# Mentions vs Citations vs Recommendations: The AI Search Metrics That Win You Buyers

> Mentions, citations and recommendations measure different things in AI search. Real audit data on which metric wins buyer shortlists, demos and enquiries.

Source: https://www.citedintel.com/answer-engine-optimization/mentions-citations-recommendations-ai-search-metrics
Published: 2026-08-22
By Cited (citedintel.com) — the generative engine optimization (GEO) platform.

---

A prospect asked me a question on a call last week that I could not answer with a one-liner: what is the difference between a mention and a citation on your platform, and is share of voice built on mentions or on recommendations? I gave a long answer, then went back and pulled the data. This note is the short answer I wish I had given, with the numbers behind it.

The three words get used interchangeably across the [generative engine optimization](https://www.citedintel.com/generative-engine-optimization) space, and that blur has a cost: teams optimize for a number that does not measure what they think it measures. So here are the definitions we hold to, the distribution data from our own audits, and what we have measured so far about how each metric connects to demos.

## The three metrics, separated

When an assistant answers a buying question, three different things can happen to your brand, and each one is a different metric.

A mention becomes a recommendation when it reaches the shortlist; citations show who supplied the underlying trust.

| Metric | What it means | Where it lives | What it tells you |
| --- | --- | --- | --- |
| Mention | Your brand is named anywhere in the answer text | The answer the buyer reads | You are in the conversation at all |
| Citation | A URL the assistant references as a source for its answer | The footnotes and link cards | Which pages the assistant trusts — yours or a third party's |
| Recommendation | A mention placed where buyers look: among the first names in the answer | The top of the shortlist | You are in the consideration set, not the tail |

The two axes are independent. A brand can be mentioned without its site being cited, because the assistant learned about it from review platforms and community threads. A brand's site can be cited on an answer that recommends a competitor. And a brand can be mentioned tenth in a list of fourteen, which is technically visibility and practically nothing.

Our share of voice metric is built on mentions, counted per answer: the share of AI answers on buying questions where the brand is named. One deliberate rule inside it: a question that names your brand never counts toward your score, because asking an assistant about yourself guarantees a mention and measures nothing about discoverability.

## Where mentions land inside an answer: our audit numbers

Here is why the mention-versus-recommendation split matters more than most dashboards admit. We went through the ranked brand appearances recorded across our audits — just over 63,000 of them — and looked at where in the answer each mention actually sits.

- The median AI answer names **10 brands**. The 90th percentile names 18.
- **28%** of all mentions land at positions 1–3 of the answer.
- **17%** land at positions 4–5.
- **55%** land at position 6 or later — past the point where most readers are still comparing options.

More than half of everything a mention-counting dashboard celebrates is a name-drop in the tail of the answer. Buyers do not shortlist the ninth name. Consideration-set research has said for decades that people carry three to five options into evaluation, and the answer format follows the same shape: the first few names get sentences and tradeoffs, the tail gets a comma.

When we built our top-of-answer metric we tested cutoffs against this distribution. A top-10 cutoff keeps 77% of all mentions, so it barely separates signal from noise — it would echo the headline number. A top-5 cutoff keeps 45%, which is a real dividing line between the shortlist and the tail. That is the cutoff we shipped: alongside your share of voice, we now report how often you appear among the first five brands named, on the same denominator, so the two numbers compare directly.

A live example from a healthcare AI brand we track: mentioned in 15% of the answers checked, inside the first five names on 12% of them, and typically placed third when named — while the category leader typically places second. Read together, those numbers say the brand is not just present but shortlisted, and the gap to the leader is one position, not one universe. Read as a single mention count, all of that nuance is gone.

## Citations are a different axis entirely

Citations answer a different question: when the assistant defends its answer, whose pages does it lean on? Across our audits the pattern is consistent — review platforms, community threads and comparison pages dominate the citation lists on buying questions. The assistant needs quoteable tradeoff language, and it takes that language from whoever published it, which is often not the brand being discussed. We wrote about the failure mode this creates in [the invisible brand problem](https://www.citedintel.com/answer-engine-optimization/invisible-brand-problem-zero-ai-mentions): strong products with thin public evidence lose to weaker products that are easier to quote.

