The Brand as Told by Machines

For years, a luxury house’s marketing team has worked by selecting contexts: which publications, spokespeople, images, partners and stories to make visible. Generative assistants add an intermediary that synthesises public information into a direct answer. The brand chooses neither the question nor every source nor the final wording.

The strategic question, then, is not “How do we make the model say what we want?” A generated answer can change across sessions, products, languages and model versions. The useful question is: “What information can it find, what does it treat as credible, and which signals does it fail to interpret?” In luxury, GEO is the systematic work on those conditions.

The scale of the phenomenon deserves attention, but published measures require care. The fifth edition of the Comité Colbert and Bain Luxury & Technology study says that 22% of surveyed houses rank AI among their top three technology priorities, up from 5% in 2024. The same presentation reports that 90% of links cited by models point to sites outside the brands, and that seven in ten large houses are underrepresented relative to market share in the sample analysed. These figures are reported by Comité Colbert and Bain on the basis of the analysis presented in their 2026 report. The model-visibility component was conducted with meikai.ai.

In the September 2026 edition of the Luxury AI Visibility Index, Luxe Digital ranked Brilliant Earth first and Cartier second in jewellery: 36,801 and 22,624 citations of their official domains, respectively, collected across five AI systems. It is a useful example because it separates visibility in answers from size, desirability, quality or economic value. The index is a proprietary monthly measurement: it counts official-domain citations across five systems, gives each system equal weight and applies logarithmic scaling; it is not a universal measure of reputation or product quality.

The result suggests a diagnostic question, not a conclusion about the two brands’ value: which questions produced those mentions? Which systems were observed? In what language and at what point in time? A house may be highly visible because it appears in purchase comparisons yet be described in generic terms; it may also have a smaller presence and a more accurate representation. Before optimising visibility, the house needs to define what a positioning-consistent answer would look like.

Research by Dubois, Hess, Dawson and Jaiswal, published in Harvard Business Review in June 2026, helps explain the risk: models tend to recognise explicit signals — names, stated characteristics and direct information — better than luxury’s implicit codes, such as rarity, cultural associations or restraint in language. Making a fact legible does not mean turning every nuance into promotional copy. It means giving the model accurate material without flattening the story.

Two recent studies produce apparently opposing percentages. Muck Rack says 84% of links examined in its “What Is AI Reading?” project came from earned editorial sources, while Yext says 86% of citations in its study could be traced to “brand-managed” sources. A direct comparison would be misleading: the samples, questions, source classifications and research aims differ.

Table 1. Two studies that are not directly comparable: scope and taxonomy as reported by their authors.

The two samples start from different questions. Muck Rack says it focuses on unbranded exploratory questions and excludes transactional queries that already name a brand. Yext says it includes four combinations of intent — branded or unbranded, objective or subjective — across four industries, with location and business context. The comparison helps show how the source mix changes with the questions; it cannot establish which percentage applies to every luxury conversation. “Earned” and “brand-managed” are also different categories: Yext includes websites, local listings and reviews/social among managed sources.

The practical implication is not to decide which percentage to believe. Ask every provider or study: which prompts, models, time window, countries and languages, counting unit and source definition? Without that information, the number is not a usable baseline.

A house cannot instruct models or secure lasting control over their answers. It can improve the information conditions from which those answers are formed. The work has four parts.

**Canonical facts.** History, materials, techniques, availability, services, repair policies and provenance should be presented accurately, kept current and made consistent across official channels. Structured content helps people and automated systems; it does not guarantee that a model will retrieve or cite it.

**Legible codes.** A house can explain in clear language what might otherwise remain implicit: why a technique is distinctive, what continuity connects a collection to the archive, or what a material choice means. Clarity does not require making every element public or abandoning discretion; it requires deciding which interpretations deserve a public foundation.

**Independent sources.** Editorial relationships, cultural institutions, scholars, archives and authoritative retailers all contribute to public representation. The task is not to manufacture citations or multiply promotional pages, but to enable accurate coverage and provide verifiable information. Third parties retain their independence: a house can offer access and documentation, not dictate judgement.

**Listening and correction.** Regular monitoring can identify factual errors, omissions and incongruous associations. Responses should be proportionate: update an official page, clarify a fact, alert the source to an error, or accept an interpretation that does not require correction.

This approach requires a “machine-readable brand code”: a foundation of facts and associations that helps systems reconstruct the brand without reducing it to a generic profile. In short: do not govern answers; govern the source ecosystem.

The first step should not be an optimisation campaign, but a baseline. A straightforward comparative protocol can be run without mistaking research for a definitive commercial audit.

1. Select five houses across different levels of awareness and price, including at least one independent Italian brand. Document the reasons for selection.

2. Prepare fifteen unbranded questions spanning occasions, style, savoir-faire, durability, gifting and category comparisons. Avoid wording that implies the answer.

3. Run each question on ChatGPT, Gemini and Claude in the same language and time window; retain each prompt, date, full answer and cited source. Repeat part of the sample to observe variation.

4. Code presence and order of mentions, factual accuracy, positioning consistency, important omissions and source quality. Separate observations from researchers’ interpretations.

5. Repeat the protocol at regular intervals, keeping prompts and coding rules stable. Record model or interface changes that may affect results.

Table 2. Proposed measures for a comparable baseline.

The protocol described here is a research proposal, not an experiment already conducted: this paper does not present results collected from this sample.

A diagnosis should lead to a short list of interventions, not a catalogue of SEO tactics. If a fact is wrong, identify the most authoritative primary source and update channels controlled by the house. If an essential description is missing, decide whether it should be made public. If a third-party source is inaccurate, contact it with evidence without asking for favourable treatment. If the answer is correct but differs from the brand’s internal language, distinguish a genuine distortion from a difference in tone.

GEO should not become a generic content factory. Publishing dozens of pages built to answer every query can increase repetition and weaken the house’s voice. The priority is a limited, reliable information base that is maintained over time; then observe whether answer quality changes. Any effect should be measured using a design that separates interventions from model changes and normal answer variation.

GEO is neither traditional reputation management nor a shortcut to buying attention. It is a discipline for listening to and governing the public information that feeds a new intermediary. A reasonable objective is to reduce avoidable errors and make brand codes more recognisable when a relevant question begins the conversation.

For the CMO, the decision is not whether to “do GEO”, but which facts and interpretations deserve an official source, which external relationships need care, and which measures can genuinely guide action. For the CEO and creative leadership, the challenge is to hold legibility and selectivity together: to be understood without becoming interchangeable. The operating model that enables this work is explored in Organising the Maison for AI.