Three models, three companies
Meta description: Unsupervised updates propagate through the answer layer because the public record was never made consistent — and no one checks what changed.
The exercise is simple and the result is not. Someone in the brand team — usually after a board member asks — types the same prompt into three AI engines: what is this company, what does it do, who is it for. Three answers come back describing three different companies. One is roughly current. One describes a positioning the company abandoned two years ago. One files the company in a category it has never claimed. The brand book is not the problem — the brand book is consistent. The answers are not, and the brand book has no mechanism that reaches them.
[AWAITING MIKLOS — scene: the same prompt in three engines returns three positions, one of them a retired offer]
What do brand teams assume is portable into a model?
Brand teams assume the identity is portable: the positioning line, the messaging hierarchy, the approved descriptors — crafted in the guideline, deployed on the site, and therefore present wherever the company is described. This assumption held when description flowed through channels the team could see. Models do not read guidelines. They assemble descriptions from the sources they weight: third-party coverage, directories, reviews, forums, older pages the company forgot to retire. The brand’s self-description is one voice in a crowd, and in the cases I watch, often not the loudest.
Where do the engines actually pick up the description?
Miklos Roth, author of Signal Over Noise
The source of record I look for is a consistent public description: the company’s own site stating the current facts, third-party references agreeing with them, and outdated listings corrected rather than abandoned. Sometimes the intervention is on the site — an accurate clarification or a better structured page. Sometimes it is outside the site: improving the business’s public representation, seeking appropriate editorial coverage, or correcting an outdated listing. A large quantity of low-quality references may create noise rather than improve understanding.
What I can report is that the sources are discoverable and routinely surprising. In the cases I have examined, the dominant source for a model’s description was sometimes a years-old analyst paragraph, a directory entry nobody owned, a Wikipedia-adjacent summary, a press release from a partnership that ended. The company’s current site — the artifact the brand team maintains — was one input among many, and its careful current language competed with stale language that had a decade of repetition behind it. Repetition, I have noticed, often beats correctness.
Why does a new guideline PDF not reach the answer?
The false explanation is a brand refresh: the descriptions are inconsistent because the brand has drifted, so tighten the guideline and re-announce. I have watched refreshes launch and the models continue describing the old company, because the refresh changed the owned surfaces and the models were listening elsewhere. A guideline governs what the company says. The incoherence lives in what is said about the company — a space with no guideline, no owner, and in most companies I visit, no monitoring.
What pattern marks coherence failure versus ordinary variation?
Ordinary variation is phrasing: different engines describe the same company in different words. Coherence failure is substance: different engines describe different companies — different categories, different offers, different vintages of the truth. I have learned to check one specific thing first: whether any engine describes something the company stopped being. A retired offer still answering is the cleanest marker, because it has a date. The description is not wrong about the present. It is faithful to the past, and nobody told the past to stop talking.
Before claiming a fix, I would revisit the target question and related questions rather than evaluate one isolated answer, retaining the before-and-after wording, the date, and the test conditions. Even if the new fact begins appearing, attribution should be modest: recommendation systems, retrieval sources, model changes, and query wording may all contribute. I cannot guarantee that an assistant will repeat a corrected answer; I can help a company improve the accuracy and availability of the information from which people and systems may form an answer.
[AWAITING MIKLOS — scene: second scene]
What question should brand and digital ask together?
The incoherence sits exactly between two functions: brand owns the identity, digital owns the surfaces, and the answer layer is a surface that displays the identity — owned by neither. In the companies that eventually address this, I have noticed the work begins with a question the two functions ask jointly, for the first time, about where the company’s description actually lives. Until that question is asked together, each function fixes its own half and the answers stay incoherent.
The question that turns observation into ownership: for each fact that matters to the business, which public page is the source of record, who is responsible for keeping it true, and when was it last checked against what the answer layer actually says?
PR excerpt: [PR — awaiting Miklos material; cut 80–120 words from his paragraphs once available.]
By Miklos Roth — AI marketing consultant, Budapest. More: his profile.