Watching your citation mix tells you two things a mention count cannot. First, whether your own site earns any of the trust: if every mention of you rides on third-party pages, your visibility is rented, and a single delisting or a stale review page can move it. Second, which formats win in your category: if comparison tables get cited and feature pages do not, your content roadmap writes itself. Our [review of GEO tools](https://www.citedintel.com/best-geo-tools) exists partly because of this dynamic — comparison content is what assistants reach for when a buying question needs a defensible answer.

One odd discovery from our own search console data this month: some of the highest-ranking queries our pages appear on are not from humans. They carry search operators no person types — quoted brand names with long strings of site exclusions — and they retrieve our competitor comparison pages at positions four to eight. These are assistants and research agents gathering material mid-answer. They never click, so they leave no traffic, but they read. If you only measure sessions, this entire layer of retrieval is invisible; it shows up as citations in AI answers, not as visits in analytics.

## From answers to demos and enquiries: what we can measure

Does any of this move revenue? I will not claim a straight line from a mention to a closed deal, because nobody can measure that line today and you should be suspicious of anyone who says they can. Here is what we can measure, starting with our own funnel.

Every signup on our platform answers a required, write-once question: where did you hear about us? The early distribution: a friend or colleague leads at 36%, social at 26%, Google search at 21% — and one in twenty signups already self-reports finding us through an AI assistant. We are a young brand with modest visibility, and the AI-sourced share is the one we expect to compound, because the behavior driving it is compounding: buyers increasingly start vendor research inside an answer instead of a results page. Microsoft's advertising research points the same direction — engagement on AI-assisted search surfaces runs meaningfully higher than on classic results pages.

The mechanism is a chain with a measurable first link: buyers ask assistants for options, the first few names in the answer become the consideration set, the consideration set books the demos, and demos become pipeline. Mentions measure whether you enter the chain at all. Recommendations — top-of-answer placement — measure whether you enter it where decisions form. Citations measure whether you control the sources that keep you there.

The practical move I now suggest to every prospect: add the same source question to your own demo form, with an AI-assistant option, and watch the share over a quarter. It costs one form field and settles the revenue argument with your own data instead of anyone's promises.

## Which fix moves which metric

The reason to keep the three metrics separate is that different work moves each one.

- **Low mentions.** The assistant has no evidence to name you with. This is a coverage problem: comparison pages, buying guides and third-party review presence. Start with the gaps on the buying questions you lose outright.
- **Mentions without recommendations.** You are in the tail of the answer. Assistants elaborate on brands they can defend with specifics — depth beats breadth here: sharper differentiation on the conversations that matter, plus review evidence that supports a stronger placement.
- **Mentions without first-party citations.** The assistant trusts other people's pages about you more than yours. Publish the checkable, structured content it can quote: comparisons with real tradeoffs, documented specifics, pages that answer the question as asked. Run your pages through a [readiness check](https://www.citedintel.com/free-geo-tool) to see what assistants can and cannot read.

If you are evaluating platforms on this axis, look at whether the tool separates these layers or sells you one blended score. We keep a [comparison of GEO platforms](https://www.citedintel.com/compare-geo-tools) current, and for enterprise teams we have a separate breakdown of [Profound alternatives](https://www.citedintel.com/answer-engine-optimization/top-profound-alternatives-for-enterprise) that covers how each tool handles measurement depth. For the wider discipline split, our guide on [GEO versus AEO](https://www.citedintel.com/geo-vs-aeo) maps where retrieval-time optimization ends and model-memory optimization begins.

The one-line version of this whole note: mentions tell you that you exist, recommendations tell you that you compete, and citations tell you why. Track all three, and never let a dashboard average them into one flattering number.