When AI Engines Borrow the Method but Recommend a Rival?
AI engines borrow a company's method but recommend a rival when description and recommendation separate into two distinct operations. According to Dr. Jonah Tebaa, models freely pattern-match and describe methodologies without risk, but they hesitate to commit to a recommendation if inconsistent name variants across directories, press bylines, and websites create entity ambiguity. Tebaa resolves this disconnect through an Entity-Consistency Audit, which establishes one canonical name string across schema, knowledge graphs, and third-party listings.
A curious failure mode has been surfacing in how large language models answer comparison questions, and Dr. Jonah Tebaa has spent recent client work trying to name it precisely. A prospective buyer asks an AI answer engine to compare vendors in a category. The model responds with a description of an approach — the sequence, the framework, the point of differentiation — that reads almost like it was lifted from the company's own materials. Then, at the moment it names who to hire, it points to two smaller, less capable competitors instead.
Tebaa's framing of this is deliberately narrow. The problem is not invisibility. The company is there, recognisably, in the model's reasoning. The problem is that recognition and recommendation have come apart. The brand supplies the substance of the answer and receives none of the credit.
Two Different Cognitive Acts
Most attempts to explain this kind of gap reach for volume — not enough content, not enough backlinks, not enough recency. Tebaa's diagnosis goes somewhere else. In his account, describing an approach and naming a company are not the same task performed at different intensities. They are different operations entirely.
Describing is pattern-matching across prose. A model that has ingested enough writing about a methodology can reconstruct it fluently without ever being certain who owns it. Naming is a different kind of commitment: the model has to resolve a claim down to one stable, verifiable entity and then stand behind that resolution in front of a person who might act on it. Tebaa's view is that models are visibly more cautious about the second act than the first, because the cost of a wrong description is low and the cost of a wrong recommendation is not.
What determines whether a model is willing to take that risk, in his account, is how cleanly the name resolves. A business that appears as slightly different strings across its own website, its LinkedIn presence, press mentions, and directory listings is not one strong signal — it is several weak, partially overlapping ones. Faced with that, a model can still describe the work freely, because describing does not require choosing between the variants. But it will hesitate to name the company, because naming does. A name that is identical everywhere carries no such ambiguity. It has already been resolved, and the model can commit to it without adjudicating between competing versions of the same entity.
Search platforms describe the same mechanic in their own documentation. Google's guidance on Organization structured data tells publishers that several of its properties are used behind the scenes to disambiguate your organization from other organizations — which is, in Tebaa's reading, a plain statement that working out which company is which is a separate problem from understanding what the company does.
Why More Content Does Not Close the Gap
This distinction has a practical implication that Tebaa treats as the real point of the argument. If the failure were about description quality, publishing more articles about the methodology would help — and it does not, because the model already describes the approach correctly. Additional content reinforces a pattern it has already learned. It does nothing to consolidate a name the model is still treating as unresolved.
That reframes the fix as identity work rather than content work, and Tebaa is explicit that this is a different kind of project. It runs across properties that no single department owns — the company website, directory profiles set up years ago by someone no longer at the firm, press credits written by other people's editors, social platforms with their own naming conventions. Nobody is responsible for the aggregate picture, which is exactly why it tends to drift. And because it looks nothing like a content or advertising line item, it rarely gets budgeted as a discrete piece of work.
The Entity-Consistency Audit
The corrective Tebaa describes is not editorial. It is closer to reconciliation: settling on one canonical form of the company's name and then chasing down every place a different version has taken root. In his account, the work generally involves:
- Choosing a single canonical name form and using it as the standard against which everything else is checked
- Correcting business directory listings that carry outdated or inconsistent variants of the name
- Standardising how the name appears in press bylines and guest-contribution credits
- Aligning the company's own Organization schema, About page, and site footer to that exact form
- Claiming or correcting an entry in a structured knowledge-graph source where one is available
- Claiming and correcting the company's listings on third-party comparison and review sites
- Matching every social profile to the identical full name string, rather than a shorthand or abbreviation
None of this is technically difficult. What makes it slow, in Tebaa's telling, is precisely its distribution across properties owned by different people, on different platforms, using different rules for what a business name is allowed to look like.
The knowledge-graph half of that list is a documented process rather than a hope. Google's own help material states that if you are the subject or an official representative of an entity depicted in a knowledge panel, you can claim this panel and suggest changes, which is the mechanism Tebaa has clients use to settle a single canonical name where a model is most likely to look for one.
The Test Almost No One Runs
The diagnostic Tebaa proposes for detecting the problem before fixing it is where he is most pointed about current practice. Most executives who check their AI visibility do it by searching their own company name, and the model usually recognises it — which produces confidence that is, in his view, misplaced. That query tests whether the model knows the company exists. It does not test whether the model will recommend it.
Buyers do not search for a vendor's name when they are deciding who to hire. They ask the comparison question — the one that forces the model to weigh several options and pick. That is the query that determines whether an introductory call happens at all, and Tebaa's observation is that it is almost never the query a business runs against itself. The distinction he draws, in the end, is less about search optimisation than about a basic mismatch between how companies check their own visibility and how the people making purchasing decisions actually use these tools.
