Mention vs recommendation in AI responses

A brand appearing in an AI-generated answer does not necessarily mean that the system recommends it.

This distinction is easy to overlook, especially when a dashboard reports a single visibility figure or a study is reduced to a list of brand names. A company may be named as an example, cited as a source, compared with competitors, described as unsuitable for a particular use case or included only because the user mentioned it in the question. None of these observations is equivalent to a recommendation.

For teams monitoring brand representation in AI systems, the difference matters. A mention establishes presence. A recommendation indicates that the system has connected the brand with a user’s need, decision context or selection criteria. These are related but separate events, and they should be measured separately.

As explained in our guide to AI brand visibility, visibility is the observable presence of a brand, product, domain or related entity in a defined set of AI-generated answers. Recommendation is one possible form of that presence. It is not its default meaning.

What is a brand mention in an AI response?

A brand mention occurs when an AI system refers to a brand, company, product, service, domain or another defined entity connected with the brand.

The definition may sound straightforward, but a useful study needs a more precise rule. A reference should count only when the answer identifies the intended entity correctly enough for the observation to be meaningful. A similar company name, an incorrectly attributed product or a confused subsidiary should not automatically be counted as a valid mention.

For example, an AI response may mention a company in several different ways:

• as a provider in a category list;

• as an example used to explain a topic;

• as a competitor of another brand;

• as the publisher of a cited source;

• as a company connected with a past event or controversy;

• as an option the user should avoid;

• because the prompt named the brand directly;

• as a product without clearly linking it to its parent brand.

All of these may show that the entity is present in the answer. They do not, however, show the same thing about discoverability, suitability or preference.

A practical operational definition is:

A valid brand mention is a correctly identified reference to a defined brand entity within an eligible AI-generated answer.

The words correctly identified and eligible are important. An answer may contain the right name but still refer to another organisation. It may also contain the brand because the question explicitly required it, which is valuable for recognition testing but much less informative about unbranded discovery.

Not every mention has the same meaning

A single mention rate can conceal important differences. Consider the following answers to a question about software for monitoring how AI systems describe a brand.

The point is not that one category of mention is always more valuable than another. The right interpretation depends on the research purpose. A source-only reference may matter for a publisher seeking evidence that its domain enters an AI system’s source environment. A prompted mention may matter when a business wants to test entity recognition or factual accuracy. A spontaneous mention is more relevant when the question concerns discovery without a brand name in the prompt.

The mistake is treating each occurrence as evidence that the AI system has selected or endorsed the brand.

What is a recommendation?

A recommendation occurs when an AI response presents a brand as a suitable option for the user’s stated or reasonably inferred need, use case, constraints or decision criteria.

A recommendation can be direct:

“For a mid-sized marketing team that needs to monitor brand mentions and recommendations across several AI systems, consider Semantio.”

It can also be conditional:

“If your priority is repeated research using a fixed scenario panel, Semantio may be a suitable option.”

Or comparative:

“Among the tools listed, Semantio is the stronger fit for teams that need to analyse full AI responses rather than track a single headline score.”

In each case, the system goes beyond naming the brand. It links the brand with a selection context and indicates some degree of suitability.

That does not mean the recommendation is necessarily correct, fair or based on complete information. AI systems may make unsupported recommendations, overlook relevant competitors, use outdated information or infer capabilities that a brand does not offer. The fact that a system recommends a brand is an observation about the generated answer. It is not independent proof of product quality or market superiority.

This is one reason why recommendation data should remain connected with the full answer, the scenario, the system and the date of the observation. A recommendation without context is difficult to audit.

A useful distinction: mention, shortlist inclusion and recommendation

In practice, it helps to separate at least four levels of brand presence.

Mention

The brand is present in the answer. No preference or suitability is implied.

Example: “Several companies publish research on AI brand visibility, including Brand X.”

Shortlist inclusion

The brand is included among options that the user may consider, but the answer does not clearly indicate a preference or fit.

Example: “You could compare Brand X, Brand Y and Brand Z before choosing a provider.”

Shortlist inclusion may be valuable. It indicates that the system has brought the brand into a decision set. But it should not automatically be classified as a recommendation.

Recommendation

The answer presents the brand as suitable for a defined need, audience or use case.

Example: “Brand X is a reasonable option for organisations that need quarterly monitoring across several AI platforms.”

Primary recommendation

The answer gives the brand a more prominent position than other options or explicitly presents it as the preferred choice under the stated conditions.

Example: “For this use case, I would start with Brand X because it supports scenario-based analysis and preserves the underlying responses.”

