” should not be used as synonyms. Referral analytics observe visits that arrive with a detectable source. Visibility measurement observes answers, including the many answers that produce no click at all.
Citation quality is not binary
Academic work on citation-aware generation separates answer quality from citation quality for good reason. A link can be present yet fail to support the nearby claim. It can support only half the sentence, point to an outdated page, or attribute an assertion to a source that merely repeats it.
For commercial monitoring, sample citations manually. Open the page. Check whether it supports the named claim, whether the claim concerns the correct business, and whether the source is primary or derivative. Entity mistakes matter here: a citation to a similarly named firm is not “close enough.” It is a measurement error and potentially a reputational problem.
Call this citation correctness or entailment, not authority. Authority is a judgment about the source. Correctness asks a narrower question: does the cited material actually back the claim attached to it?
Owned, earned, and directory citations tell different stories
Group cited domains by relationship to the business. Owned sources include the firm’s site and controlled profiles. Earned sources include editorial coverage, professional bodies, research, and independent reviews. Directories sit in a useful middle bucket because their editorial standards and business models vary wildly.
An owned citation tells you the assistant can retrieve your evidence. An earned citation tells you somebody else’s page is helping establish the claim. A directory citation may be doing identity resolution, local discovery, reputation work, or all three. The category is less important than preserving enough detail to inspect what happened.
Do not translate this into “get mentioned everywhere.” Ten cloned profiles do not equal one credible source that supplies a missing fact. Source strategy should follow the evidence gap: identity, specialization, geography, proof, or comparison.
Use a scorecard, not a smoothie
Report recommendation rate: the percentage of eligible runs in which the brand is actually offered as an option. Report mention rate separately when neutral or negative appearances are in scope. Report owned citation rate and earned citation rate as their own series. Preserve per-model results and raw answers.
Then add a small citation audit: a sample of links checked for support, freshness, entity match, and claim type. This catches the wonderfully awkward case where a chart climbs because assistants are citing a page that contradicts the brand.
If leadership insists on one headline number, make it clearly derivative and keep the component metrics beside it. Composite scores hide trade-offs. A firm with broad recommendation presence but weak owned citations has a different job from a publisher with heavy citations and no commercial mentions.
Choose the fix that matches the missing event
No mentions and no citations suggests a discovery or authority problem. Mentions without owned citations suggest the market knows the name but the site is not supplying reusable proof. Owned citations without recommendations suggest useful information that does not establish fit, differentiation, or trust. Recommendations supported by inaccurate third-party claims require correction, not more content.
This is the practical payoff of fussy definitions. You stop prescribing a blog post for every illness. Sometimes the missing asset is a service page. Sometimes it is a professional listing, a case result, an accessible page, a corrected entity record, or evidence another credible source is willing to publish.
Key takeaways
- A brand mention, a recommendation, and a source citation are separate observable events.
- A displayed citation can be wrong or incomplete; audit whether the page supports the attached claim.
- Separate owned, earned, and directory sources so the evidence gap remains visible.
- Use component metrics to choose the fix instead of blending everything into one visibility score.
Sources and further reading
Primary documentation and research used for this field note. Product behavior changes; check the linked source before treating any implementation detail as permanent.
- 1. Enabling Large Language Models to Generate Text with Citations — EMNLP 2023 / ACL Anthology
- 2. Ground Every Sentence: Improving Retrieval-Augmented LLMs with Interleaved Reference-Claim Generation — NAACL 2025 / ACL Anthology
- 3. ChatGPT Search: viewing citations and sources — OpenAI
Next step
Atlas shows the public map. A Viclaro audit turns that map into the prompts your firm is losing and the page edits most likely to change the next scan.
See the public recommendation map.