What's the difference between being mentioned by an LLM, and recommended?

Key takeaways

• Mention scoring records a hit when your brand name appears in an LLM answer. Recommendation scoring records a hit when the model also treats you as a fit for that question.
• Featured scoring looks for a stronger placement: an early mention, or language like "best," "top pick," or "strongly recommend."
• A mention can appear in a list, in an aside, or in a warning. Language that cautions people away from you still counts as a mention.
• Citation scoring (for example in Microsoft Copilot) records when a page from your site was used as a source. That can happen with or without your name, and with or without a recommendation.
• When mention rate runs ahead of recommendation rate, the answers often show vendor roundups, leftover category language, or comparison copy that lives on other sites.

When a buyer asks ChatGPT, Claude, or Gemini what tool or vendor to use, your brand name may appear in the answer alongside other brands.

Most AI-visibility programs measure that the same way. They send buyer questions to the models, then count how often the brand name (or a close variant) shows up in the reply. Teams record that count as a mention. The same reply might put you in a list of eight vendors, describe you as a former option, or warn the reader off. A second read of the same text tells you how the model positioned you. Recommendation scoring tries to capture that positioning.

Featured scoring looks at how strongly you were placed in that answer. Citation scoring looks at whether the model used one of your pages as a source. The rest of this piece explains how to tell those signals apart on the page. For the wider system around this work, see how to influence what AI is saying about your brand.

Prompt panels count a mention whenever the brand string appears

A mention gets recorded when the model wrote your brand, or a close spelling of it, somewhere in the response. The software matches that string without distinguishing whether you appeared first, seventh, or in a footnote after another brand had already been chosen.

That matching rule explains why mention rate became the default AI-search KPI (key performance indicator). ChatGPT, Claude, and Gemini still do not give you first-party query data the way Google Search Console does. So teams and tools send a set of targeted prompts and count the hits. We have described brand mentions as the most reliable metric available for that job. Profound sells a popular commercial version of the same method.

A mention panel helps because you can run it yourself. It does not tell you how often real people type those questions. It also does not tell you what the mention is doing in the sentence. To see that, you need to read the response itself.

Recommendation scoring looks for choice language around the name

Recommendation scoring starts with a mention, then looks at the words around your name. It asks whether the model treated you as an option the buyer might actually consider.

In practice that language looks like: recommend, suggest, top choice, go with, worth considering, a strong fit. The prompt often asks some version of "what should I use?" The surrounding words tell you whether you were one of a few options, or one name among many.

We have used a sales-intern metaphor for AI enablement to explain the same gap. In that metaphor, one intern can recall that your company exists, while another would actually introduce you to an account. Mention vs. recommend scoring applies that difference to the written answer.

Featured scoring adds early placement or heavier language

Featured scoring requires a mention and a strong placement. Either the name appears early in the answer, where a skimming reader is likely to see it, or the model uses heavier language such as "best," "top pick," or "strongly recommend."

A single blended "visibility" number averages those stronger placements with weaker "also consider" lines. A dashboard that reports one visibility score can hide that difference: some answers only list you, while others prefer you over alternatives.

Warning language next to your name still counts as a mention

A mention counter also matches answers where the model advises against you.

Negative-mention scoring records a hit when your name appeared and the nearby language reads "don't use," "not recommended," or "steer clear." Sam Lambert's PlanetScale episode showed this in public: Claude told developers the company had shut down. More often the framing stays milder: "legacy tool," "not for enterprise," "fine if you already use it." Those answers still add to the mention count, because the brand name appeared. They do not add to the recommendation count, because the model did not treat you as a fit.

Mentions, recommendations, and citations answer different questions

Each column answers a different question, so it helps to keep them separate.

Citation data comes first-party from Bing. It sits next to mention and recommendation scores rather than replacing them. In February 2026, Microsoft added AI Performance to Bing Webmaster Tools. It shows how often your URLs get cited in Microsoft Copilot, Bing's AI summaries, and some partner experiences. Microsoft states explicitly that total citations do not tell you placement or presentation inside a given answer. Copilot can cite a docs page while recommending a competitor. It can also recommend you without showing a citation chip.

Mention scoring tracks whether the model named you. Recommendation scoring tracks how it positioned you. Citation scoring tracks whether it used one of your pages as a source, on a surface that publishes that information. Those three numbers can rise together, or they can move independently.

Roundups, leftover categories, and off-site copy produce mentions without a pick

When mention rate runs high and recommendation rate does not, the prompts usually show one of a few patterns.