Not every research design needs all four categories. A smaller study may use only mention and recommendation. But the classification rule should be decided before results are reviewed, particularly if the study will be repeated over time.

Otherwise, a change in the reported recommendation rate may reflect a changed interpretation rather than a change in the answers themselves.

Recommendation is always contextual

There is no universal recommendation rate that belongs permanently to a brand.

A brand may be recommended frequently for one audience and rarely for another. It may appear in broad category questions but disappear when the user adds a market, budget, technical requirement or location. It may be named in awareness-stage scenarios while competitors receive stronger recommendations in comparison or decision-stage scenarios.

Consider these questions:

1. “What tools can help a marketing team understand how AI systems describe its brand?”

2. “Which solution is suitable for a communications agency running repeatable AI brand studies for several clients?”

3. “What is the cheapest way to check whether ChatGPT has mentioned my company?”

4. “Which platforms can monitor social media conversations about a consumer brand?”

They are related, but they do not ask for the same solution. A recommendation in one scenario cannot be transferred automatically to another.

This is why a defensible study begins with scenarios, not with a preferred answer. The scenario should describe a realistic user need, relevant constraints and the decision stage being examined. It should not be written merely to create an opportunity for the researched brand to appear.

A well-designed panel may include category discovery, problem recognition, provider comparison, specific use cases and questions for defined audiences. The resulting data can then show not only whether the brand appears, but in which circumstances it is treated as relevant.

Prompted and spontaneous recommendations

The distinction between prompted and spontaneous presence also applies to recommendations.

A prompted recommendation happens when the user names the brand in the question:

“Is Brand X a good choice for monitoring AI visibility?”

The answer may recommend the brand, reject it or describe its strengths and limitations. This can be useful for testing how the system evaluates a known option. It does not show whether the brand would be discovered without prior awareness.

A spontaneous recommendation happens when the brand enters the answer even though the user did not name it:

“Which tools should a marketing team consider when it wants to research how AI systems represent its brand?”

This is usually a stronger signal of unbranded discoverability. The system itself has selected the brand for inclusion. Still, the result must be read carefully. The recommendation may depend on the wording, language, market, system, current web access or other research conditions.

Research on language-model evaluation has repeatedly shown that seemingly small changes in prompt formulation can affect outputs. That is one reason why a study should not treat a single convenient question as a complete picture of a platform’s behaviour. The paper State of What Art? A Call for Multi-Prompt LLM Evaluation is useful background here: it does not concern brand research specifically, but it demonstrates why one prompt wording can produce overconfident conclusions.

What should count as a recommendation?

A measurement framework needs a clear rule. The exact classification can vary by study, but the rule should be explicit and consistently applied.

A response may be classified as a recommendation when it meets both conditions:

1. it identifies the brand as an option relevant to the user’s need, situation or criteria; and

2. it expresses suitability, preference, fit or a clear invitation to consider the brand.

Signals of recommendation can include phrases such as:

• “consider Brand X”;

• “Brand X is a good fit for…”;

• “Brand X may be suitable if…”;

• “I would choose Brand X when…”;

• “Brand X is one of the better options for…”;

• “start with Brand X if your priority is…”.

The wording alone is not always enough. “You can consider Brand X” may be a weak shortlist inclusion if the answer gives no reason for relevance. Conversely, a response may contain a substantive recommendation without using the word recommend.

For this reason, classification should consider the whole answer rather than an isolated phrase. A system can first list several brands, then explain which one suits the specified requirement. It can also recommend a brand conditionally while noting an important limitation. Both the positive fit and the caveat should be preserved in the evidence.

What should not count as a recommendation?

The following situations should usually be kept separate from recommendations:

• the brand is merely named in a list without any indication of fit;

• the brand is cited as a source;

• the answer describes the brand factually after the user named it;

• the brand appears as a competitor or point of comparison;

• the system says that the brand is not suitable for the stated need;

• the product is named but incorrectly attributed to the brand;

• the brand is included because the prompt required every known provider to be listed;

• the answer is too vague to establish whether the system presents the brand as relevant.

The last category matters more than it may seem. AI responses often use compressed language, particularly in tables and short lists. If a researcher cannot determine whether suitability is implied, the record should be marked as ambiguous or reviewed under a documented rule. Forcing every answer into a binary classification can create a misleading level of precision.

Why a high mention rate can coexist with a low recommendation rate

A brand may appear often yet receive few recommendations for several reasons.

It may be widely recognised but associated only with a narrow product or audience. It may be mentioned mainly in prompted queries. It may be used as a reference source rather than an option to choose. It may appear in broad informational scenarios but not in decision-oriented ones. It may also be mentioned alongside competitors without being presented as a fit for the user’s need.

