Cited vs Recommended: The Two AI-Visibility Metrics Teams Keep Confusing
Being cited and being recommended are different outcomes. A citation puts your URL in an AI answer’s source list; a recommendation puts your brand name in the sentence the buyer acts on. Semrush’s June 2026 ghost-citations study found 61.7% of domain appearances were citations where the brand was never named in the answer. Track link-citation share and brand-mention share in separate columns, because only one of them reliably moves pipeline.
Most AI-visibility dashboards roll everything into a single score, which hides the question that matters: when an engine answers “best AI sales agent platform for an Australian mid-market team”, are you the source it leans on, or the brand it names? Separate events, separate causes, very different commercial value.
The two outcomes, defined precisely
A citation is a link: your URL appears in the engine’s source list or as a footnote, because a retrieval step pulled your page in as evidence. A recommendation is prose: the model writes your brand name into the answer as an option worth considering, with or without a link.
You can have either without the other, and both happen constantly. Semrush, working with Kevin Indig’s Growth Memo, published the ghost citations study on 9 June 2026, logging “3,981 domain appearances for 115 prompts run across 14 countries and four leading AI search engines”. The headline finding: “Almost 62% (61.7%) were ghost citations. AI platforms used the page as a source link, but the brand name never appeared in the actual answer.”
The same study splits it the other way: “74.9% of all brand appearances included a citation, but only 38.3% of appearances included a brand mention.” Read that as a warning. Your page can do the evidentiary work in an answer that recommends somebody else. You paid the content cost; a competitor got the buyer.
The inverse is just as common. Semrush’s earlier AI Visibility Study (3 September 2025) reported that “only 6-27% of the most mentioned brands also rank as top sources that AI models cite, depending on industry and platform.” The brands models talk about are largely not the domains models quote.
Why the gap exists
Citations come from retrieval; recommendations come from a blend of retrieval and what the model already holds about your category. Engines also reach for structurally different evidence. Yext analysed “more than 6.8 million citations across 1.6 million responses from Gemini (Google), ChatGPT (OpenAI), and Perplexity” in research published 29 October 2025, finding “52.15% of Gemini citations came from brand-owned websites” whilst for ChatGPT “48.73% of citations came from third-party sites like Yelp, TripAdvisor, and MapQuest.”
That asymmetry explains a lot of confused reporting. On Gemini you may be cited because your own site was retrieved. On ChatGPT you are more often recommended because someone else’s page — a roundup, directory or review thread — was retrieved with your name inside it.
Peec AI’s listicle rank effect study (published 14 May 2026) put numbers on that second path. Analysing “over 5.7 million data points across three distinct industries and eight AI engines”, it reports that “B2B SaaS shows the highest absolute visibility lift for being #1 (+16.5 percentage points (pp))”. For brands ranked first in a frequently AI-cited third-party listicle, it also estimated an improvement in answer position of “1.17 positions earlier” in B2B SaaS. Your rank inside a third party’s article is a lever on your recommendation outcome, and not one you directly control.
The economics are not equivalent
Recommendation is the one with a measured downstream effect. Similarweb’s The Downstream Impact of AI Visibility reports users are “2.5x More likely to visit an AI-recommended brand”, that “55.9% Of AI-influenced traffic arrives via search” rather than as a direct AI referral, and “2x Deeper engagement from AI-influenced visitors”.
That middle figure breaks most attribution models. Most of the value of being recommended surfaces as branded search: someone reads the answer, remembers the name, and searches for it later. Judge AI visibility by referral sessions alone and you are measuring the minority channel, then concluding the whole thing does not work. We covered that analytics gap in measuring AI referral traffic in GA4.
Citation without mention is not worthless. It builds the corpus that later supports being named. But it is a slower asset, and it can subsidise a rival’s answer.
Five metrics, and what each one is honestly worth
| Metric | What it actually proves | How to measure it | How easy to game |
|---|---|---|---|
| Link-citation share | Your page was retrieved and used as evidence. Says nothing about whether you were endorsed. | Across a fixed prompt set, count answers where your domain appears in the source list, divided by total answers. | Moderate. Publishing volume, freshness and parseable structure all shift it without changing buyer perception. |
| Brand mention share | The model named you as an option a buyer could choose. The closest proxy for being in the consideration set. | Same prompt set; count answers naming your brand in the prose, linked or not. Record separately from citations. | Hard. Depends on third-party corpus (reviews, roundups, forums) that you influence indirectly at best. |
| Recommendation rank | Very little. Order within a single answer is largely noise (see below). | Position of your name in the answer’s list, averaged over many repeated runs — never a single poll. | Not so much gameable as unstable. Treat any vendor selling “your rank in AI” with suspicion. |
| Referral sessions | Somebody clicked through from an AI surface. A real but minority slice of AI-driven demand. | Server logs or analytics, filtered by referrer and verified user agent, not UA string alone. | Easy to misread. Bot and scanner traffic inflates it; blocked referrers deflate it. |
| Assisted conversions | AI exposure preceded real pipeline. The only metric a CFO should care about. | Self-reported “how did you hear about us” on forms, plus branded-search volume trend, joined in the CRM. | Hard to game, and hard to measure. Small samples make it noisy for months. |
Why tracking tools conflate them
Partly because a single score sells better, and partly because separating the two means parsing the answer text rather than just the citation payload.
The rank column deserves particular scepticism. In research published 28 January 2026 by SparkToro’s Rand Fishkin and Gumshoe.ai’s Patrick O’Donnell, “600 volunteers ran 12 different prompts through each of the 3 tools a combined 2,961 times” across ChatGPT, Claude and Google’s AI. They found “there’s a <1 in 100 chance that ChatGPT or Google’s AI, if asked 100X, will give you the same list of brands in any two responses”, and on ordering “it’s more like 1 in 1,000 runs before you’d see two lists in the same order.”
