Cited, Mentioned, Recommended: Three AI Outcomes, Not One

AI visibility gets muddled when citation, mention, and recommendation are treated as the same thing.

Ask a team how visible they are in ChatGPT, Gemini or AI Overviews and you usually get one number. Ask what it counts and the answer gets vague.

A citation means your content was used as a source. A mention means your brand appeared in the answer. A recommendation means the AI actually included or favored your brand as an option.

Those are different outcomes.

You can be heavily cited and still have very little brand visibility. You can be mentioned without being recommended. And recommendation is often much closer to the business outcome marketers care about, because it reflects entry into the buyer's consideration set.

This post is about keeping those three apart: why they diverge, how to define and count each one consistently, and what happens when you break them down far enough that the data stops being a score and starts being market research.

---

The evidence that they come apart

The clearest published measurement of the gap between the first two outcomes is Semrush's ghost-citation study, published June 9, 2026. It was written by Margarita Loktionova and run in collaboration with growth advisor Kevin Indig and Growth Memo, using Semrush's AI Visibility Toolkit. Semrush logged 3,981 domain appearances for 115 prompts run across 14 countries and four AI engines: ChatGPT, Google AI Overviews, Gemini and Google AI Mode.

An "appearance" means a domain showed up in an answer at all, as a source link, as a brand mention in the text, or both. Every appearance fell into one of three buckets:

• 61.7% were ghost citations. In Semrush's words, "AI platforms used the page as a source link, but the brand name never appeared in the actual answer."
• 13.2% were cited and mentioned.
• 25.1% were mentioned without a citation.

The study is usually summarized as "62% of AI citations don't lead to a brand mention," which is how Semrush headlined it. The precise denominator is all domain appearances, not citations alone. Working from the published buckets, citations occur in 74.9% of appearances, so the share of cited appearances that never named the brand is closer to 82% (61.7 divided by 74.9). That is my arithmetic from Semrush's figures, not a number Semrush reports, and it makes the point stronger rather than weaker.

The split also varies sharply by engine and by query type:

• Gemini named brands in 83.7% of appearances but linked sources in only 21.4%. ChatGPT did the reverse: an 87% citation rate and a 20.7% mention rate.
• Informational queries ("what is", "explain", "how does") had an 89.3% citation rate and an 18% mention rate. Comparative queries ("best", "vs", "recommend") had a 43.3% mention rate, which Semrush describes as 2.4 times the informational rate.
• In 454 prompt-and-domain combinations tested across multiple engines, the engines disagreed on whether to name the brand in 100 of them (22%).

Semrush's own conclusion is the one I would put on the wall: "Citations and mentions are separate outcomes," and "for accurate measurement, check each engine separately for your top branded and non-branded topics."

Note what it does not measure: whether a named brand was recommended. That is the gap the rest of this post is about. (Its other limits are covered at the end.)

Source: Semrush, Why 62% of AI citations don't lead to brand mentions [Study], by Margarita Loktionova (June 9, 2026)

---

Why the three outcomes diverge

They diverge because each one comes from a different decision inside the system.

A citation is a retrieval decision. Google describes AI Overviews and AI Mode as using a "query fan-out" technique, issuing multiple related searches across subtopics and data sources, and says this lets them "display a wider and more diverse set of helpful links associated with the response" than a classic search. A page gets linked because it helped answer a sub-question. That says nothing about whether the answer needs to name who wrote it. A clear definition of a category term on your blog is exactly the kind of page that gets retrieved for a "what is" question and then dissolved into an unattributed paragraph.

A mention is an attribution decision. The model names an entity when the entity is part of the answer, not just the source of it. On an informational question, the company behind the explainer is rarely part of the answer. On a comparative question, companies are the answer. That is the mechanism the Semrush intent split shows.

A recommendation is a judgment. The model has to decide that a brand fits the user's stated need and say so. That judgment draws on what the wider body of evidence says about the brand: which category it belongs to, which use cases, which company size, which strengths and complaints. Your own pages are one input. Reviews, analyst coverage, comparison articles, community threads and documentation are others. A brand can be widely known, and therefore mentioned, while the evidence about its fit for a particular need is thin or contradictory, and therefore not recommended.

Outcome	Question it answers	What mostly drives it
Citation	Are we being used as evidence?	Retrievable pages that answer specific sub-questions
Mention	Are we visible in the answer?	Whether the brand is part of the answer, which depends heavily on query type
Recommendation	Are we entering the consideration set?	Corroborated evidence that the brand fits the stated need

Source: Google Search Central, AI features and your website (last updated December 10, 2025)

---

Why recommendation is closest to the business outcome

A B2B buyer using an AI assistant during vendor research is building a shortlist. The answer either puts you on that list or it does not.