The reverse is possible too. A specialised brand may have a modest overall mention rate but a strong recommendation rate within a small group of highly relevant scenarios. That may be more useful than broad visibility if those scenarios correspond with a real commercial opportunity.

Neither result can be understood from a single percentage alone.

A practical report should therefore show at least:

• the number of eligible answers;

• the number and rate of valid mentions;

• the number and rate of recommendations;

• the split between prompted and spontaneous observations;

• scenario-level results;

• system-level results;

• the full responses or relevant evidence extracts;

• the counting and classification rules.

For a defined study, the basic calculations are simple:

Mention rate = eligible answers containing a valid mention ÷ all eligible answers × 100%

Recommendation rate = eligible answers containing a valid recommendation ÷ all eligible answers × 100%

These rates should not be presented as probabilities that any user will see or receive a recommendation. They describe the proportion observed in the particular sample: selected scenarios, systems, language, market, date and research conditions.

A recommendation can still be inaccurate

Recommendation is not a synonym for positive representation.

An AI system may recommend a brand on the basis of a capability it no longer offers. It may confuse two entities with similar names, attribute a product to the wrong company or omit a limitation that is decisive for the user. It may also recommend one provider while ignoring better-suited alternatives.

This is where recommendation measurement connects with the broader question of brand representation in large language models. Presence and recommendation are observable events, but the quality of representation requires additional checks: factual accuracy, category fit, product attribution, audience fit, competitor context, sources and framing.

A recommendation should therefore be reviewed as an answer in context, not celebrated automatically as a success.

The same principle applies to brand hallucinations. If an AI system recommends a company because it has invented a product, location, certification or partnership, the answer contains a recommendation event but also a factual problem. These observations should remain distinguishable in the research record rather than being collapsed into one positive result.

How to measure mention and recommendation over time

Repeated research can reveal whether a brand’s observed presence or recommendation pattern is changing. But only if the comparative core of the study remains stable.

The same entities, scenario logic, audiences, markets, systems and classification rules should be retained where possible. Changes to the panel may still be necessary as a business evolves, but they should be documented and treated separately from the fixed part of the study.

This is the difference between a one-off observation and LLM brand monitoring. Monitoring compares comparable research periods. It does not mean asking a chatbot occasionally whether it likes a brand.

When a recommendation rate changes, the first questions should be descriptive:

• Which scenarios produced the change?

• Was the movement visible across several systems or only one?

• Did spontaneous recommendations change, prompted recommendations change, or both?

• Did the brand become a primary recommendation or merely enter more shortlists?

• Were the recommendations factually accurate?

• What happened to the competitor set?

• Did the research conditions remain comparable?

Only then can a team decide whether the result is likely to be meaningful, requires further testing or simply reflects normal response variation.

A simple model for reading AI answers

A useful sequence is:

Entity → scenario → response → classification → evidence

First, define the entity: official brand names, variants, products and relationships that allow valid identification.

Second, define the scenario: who is asking, what they need, what stage of the decision they are in and which constraints matter.

Third, preserve the response rather than only a score or a yes/no label.

Fourth, classify the observation: no valid mention, mention, shortlist inclusion, recommendation, primary recommendation, negative recommendation or ambiguous result.

Finally, retain the evidence needed to review the classification later: full answer, platform, date, language, visible sources where available and any relevant caveats.

This does not remove interpretation from the process. It makes interpretation reviewable. It also prevents a favourable-looking mention rate from being mistaken for evidence that AI systems actively direct relevant users towards the brand.

What this distinction does not mean

Separating mentions from recommendations does not mean that every mention without a recommendation is unimportant. Recognition, source presence, category inclusion and competitor co-occurrence can all be useful observations when they answer a defined research question.

It also does not mean that a recommendation rate predicts sales, preference or customer behaviour. AI-generated answers are not a representative sample of market demand, and their wording can vary across systems and time. A recommendation is evidence of what a selected system generated in a selected scenario – no more and no less.

Most importantly, a higher recommendation rate is not automatically the goal. The relevant question is whether the brand is recommended accurately, in the scenarios where it genuinely belongs, to the audiences it can serve and with a representation consistent with verified information.

That is why mention and recommendation should be measured separately. One tells us that the brand entered the answer. The other tells us that the system presented it as a relevant option. Conflating them may create a more attractive metric, but it removes the distinction needed to understand what AI systems are actually saying.

If you want to examine those observations across defined scenarios and AI systems, create a Semantio account and start with a baseline study.