Their conclusion is the operating rule: “visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric”, but “any tool that gives a ‘ranking position in AI’ is full of baloney.” Aggregate frequency is measurable; position is not.
The academic literature agrees. Ronald Sielinski’s Quantifying Uncertainty in AI Visibility (arXiv, submitted 9 March 2026, revised June 2026) argues that “single-run visibility metrics provide a misleadingly precise picture of domain performance in generative search”, and that such metrics are sample estimators, not fixed readings.
Adding the second column to a panel you already run
The panel mechanics — how many prompts, which engines, how many repeat runs, how to freeze the wording — are already set out in our post on share of model, the AI-search metric replacing rank position, and there is no point restating them. Assume you have that running. This post is about one change to the logging sheet.
- Split the appearance flag into two fields. Most sheets carry a single “did we show up” column. Replace it with domain cited (your URL in the source list) and brand named (your name in the prose), recorded independently. Never let one imply the other, and never average them into a composite score.
- Log the evidence behind rivals, not just their names. When a competitor is recommended, record which page was cited alongside them. Within a few weeks that column becomes a ranked list of the third-party documents your category’s engines actually retrieve — the only directly actionable output either metric produces.
- Report the two lines side by side, as trends. A six-to-eight-week moving line on each share. Movement of a few points week-on-week sits inside the noise floor Sielinski’s paper describes, in either column.
For the base rates that decide whether citations are even available in your category, see how often AI answers actually cite anyone — in many professional-services prompts, rarely. If it is specifically your category prompts that come back empty, the structural reason is covered in the category citation gap.
Where this sits for an AI sales-agent platform
Zian AI is an autonomous AI sales-agents platform — phone, SMS, email and WhatsApp agents, currently in waitlist and partnership beta. Being straight about our own instrumentation: the weekly category-prompt polling we run still collapses both outcomes into one “did we appear” flag, scoring a prose brand mention and a formal citation as the same event. That is precisely the conflation this post argues against, and splitting the two is on our own list rather than something we have already solved.
Being even-handed: if your brand has no branded-search baseline yet, mention share will read as a flat zero for months and weekly polling will waste your time. Fix the third-party corpus first — get into the roundups and comparison pages engines actually retrieve, then measure. Measurement is not a substitute for being worth recommending.
Frequently asked questions
Is being cited without being mentioned worthless?
No, but it is worth less than teams assume. Semrush’s ghost citations study (9 June 2026) found 61.7% of domain appearances were citations where the brand name never appeared in the answer. Those earn evidence value and sometimes a click, but they do not put you in the consideration set, and may be propping up an answer that recommends a competitor.
Which single number should we report to the board?
Brand mention share across a fixed prompt set, trended over at least six weeks, with citation share reported beside it rather than merged into it. If you need a commercial line, use branded-search volume, because that is where most AI-recommendation value surfaces.
Does our position in the list matter?
Barely, within a single answer. SparkToro and Gumshoe.ai’s January 2026 research estimated “it’s more like 1 in 1,000 runs before you’d see two lists in the same order”, and concluded “any tool that gives a ‘ranking position in AI’ is full of baloney.” Your position inside a cited third-party listicle is a different matter — Peec AI measured a +16.5pp lift for rank 1 in B2B SaaS.
Our AI referral traffic is tiny. Does that mean AI visibility is not working?
Not necessarily. Similarweb’s Downstream Impact of AI Visibility reports that “55.9% Of AI-influenced traffic arrives via search” rather than as a direct AI referral, alongside users being “2.5x More likely to visit an AI-recommended brand”. Referral sessions systematically undercount the channel.
How many times should we run each prompt?
More than once, and ideally five or more per engine. Ronald Sielinski’s arXiv paper on quantifying uncertainty in AI visibility (March 2026, revised June 2026) concludes that “single-run visibility metrics provide a misleadingly precise picture of domain performance in generative search.”
Can we buy our way into being recommended?
Not directly, and anyone selling that should be asked precisely what you are buying. The legitimate route is slow: earn accurate placement in the third-party pages engines retrieve, keep brand facts consistent everywhere, and publish material that stands up as evidence.
Sources
- Semrush and Growth Memo, the ghost citations study, 9 June 2026: 61.7% ghost citations; 74.9% of brand appearances cited vs 38.3% mentioned; 3,981 domain appearances, 115 prompts, 14 countries, four engines.
- Semrush, AI Visibility Study, 3 September 2025: only 6-27% of most-mentioned brands also rank as top cited sources.
- Yext, AI Visibility in 2025, 29 October 2025: 6.8 million citations across 1.6 million responses; Gemini 52.15% brand-owned sites, ChatGPT 48.73% third-party sites.
- Similarweb, The Downstream Impact of AI Visibility: 2.5x visit likelihood, 55.9% of AI-influenced traffic via search, 2x deeper engagement.
- SparkToro and Gumshoe.ai, research on AI recommendation inconsistency, 28 January 2026: 600 volunteers, 12 prompts, 2,961 runs across three tools.
- Peec AI, The Listicle Rank Effect, 14 May 2026: 5.7 million data points, eight engines; +16.5pp lift for rank 1 in B2B SaaS listicles.
- Ronald Sielinski, Quantifying Uncertainty in AI Visibility, arXiv, March 2026 (revised June 2026).
Every figure above was checked against the publishing organisation’s own page on 26 August 2026. Where a number is quoted, the quotation marks mark text that appears verbatim on that page.
Work with us
Zian AI is in waitlist and partnership beta. If you run outbound or support at scale and want to talk about autonomous phone, SMS, email and WhatsApp agents, Apply For Partnership.