Mentions vs citations: the two currencies of AI visibility

Ask an assistant for the best CRM for a five-person agency and read the answer twice. The first read is for names: which vendors get spoken about in the body of the answer. The second read is for links: which websites the answer lists as its sources. Those are the two currencies of AI visibility, and they move independently. A brand can be named in an answer whose sources are entirely review sites and forum threads. A publisher can be linked as a source for an answer that never mentions its own name.

The industry increasingly measures them separately. Ahrefs Brand Radar, for instance, explicitly splits mentions, where your brand is named, from citations, where your site is linked, and Reachroller applies the same split with a stricter rule: a mention counts only when the brand name literally appears in the stored answer text, so neither currency can be padded with generosity.

Conflating the two produces bad strategy in both directions. Teams that only chase citations treat AI visibility as a traffic problem and miss the shortlist decisions that happen inside the answer. Teams that only count mentions celebrate awareness while rivals collect the clicks and the conversion data. The rest of this article takes each currency apart: where it comes from, what the studies say about earning it, and which one pays in your situation.

A mention is the assistant deciding your name belongs in the answer. That decision can be made without your website ever being touched. Models learn brands from training data, where your name appears in whatever corpus text discussed you, and search-backed answers learn them from third-party pages: review sites, comparison articles, community threads, reference entries. Your own domain is just one voice among many, and often the least trusted one for a recommendation question.

That is why the mention currency is earned mostly off your own property. An engine that repeatedly reads your brand described as a credible option for a specific problem, across independent sources, starts naming you for that problem. It is also why mentions are slower to build and more durable once earned: no single page grants them, and no single page removal revokes them.

The business weight of a mention comes from where it lands: inside the answer, at the moment of shortlisting, in journeys that often end without any click at all. G2 found 69 percent of B2B buyers chose a different vendor than they originally expected because of AI chatbot output, and 33 percent bought from a brand they had never heard of before an AI named it. Those outcomes ran through mentions, and mostly through answers nobody clicked out of. How to win those clickless moments is the subject of marketing when nobody clicks.

A citation has a harder prerequisite: your page personally participated in the answer. The retrieval pipeline found it in an index, judged it relevant to the question, extracted passages from it, and listed it as a source. Every link in that chain can fail. A page that is not indexed by Google and Bing does not exist for most engines. An indexed page that answers no specific question gets retrieved for nothing. A retrieved page whose claims are vague gets outquoted by a denser rival.

Citation behavior also differs sharply by engine. Perplexity averages about 8.2 sources per answer, roughly 3.4 times ChatGPT's density, which makes it the widest door for new publishers, and its path is mapped in how to get cited by Perplexity. Meanwhile cross-platform analyses, including work by Profound, found only around 11 percent of domains are cited by both ChatGPT and Perplexity. A page that wins one engine has made no promise about the others, so citation strategy is engine-by-engine work.

The payoff for clearing the bar is the referral: a citation is the only currency that sends a measurable visitor. It also pays a second, quieter dividend. Being the cited source puts your framing, your numbers and your vocabulary inside the answer itself, which shapes how the category gets described even for buyers who never click through.

The citation studies agree on an uncomfortable shape. 5W Research found Wikipedia at 13.15 percent and Reddit at 11.97 percent of ChatGPT's U.S. citations, together over a quarter of the total, while the Wall Street Journal, New York Times and Bloomberg miss the top 20 entirely. Otterly's 2026 AI Citations Report, built on more than a million data points, confirms the same Wikipedia and Reddit dominance with a long fragmented tail. Reddit's share grew roughly 73 percent in commercial categories across 2025 and 2026. Reference content and community discussion outcite prestige media, and the full source-by-source picture is in what ChatGPT actually cites.

The long tail is the opportunity hiding in those numbers. After the giant reference domains, citations fragment across thousands of specific pages that answered specific questions well. You will not outcite Wikipedia for "what is a CRM", and you do not need to. You need to be the best indexed answer to the forty precise questions your buyers ask, which is a winnable contest for a small team.

On page structure, the evidence is honestly mixed. SE Ranking found about 71 percent of pages cited by ChatGPT carry structured data, a correlation with no proven causation, while Ahrefs' May 2026 study of 1,885 already-cited pages found adding schema produced no measurable citation lift on those pages and a statistically significant decline for AI Overviews citations. The safe reading: structure will not rescue weak content, and evidence density is the variable with experimental support behind it.

