What is brand representation in large language models?

A brand can appear in an AI-generated answer and still be represented incorrectly.

A model may mention a company but place it in the wrong category, recommend its product to the wrong audience, compare it with irrelevant competitors or repeat outdated information.

In each case, the brand is visible. The problem is what that visibility actually means.

Checking whether a brand has been mentioned is therefore not enough. Visibility is one dimension of brand representation, not its complete description.

Defining brand representation in LLMs

Brand representation in large language models is the observable way in which a particular LLM-based system identifies, describes, categorises, compares, recommends and visibly substantiates a brand in response to a defined scenario, language, time and environment.

Representation includes both what appears in the answer and any significant omissions, distortions or errors. It covers:

• the role assigned to the brand;

• the products, services and audiences associated with it;

• the attributes used to describe it;

• whether and under what conditions it is recommended;

• the competitors with which it is compared;

• the sources shown alongside the answer;

• whether the information is current and factually correct.

There is not yet one universally accepted academic definition of LLM brand representation. Research has examined its components separately, including visibility in generative engines, recommendations, brand attributes, prompt sensitivity, citations and accuracy. EMNLP research has also documented differences in how language models associate and recommend brands.

Brand representation brings these components together as a multidimensional object of research.

What are we actually observing?

Speaking of representation “in a model” is convenient shorthand. In practice, users encounter a wider system that may include:

• the underlying model;

• system instructions;

• the question and preceding conversation;

• the product interface and its settings;

• search or retrieval mechanisms;

• available sources;

• the time at which the query was made.

Not every model has internet access, uses external retrieval or displays citations. Results obtained through an API may also differ from answers produced by a consumer application carrying the same model brand.

Research into Retrieval-Augmented Generation distinguishes knowledge encoded in model parameters from information retrieved from external resources. This does not mean that every commercial system uses the same architecture. Any brand representation study should therefore specify which model, system, interface and response mode were tested.

Nor are we measuring what a model “thinks”. Models do not hold opinions in the human sense. We observe outputs and their properties. This distinction helps avoid anthropomorphic language, a concern noted in the NIST Generative AI Profile.

Brand representation and related concepts

Brand representation overlaps with familiar areas of marketing and research, but it is not identical to any of them.

A traditional search engine presents documents. A generative system can create one answer in which a brand is described, omitted, compared or recommended. Its presence should not be reduced to another type of ranking.

Seven dimensions of brand representation

1. Presence and entity identification

The first question is whether the brand appears. The second is whether the system has identified the correct entity.

A model may combine information about separate organisations or confuse a parent company, local subsidiary, product line and distributor. A name appearing in an answer is therefore not proof of correct representation. The entity and its relationships must also be accurate.

2. Role, category and position

A brand may appear as a provider, manufacturer, expert source, example, partner, competitor, alternative or warning. These roles are not equivalent. A company may be cited frequently as a source of industry information while never appearing among recommended providers. It may also be placed in a category with which it does not wish to be associated.

Representation concerns not only presence but also the function the brand performs within the answer.

3. Products, services and audiences

A system may know a flagship product while overlooking other business lines. It may miss the brand’s B2B services, assign a service to the wrong audience or describe a regional offer as globally available.

The analysis should therefore ask which products and services are attributed to the brand, which remain absent, which customer needs are associated with it and whether the stated scope of the offer is current.

Two answers can mention the same name while describing two substantially different versions of the brand.

4. Recommendation and scenario fit

Mentioning a brand is not the same as recommending it. A system may include it among many options, place it on a shortlist, identify it as the best choice, recommend it conditionally, advise against it or omit it despite a strong fit with the criteria.

Research presented at CHI 2025 found that subtle changes in wording can affect whether target concepts and brands appear in recommendations. This does not mean that every small prompt change will reverse a result. It does show why recommendation should be assessed against a particular scenario, rather than treated as a universal property of a model.

5. Attributes, framing and sentiment

A brand may be described as innovative, accessible, premium, local, safe, specialist or outdated. These attributes shape the answer, but a single sentiment label cannot capture the whole representation.

A positive answer may still reinforce unwanted positioning. For example, an enterprise provider might be praised but described exclusively as a low-cost option for small businesses. Analysis should therefore cover assigned attributes, reasons for selection, caveats, tone and alignment with the brand’s verified positioning.

Positive sentiment does not guarantee accurate representation.

6. Competitive context

Models do not always compare a brand with the competitors the company itself would identify. Answers may include direct rivals, brands from adjacent categories, providers serving different price segments, companies that co-occur in sources or organisations treated as functional alternatives.

This reveals how the system constructs the category and the available set of choices. However, a co-mentioned brand may be a partner, distributor or source rather than a competitor. Co-occurrence, comparison and genuine recommendation should be distinguished.

7. Visible sources and citations

Where sources are displayed, they may include the brand’s website, distributors, media, comparison sites, documentation, competitor pages or outdated materials.