A citation is useful. It shows your content is trusted as evidence, and it can send a click. But a ghost citation can educate a buyer about the problem and hand the shortlist to someone else in the next paragraph. A mention is better, because the buyer saw your name. But a mention can be "unlike X, which focuses on enterprise," which places you out of consideration for a buyer who just said they run a 40-person team.

A recommendation is the only one of the three that maps directly onto the decision the buyer is making.

One honest caveat before that becomes a slogan: I do not know of any published study that links AI recommendation rates to pipeline or revenue. The position is reasoned from how buying works, not measured. Treat recommendation rate as the leading indicator closest to the outcome, and keep checking it against the lagging indicators you do have, such as self-reported attribution and branded search.

So I would not reduce AI visibility to one number. At a minimum, track:

• Citations: are we being used as evidence?
• Mentions: are we visible in the answer?
• Recommendations: are we entering the consideration set?

Then break those down by engine, topic, and buyer stage.

---

How to define and count each one

Most disagreement about AI-visibility numbers is not about the engines. It is about counting rules nobody wrote down. Write them down before the first run, and change them only with a note in the report.

The unit. Count per answer: one prompt, on one engine, in one run. Every rate is answers with the outcome divided by answers collected, and the denominator travels with the rate. "Recommended in 14 of 60 answers" is a finding. A percentage without the 60 is not.

Citation rule. The answer links to a URL on your registrable domain, inline or in the source list. Normalize subdomains and tracking parameters so one page is not counted three times. Track pages about you on other domains, such as a review profile or an analyst write-up, as a separate field. They matter, especially for recommendations, but they are not your citation.

Mention rule. Your brand or an approved alias appears in the answer text. A link in the source list alone is not a mention; that is the ghost-citation case. Keep an alias list (product names, former names, common abbreviations) and a false-positive list if your brand name is also a common word. A negative mention is still a mention. Record sentiment in its own field rather than dropping it.

Recommendation rule. This is where teams drift, so be strict:

1. Only selection-type prompts qualify. "Best X for Y," "alternatives to Z," "which tool should a team like ours use." A brand cannot be recommended in an answer to "what is X."
2. The answer must present your brand as a suitable option for the stated need, or favor it. Appearing in a list framed as options to consider counts. Being picked as a top choice counts, and is worth recording as a stronger level.
3. A neutral listing does not count. "Vendors in this space include A, B, C and D," with nothing tying them to the user's need, is a mention. So is a market-map paragraph.
4. Counterexamples and warnings do not count. "Avoid X if you need Y" is a mention with exclusionary framing.
5. Echoes do not count. If the user typed your brand into the prompt, the answer naming you is not evidence of anything. Keep branded prompts in a separate set.
6. Conditional recommendations count, with the condition recorded. "X is a strong option if you are already on Salesforce" is a recommendation for that segment, and the condition itself tells you how the market frames you.

A practical scale for each answer: not present, cited only, mentioned, listed as an option, top pick.

Then have a second person code a random sample against the written rule. Where you disagree, the rule is ambiguous, so fix the rule, not the count. If a model classifies answers at scale, check its labels against human coding first.

---

A worked example (hypothetical)

What follows is a hypothetical, built to show the pattern. It describes no real company and uses no real data.

A mid-market accounts-payable automation vendor has a strong library of explainers: what three-way matching is, how approval workflows reduce late-payment fees, what to look for in invoice capture. Its single "AI visibility score" looks healthy, and the team cannot work out why demo requests that mention AI assistants are flat.

Split the same answers into three outcomes and the picture changes:

• Informational prompts ("what is three-way matching"): cited often, rarely named. The explainers work as evidence and earn the brand nothing visible. This is the ghost-citation pattern.
• Comparative prompts ("best AP automation for mid-market manufacturers"): named sometimes, usually in neutral listings, recommended rarely. When the answers do pick options, the sources behind them are review sites and comparison articles where the vendor has little presence.
• Branded prompts ("is [vendor] a good fit for a manufacturer"): named every time, by definition, and therefore excluded from the recommendation count.

The blended score had been averaging a strength (evidence) with a weakness (consideration) and reporting "fine." The explainers need a reason to name the company, such as first-party data. The comparative gap needs third-party evidence about fit for mid-market buyers, which more explainers on the vendor's own blog will not supply.

---

AEO as a market-research system

This is the part I find more interesting than the scorecard. AEO may turn out to be one of the most useful market-research systems B2B teams have.

