Brand mentions vs. recommendations in AI answers

A mention means an AI answer names a brand. A recommendation means the answer presents the brand as a choice for the question asked. The two are not the same. A brand can be named as history, as an example, as a point of comparison, or as one to avoid. Counting every name as a win measures how often a brand is talked about, not how often it is chosen.

What are the roles a brand can play in an answer?

Read enough answers and the same roles keep recurring.

• Recommended. Offered as an option to pick. "For a small agency, X is a good fit."
• Neutral mention. Named without being chosen. "X was one of the first tools in this space." or "X is mainly used by large enterprises."
• Negative mention. Presented as unsuitable. "I would avoid X for this."

A fourth case sits in between: the brand appears in a list of options with no verdict. Whether that is a recommendation depends on the framing, and a consistent rulebook matters more than the specific call.

This is the same distinction as in mentioned, recommended or cited: the three ways AI names a brand, here with a research focus.

Why does the difference matter?

Because the decision a person makes depends on the role.

If a founder sees that an assistant "mentions" their product in 30% of answers, they may feel pleased. If half of those are "not suitable for small teams", the picture reverses. Without the roles, the headline number hides exactly what you need to know.

It matters for competitors too. A rival can be mentioned often because it is the incumbent everyone compares against, while another is recommended more often. Mentions track familiarity. Recommendations track preference.

How do you tell them apart reliably?

By rules written down in advance and applied the same way every time.

1. Use the answer's own words. Judge from what the answer says, not from your view of the brand.
2. Define the roles. What exactly makes it a recommendation? Present it as a choice for the question. Do not count a name just because it comes first.
3. Treat order carefully. Being listed first is not a preference by itself. Lists can be alphabetical or by size. Count order only where the answer expresses it, such as "my top pick".
4. Verify the quote. Keep the words that justify the label, and check that they appear in the answer.
5. Use one reader for every assistant. If different people judge different assistants, differences in judging can look like differences in the assistants. See one reader for four AIs.
6. Check the reader. Compare its labels with a hand-labelled test set and publish how well they agree.

Citewake follows this approach and records the role of every brand in every answer. The methodology describes the rules.

What should you measure?

A small set of separate rates, with the same denominator.

• Recommendation rate: answers that recommend the brand, divided by all valid answers.
• Mention rate (neutral) and negative rate: same denominator, kept apart.
• Top-pick rate: answers where the brand is the expressed first choice.
• Citation rate: answers citing the brand's own site.

Reported together, they give a picture: a brand may be recommended often but rarely the top pick, or cited often but seldom recommended. Each rate shows an interval so you can judge it. See measuring recommendation rate with confidence intervals.

Is a mention ever useful?

Yes, for early signals. A young brand may go a long time without being recommended. Being named neutrally is the first sign that assistants know it exists. Tracking mentions separately shows that progress, without calling it more than it is.

It also helps when diagnosing. If a brand is mentioned but not recommended, read the answers. Why is it described as it is? Does it match what the brand wants to be known for? That is often more useful than any rate. See how to spot when AI says something wrong about your business.

What about citations?

They are a third thing. An answer can cite your page without recommending you, and recommend you without citing you. Keep it in its own column. The difference between pages an assistant found and pages it cited adds another layer; see pages found vs. pages cited.

What about answers that name no brand at all?

Some answers name no brand. The assistant may explain how to choose, list criteria, or say it cannot recommend a specific product. These are valid answers, and they belong in the denominator. Dropping them would inflate every rate, since you would be measuring only the answers where brands appeared. A good study reports how many answers named no brand, so readers can see how often the assistant declined to pick. In some categories this is common, particularly for questions that ask how to do something rather than what to buy, and the share itself says something about how assistants treat the category.

A simple habit

When someone reports that "AI mentions our brand in X% of answers", ask four things: mentions or recommendations, out of how many answers, from which assistants, and with what interval. If those cannot be answered, the number is decoration. If you want a free first look at how assistants describe a site, try the Citewake checker.