Two cautions are necessary. First, visible sources do not necessarily reveal every factor that shaped an answer. Secondly, a citation does not prove that the source supports the claim beside it. Research into verifiability in generative search engines treats the presence of a citation and its actual support for a statement as separate questions.

An audit should therefore examine both the origin of a source and its relationship to the claim being made.

Not every problem is a hallucination

Hallucination is the best-known label for generative AI errors, but it should not describe every unfavourable result. NIST uses the term “confabulation” for confidently presented false or erroneous content. Brand representation research requires a broader set of categories:

• factual hallucination – a claim conflicts with the facts;

• unsupported claim – available evidence does not substantiate it;

• outdated information – a formerly correct statement is no longer current;

• entity conflation – information about different entities is combined;

• classification error – the brand is placed in the wrong category;

• omission – a significant fact or part of the offer is absent;

• distortion or oversimplification – the statement changes the meaning without being wholly false;

• recommendation mismatch – the brand is recommended despite failing the criteria;

• positioning misalignment – the representation diverges from the brand’s verified identity and current positioning.

Hallucinations are one part of the wider category of brand representation issues. A false claim, missing product, confused entity and misplaced recommendation require different responses.

Why can brand representation change?

Generated answers depend on their conditions: wording, user intent, language, location, conversation history, model, interface, access to search, available sources, date and the variability of generation itself.

A large study published in Transactions of the Association for Computational Linguistics found that paraphrased instructions can produce materially different model results. It covered many models and tasks rather than brands alone, but provides strong evidence of prompt sensitivity.

Three phenomena should therefore be separated:

• scenario or prompt sensitivity – the question is changed;

• output variability – the same scenario is repeated under the same conditions;

• change over time – the test is repeated after models, sources or brand information may have changed.

A single answer is evidence of what a particular system returned under defined conditions. It is not evidence of a stable model opinion or a guarantee of future answers.

Monitoring over time is not automatically a test of stability. It shows differences between measurements but does not replace repeated runs under identical conditions.

How should brand representation be studied?

A responsible study should begin by defining the entity and the reference against which results will be assessed.

1. Define the brand identity

Create a current record of the organisation, products, services, audiences, markets, attributes, competitors and facts requiring scrutiny. Separate facts from aspirations and promotional claims.

2. Design realistic scenarios

Scenarios should reflect genuine information or decision needs: finding a provider, comparing solutions, checking reputation, choosing a product or identifying alternatives. This is not a conventional keyword list. The same subject may require different scenarios for different audiences and decision stages.

3. Record the research environment

Document the model, interface, language, date, internet access and number of runs. Without these details, later comparison becomes unreliable.

4. Preserve the evidence

Aggregated results should not replace full answers. Keep logs, relevant extracts, sources and citations so that every conclusion can be traced back to its evidence.

5. Analyse dimensions separately

Presence, recommendation, sentiment, attributes, competitors, sources and accuracy answer different questions. Combining them into one score may hide the main issue. A highly visible brand may still be assigned to the wrong category, lose recommendations, be described using outdated information or be associated with the wrong audience.

6. State the limitations

Conclusions must match the scope. A study of five models does not describe every AI system; a test in one language does not establish identical representation in other markets. Brand Semantics provides a broader guide to running an AI visibility audit across AI search platforms.

What can this research tell a brand?

A well-designed study can show whether a brand is recognised, which role it plays, which needs and audiences it is associated with, when it is recommended, which competitors appear and where errors or distortions occur.

It cannot, on its own, establish what every user will see, how people perceive the brand, whether an answer changed a purchasing decision or what its direct sales impact will be. Those questions require additional traffic, conversion, attribution or qualitative research.

How does Semantio analyse brand representation?

Semantio is a research platform for analysing and monitoring how large language models describe, compare and recommend brands – and where they generate errors, inaccuracies or other representation issues.

Research can cover brand mentions, role, recommendations, attributes, sentiment, products, audiences, competitors, sources and citations where available, as well as the consistency of claims with the brand’s defined Identity.

Users can move from aggregated findings to individual answers and their supporting evidence. Brand representation is therefore not reduced to a single ranking or opaque score.

Create an account and see how language models represent your brand.

Frequently asked questions

Is brand representation the same as AI visibility?

No. Visibility asks whether a brand appears. Representation also covers its role, recommendations, attributes, offer, audiences, competitors, sources and factual accuracy.

Does an LLM hold a stable opinion about a brand?

No such conclusion should be drawn. An answer is the output of a particular system under particular conditions and may vary by scenario, model, language, sources and date.

Can a highly visible brand still be represented poorly?

Yes. It may frequently appear while being placed in the wrong category, associated with an outdated offer or recommended to the wrong audience.

Is every incorrect answer a hallucination?

No. The issue may instead be outdated information, entity conflation, misclassification, omission, oversimplification or a recommendation that does not fit the criteria.