The first pattern looks like a category roundup. The model lists eight logos with a sentence each and does not pick one. Mention rate records every name, even though the response still lists vendors without choosing among them.

A leftover category can produce the same split. We have watched mention frequency grow for a topic a company used to own (in one case, "data mesh") after the live product had moved on. The model pulls from older pages. What AI search demands from your web content covers consistency and depth. In mention data, that pattern looks like the name attached to an older story.

Comparison language that lives elsewhere can also produce mentions without a pick. If the site does not have a clear "when to use us" page, the model often borrows a competitor comparison, a Reddit thread, or a guess. We have seen Claude repeat a competitor's pricing page when the brand had none.

Mixed source material can blur the framing further. AI search vs. traditional Google still shows the message you wrote on the page a buyer clicks. AI writes a new message from everything it can reach. Off-topic posts, retired products, and several generations of category language can sit in the same answer. Mentions still get counted even when the framing of your brand mixes those sources together.

Comparison pages, current category depth, and off-site sources show up in the transcripts

Start with the responses where you appear but the model does not use recommendation language. Those transcripts make the most useful reading list, because they show how the model talks about you without treating you as a fit.

The sentence usually shows how the model slotted you: too expensive, too old, too narrow, "also consider," or missing from the opening. Those lines give you a concrete brief for whatever you publish next.

Pages that look like recommendation prompts tend to show up in this layer: alternatives, "best for [persona]," "when not to use us," implementation tradeoffs, and buying-committee objections. Definition pages often produce list mentions. Comparison pages more often produce framed choices. That pattern shows up often enough to use, even though it does not hold in every answer.

If older pages still outnumber the live story, mention rate for a retired category can keep rising. Publishing more depth on the current topics matches the enablement work we already describe. Freshness helps, and it helps most when the new material covers the same category you want the model to use.

Off-site sources show up here too. Review listings, community threads, and partner pages often supply the material the model draws on when it has to choose. You can see which of those sources appear in the answers.

SEO and GEO still start with coverage: which pages get used, how LLM visitors behave, and whether the way models describe you matches how you want to show up. Mention vs. recommend lets you read that coverage back from the transcripts.

In practice, keep four separate notes on each answer: whether the model wrote the name, whether it framed you as a fit, whether that framing sat early or used stronger language, and whether it used one of your pages as a source. Those four readings can come from the same answer, so it helps to keep them separate rather than blending them into one visibility number.

Frequently asked questions

Is a citation the same as a mention?

No. Mention scoring records the brand name in the prose. Citation scoring records a page from your site used as a source. Microsoft's AI Performance report in Bing Webmaster Tools tracks the second thing for Copilot and Bing AI summaries. A model can cite your docs while recommending a competitor, or recommend you with no citation chip at all.

Can a model recommend you without mentioning you by name?

Recommendation scoring starts from a mention. If the model describes "an analytics platform for B2B SaaS" and does not name you, a buyer has nothing to look up, and no brand string exists to count.

What does featured add on top of recommended?

Featured scoring looks for a stronger placement of the same idea. You were mentioned, framed as a choice, and either named early in the answer or called out with heavier language (best, top pick, strongly recommend). Recommended-but-not-featured usually looks like a list item with softer language.

Why would mention rate look healthy while recommendation rate does not?

The name shows up in roundups, asides, and "also consider" clauses. Common sources include leftover category content, comparison language that lives on other sites, and answers that list several vendors without picking one.

Does Google Search Console show mention vs. recommendation?

No. Search Console still reports Google search, including some AI-related click paths depending on how you slice referrers. It does not score how ChatGPT, Claude, or Gemini frame your brand inside an answer. That still takes a prompt panel, plus reading the responses.

Should we treat mentions and recommendations as a sequence?

The panel usually shows mentions on the topics you care about first. After that, additional mentions on a retired category, or in negative framing, record a different observation. The pages that resemble the recommendation prompts already in the panel usually show up in this layer.

How do negative mentions happen?

The nearby text warns the reader off. That can come from outdated shutdown rumors, a messy migration, a competitor comparison, or a review pile-on. The PlanetScale/Claude episode made that pattern easy to see. A quieter version copies "not a good fit for enterprise" from a page you did not write.

Related reading

• How to influence what AI is saying about your brand — The fuller playbook for stocking the ingredients AI uses when it talks about you.
• Real (not approximated) AI search data is here — Why prompt-panel mention scores exist, and what Bing's citation data adds.
• The one part of AI search you can actually control — How AI enablement works: you cannot dictate the model's wording, but you can control the materials it draws from.