When nobody writes about you, AI answers with your namesake's reputation

Measured on 6 August 2026. 12 entities, 2 questions each, 6 engines, 144 answers.

Ask an assistant about a company and you get an answer whether or not the material to write it exists. We wanted to know what changes between an entity the press has covered for decades and one almost nobody has written about. The answer is not that assistants become kinder when the record is thin. It is that they keep answering, and the substance comes from somewhere else.

What we asked

Every entity was put to the six engines twice, in the same words:

• a neutral question, What should I know about {entity}?
• an adversarial question, What are the main criticisms and risks of {entity}?

The 12 entities were sorted into three strata by how much independent third-party coverage exists about them:

• Thick coverage: Boeing, Shein, Ryanair, Nestlé. Decades of press, documented disputes.
• Medium coverage: Ledger, Hetzner, Backblaze, Fastmail. Known inside their niche, reviewed by trade media and user communities.
• Thin coverage: four brands operated by the publisher of this study. They were chosen deliberately: measuring this stratum means recording whatever criticism an engine invents, and we would rather it land on names we answer for than on a third party's.

Engines: ChatGPT, Claude, Gemini, Perplexity, Mistral, Grok. One pass per question and per engine. The unit below is the cell, one entity as seen by one engine, so 24 cells per stratum, each covering both questions.

The better documented an entity, the less an assistant has anything nice to say

Read the two columns together, because each says something the other does not.

On the thick stratum, the neutral question already returns criticism. Nobody had to ask for it. Across 24 cells only two carried anything an analyst scored as positive, and every single cell carried something negative. Shein drew a negative tone from all six engines on both questions.

On the medium stratum the pattern is the textbook one, and it is almost mechanical: the neutral question yields a positive tone, the adversarial question yields a negative one, 19 cells out of 24 showing exactly that pair. These are entities with a real record, positive and negative, and the question decides which half surfaces.

The thin stratum inverts the picture. Two thirds of its cells carry a positive tone and only a third carry any negative at all, even though half the questions asked for criticism by name. On the surface, that is the deference one might expect toward a small unknown brand. It is not. It is what happens when the question finds nothing and the engine answers anyway.

What the adversarial question does to a name nobody has written about

The eight negative cells in the thin stratum are worth more than the count. In each case we read the answer and its cited sources. None of them contain a criticism of the entity that was actually asked about.

Mistral, asked for the criticisms and risks of Atoutpersona, a French personality-test site, answered about Persona, an American identity-verification company, on the strength of a shared substring. It opened with "likely referring to Persona", then attributed to the name it was given a February 2026 exposed-code incident, extensive biometric data collection, and a controversy over investor ties to Peter Thiel and Palantir. Its closing summary drops the hedge and speaks of "the main criticisms and risks of Atoutpersona". Its two sources are a Trustpilot page for withpersona.com and the Wikipedia article "Persona (identity verification service)". Neither mentions Atoutpersona.

Perplexity, asked the same question about Stileex, a Madagascar data publisher, converted the absence of material into the finding itself: "low transparency", "unclear legitimacy", "limited independently verifiable information". It then imported red flags from Stylex, Stylix, StealthEX and Stelix, listing missing team information, copied contact details and billing complaints. Of the twenty sources it cited, two concern Stileex. The other eighteen concern a React styling library, a crypto exchange, an office furniture maker, a continuous glucose monitor, a 3D printing service and a World Bank report on water in Madagascar. To its credit Perplexity states that these findings may not apply, and lists what it cannot confirm. The tone of the answer is still that of a warning.

A third engine, asked about another thin-coverage brand, produced no entity-specific criticism at all and instead returned eight families of category risk, each one sourced to a real article about the category rather than about the company.

Three engines, three different routes, one mechanism. When the corpus is thin, the adversarial question does not return "little is known". It returns criticism assembled from the nearest available material, carrying real URLs.

Why this matters more than the sentiment score

An entity with thin coverage is not protected by its obscurity. It is exposed by it, in a way the count of negative answers understates: the thin stratum looks gentler on paper and carries the least accurate answers in practice. A brand can be positively described in five answers and, in the sixth, inherit a data breach that belongs to a company on another continent.

Two properties of that failure make it worth watching rather than shrugging at. It is sourced, so it reads as verified to anyone who does not open the links. And it is stable, because it comes from the same public material every time the question is asked, which means it repeats for every prospect, recruiter or partner who asks.

The lever is the material itself. What decides the answer about a thinly covered name is whether there is anything to find under that exact name, and whether it is unambiguous enough that the nearest namesake does not win. That is measurable under this protocol, name by name, engine by engine.

Method

Six engines: ChatGPT, Claude, Gemini, Perplexity, Mistral, Grok. Twelve entities, two questions each, one repetition per question and per engine, 144 answers collected on 6 August 2026 in a single window. Tone is set per keyword mention by an AI analyst, and a cell counts as carrying a negative when at least one mention in either answer was scored negative, which keeps the reading stable regardless of how many times an answer names the entity. Full answers, cited sources and the pages read without being cited are archived for every measurement.

Three properties bound what this measurement establishes, and each is a choice rather than an accident. One pass per cell means a cell carries the variability of a single draw, so the strata are read as blocks and not entity by entity. The stratification is set from known coverage rather than from a measured index of it. The thin stratum is drawn from brands the publisher operates, which is what makes it possible to publish its source lists in full.