Mentions move revenue by moving shortlists. Forrester's 2026 Buyers' Journey Survey found 94 percent of buyers used AI during their most recent purchase and 55 percent compared vendors inside AI tools, and the G2 numbers above show those comparisons changing outcomes for 69 percent of buyers. None of that influence registers in your analytics, because the deciding moment produced no visit. If your sales team hears "we saw you recommended by ChatGPT", that was the mention currency paying out.

Citations move revenue through a smaller, hotter pipe. AI referrals are still around 0.18 percent of all sessions in some panels, and a meaningful share of AI-sourced traffic hides in direct. But the visitors who do arrive convert unusually well: WebFX measured about 1.2 times organic across 2.3 billion sessions, Adobe found AI traffic converting 42 percent better in March 2026, and Semrush reported roughly 4.4 times standard organic. Ahrefs found up to 23 times, a figure widely treated as an outlier and best cited with that caveat. The studies disagree on magnitude and agree on direction: this traffic is scarce and valuable.

So the honest allocation depends on your motion. Sales-led B2B brands should weight mentions first, because their revenue is decided in shortlists rather than sessions. Self-serve and content businesses should weight citations first, because the referral click is their conversion event. Almost everyone ends up needing both, which is the argument for measuring them as separate lines from day one rather than discovering the distinction in a quarterly review.

Mentions are earned in other people's text. The playbook is third-party presence: genuine participation in the communities engines read, coverage in the niche listicles and comparison articles that answer buying questions, accurate entries on the review platforms your category trusts. This is closer to digital PR than to SEO, and it is slow, compounding work.

Citations are earned on your own pages. The experimental evidence here is unusually direct: the Princeton-led GEO study at KDD 2024 found adding quotations, statistics and cited sources boosted a page's visibility in generative engine responses by up to roughly 40 percent, with the best methods improving about 22 percent on position-adjusted word count and 37 percent on subjective impression, while keyword stuffing landed near the bottom. Publish pages that answer one question each, load them with checkable evidence, get them indexed, and give the engines a week or two.

The two playbooks reinforce each other. The comparison page that earns a citation also teaches engines your name for future mentions; the Reddit thread that earns a mention often becomes a cited source itself. Reachroller's fix loop automates the citation side: every lost question generates a publish-ready page with slug, title tag, meta description, schema markup and indexing steps, and a recheck confirms whether the answer flipped. The mention side stays human work, and digital PR for AI visibility is the map for it.

Because the two currencies move independently, every answer to a tracked question lands in one of four states, and each state is a different diagnosis. Mentioned and cited is the strongest outcome: the engine recommends you and grounds the recommendation in your own page, so your framing controls the narrative and the referral click has somewhere to land. Treat questions in this state as assets to defend, and recheck them on a schedule, because answers churn even when nothing about the web changes.

Mentioned but never cited means the engine learned your name from third parties or from training data while your own pages contribute nothing. It is a good problem with an obvious next move: publish a citable page for the exact question, because you are already past the harder hurdle of being considered category-relevant. Cited but never mentioned is the stranger state: your content is feeding answers that recommend other people, common for glossary pages and research posts that explain a topic without connecting it to your product. The fix happens inside the cited page, tightening how your brand is positioned next to the facts the engine keeps borrowing.

Neither mentioned nor cited is the median starting state, the one G2's 51 percent zero-citation figure describes. The sequencing advice for it runs against instinct: chase the citation first, because it is the currency with a two-week feedback loop, and let the early citations begin teaching the engine your name for the slower mention game. A quarter of disciplined publish-and-recheck cycles moves questions out of this state faster than any amount of brand advertising aimed at the model's memory.

Both currencies inherit the measurement problems of AI answers generally: responses vary run to run, so single checks are anecdotes, and questions containing your brand name produce mentions by construction. The fixes are the same as for share of voice: a fixed unbranded question set, repeated runs per engine, stored answer text, and strict counting. The full framework is in AI share of voice.

Report the two currencies as separate lines with separate denominators: the share of answers that mention you, and the share of answers that cite your domain. Watching the lines diverge is diagnostic. Mentions without citations means engines know you from third parties and your own pages are not pulling weight. Citations without mentions means your content informs answers that recommend other people, which usually calls for clearer positioning inside the cited pages themselves.

Calibrate against the market before judging the numbers: G2 found 51 percent of B2B tech brands hold zero citations across ChatGPT, Perplexity and Gemini, so any consistent citation presence already clears the median. Reachroller starts every account with exactly this baseline, both currencies measured with evidence attached, from $29 per month with a three-day, 50-credit trial and no card. The scoring rules live on the methodology page, because a metric you cannot audit is a metric you cannot trust.