Most teams start with a visibility question: are we showing up in ChatGPT, Gemini or AI Overviews? But once you look across enough prompts, topics and buyer journeys, something more interesting appears. AI responses start revealing how the market itself is being framed:

• which problems buyers associate with the category, in the language the engines use;
• which competitors repeatedly enter the conversation, including ones you do not consider competitors;
• what attributes are used to compare them: pricing model, integrations, company-size fit, support;
• what objections surface later, and whether they describe the product you sell today or the one you sold two years ago;
• which sources appear to influence those answers, which is where corroborating evidence about you has to exist.

For a B2B company, that can become a surprisingly rich source of market intelligence, and it only becomes visible once the three outcomes are counted separately and cut along the dimensions buyers actually differ on.

By engine, because the Semrush data shows engines behaving in opposite ways on citation and mention. By topic, because the outcome mix tells you which problems the market associates with you and which with someone else. By buyer stage, because being cited at the problem stage and absent at the comparison stage is a specific, fixable finding.

By persona. A CMO may ask about business impact and strategic fit. A practitioner may care about workflow. A technical buyer may ask about integrations, security or architecture. Procurement may focus on implementation risk, pricing or vendor viability. If your brand is visible for practitioners but disappears once the conversation moves into security, ROI or competitive evaluation, that is telling you much more than "our visibility score is low."

It might expose weak documentation. Missing proof. A positioning gap. Stronger third-party validation for a competitor. Or a recurring objection that sales handles every week but marketing has never addressed publicly.

Practically, that means building the prompt set as a grid rather than a list: personas down one side, buyer stages across the top, with each cell holding a few phrasings of the questions that persona would really ask at that stage. Then read recommendation rate cell by cell. The empty cells are the findings.

Agencies see another version of the same opportunity. Across multiple clients, the patterns can reveal where a category is shifting, which sources matter, which competitors are gaining ground, and which questions are becoming more important in the buying process.

---

What that changes about the operating loop

That changes how I think about the AEO operating loop. Instead of simply track, optimize, track again, it starts to look more like this:

1. Observe the market across prompts, personas, stages and engines.
2. Identify gaps in positioning, evidence and content.
3. Act across the right functions.
4. Measure what changed, outcome by outcome.
5. Learn again.

Step three is the one that breaks the old model. A citation gap on informational questions may be a content problem. A recommendation gap on security questions may belong to product marketing and documentation. A competitor favored on the strength of third-party reviews is a customer-marketing and PR problem.

SEO remains foundational, but the output increasingly touches product marketing, brand, content, PR, customer marketing, sales and analytics.

The more I work on this problem, the less AEO looks like a narrow search discipline and the more it looks like continuous customer and market intelligence with an execution layer attached.

---

Honest limits

Answers vary from run to run. The same prompt on the same engine can return different sources, different brands and a different order on consecutive runs. A single run is an anecdote. Run each prompt several times, report rates over runs, and treat small movements as noise until they persist. I covered why this happens in Why Your AI Citations Keep Changing, and how to size a prompt set so real differences can be separated from variance in How Many Prompts Do You Need?.

The prompt set is a sample you designed. Every rate is conditional on which questions you chose to ask. A set weighted toward informational prompts will flatter citations; one weighted toward comparative prompts will produce more mentions. Publish the mix next to the rates and keep the set fixed across periods you intend to compare.

Recommendation coding involves judgment. A written rule and a second-coder check reduce it; they do not remove it. Report your rule alongside your numbers.

Market-research readings are inferences. A competitor that recurs in answers is recurring in answers. Whether buyers make the same association is a hypothesis to test against win-loss notes and customer interviews, not a conclusion.

The external evidence is thin. The best published measurement of the citation-mention gap comes from a vendor using its own tool. It states no collection window, does not name the 14 countries, and reports no sample sizes per engine, country or intent, so its engine-level percentages cannot be compared with much confidence. I am not aware of any public measurement of recommendation rates with a disclosed method, or of any study linking recommendation rate to revenue. For a fuller checklist on judging numbers like these, see How to Read an AI Visibility Statistic Without Getting Burned.

---

The takeaway

Stop asking "how visible are we in AI search?" as one question. Ask three:

1. Are we being used as evidence? Citations.
2. Are we visible in the answer? Mentions.
3. Are we entering the consideration set? Recommendations.

Define each in writing, count per answer with the denominator attached, and break each down by engine, topic, buyer stage and persona. Done properly, the breakdown tells you what a single score never will: how your market frames the problem, the competitors and you.

If you want to see how the three outcomes split for your own brand on the questions your buyers actually ask, that is what Genmark AI is built to measure.

---

Sources

• Semrush, Why 62% of AI citations don't lead to brand mentions [Study], by Margarita Loktionova (June 9, 2026)
• Google Search Central, AI features and your website (last updated December 10, 2025)