Eligibility and selection are two different problems
Most advice collapses them. They fail differently and they are fixed differently.
Eligibility is whether the model can reach, read and resolve you at all — crawler access, server-rendered text, one unambiguous business identity. If you fail here, no amount of positioning helps, and the diagnosis is in how to appear in ChatGPT.
Selection is what happens next. Several eligible suppliers are in the retrieved set. The model names two or three. This page is about that step.
The tell is simple. Ask your category question and see whether competitors are named while you are not. If they are, you are eligible and losing the selection.
What actually correlates with being named
Kargaev (2026) built a signal hierarchy for generative engines and the ordering is unusually clear. Brand entity mentions score NIS 0.918 — the dominant signal. Statistics present in the content score 0.747. Citations within the content score 0.671. Content length scores 0.043, and page speed 0.000.
Two things follow, and the second is the one most teams resist.
The first: a model recommends what it can describe. Entity mentions dominate because they are how the model builds a confident description — this company, this category, this location, corroborated across sources that do not belong to the company.
The second: the levers that classic SEO trained everyone to pull are near-null here on their own. Speed and length still matter for readers and for classic ranking. They simply do not decide whether an AI names you. A fast, long, beautifully optimised page about an unresolvable company does not get recommended.
Why corroboration beats assertion
A model weighing whether to name you is, in effect, asking how confident it can be that the claim is safe. Wallat, Heuss, de Rijke and Anand (2025) describe the faithfulness requirement in retrieval-augmented generation: a system that cites sources is expected to ground its claims in those sources accurately. Grounding is easiest when several independent sources say the same thing.
This is why your own website is a weak recommendation signal no matter how well written. It is one source, and it is you. A trade publication, a directory in your sector, a review platform and a customer's case study saying compatible things about the same entity is a different evidentiary situation.
de Oliveira (2026) describes the compounding version of this as the authority loop: stronger semantic associations raise future citation probability, which produces more associations. It is slow, and it is the reason established players hold position — but it is also why the first few genuine external mentions matter more than the next fifty pages of content.
Specificity is the lever a smaller company actually has
MIT's measurement of AI search (Aral, Li, and Zuo, 2026) found answers concentrating on the top 1,000 sites far more than classic search, with markedly lower response variety. Against the broad category question, a small supplier will not win.
The move is not to fight for the broad question. It is to be the unmistakable answer to a narrower one — a specific industry, a specific constraint, a specific geography. "Marketing agency" has no defensible answer. "Agency that works on Dutch and German AI visibility for industrial suppliers" has one, and the model will take it, because a specific match is a safer claim than a general one.
This is a real trade. You give up breadth you were not winning anyway and take ownership of a question you can win.
The work, in order
- Resolve the entity. One legal name used consistently, one description, one address, Organization schema, identical details wherever they appear. Contradictions are what make a model hedge.
- Write the position in plain statements. Who you serve, what you do, where, and what you do not do. The last one is what makes you easy to select.
- Earn corroboration. Sector publications, associations, review platforms, named client work. Three good ones outperform thirty directory entries.
- Answer the category question on your own site. Not "our services" — the actual question a buyer types, answered in the first paragraph.
- Measure with a fixed prompt set, monthly, and read the trend rather than any single run. Response variety is low but not zero, so one test is noise.
Frequently asked questions
How do I get recommended by ChatGPT?
Make yourself resolvable as one entity, state a specific position in plain language, and get independent sources to corroborate it. Entity mentions are the strongest measured generative-engine signal (Kargaev, 2026); corroboration is how they accumulate.
How do I get my business recommended by ChatGPT?
The same mechanism, applied narrowly. A single business wins by owning a question small enough to be answerable — its sector, its region, its constraint — rather than competing for the category term.
Can I pay to be recommended by ChatGPT?
Not for the organic recommendation. Paid placements in AI products exist and are labelled separately; they do not change how the model composes the answer above them.
How long does it take?
Access and entity fixes can change what the model says within weeks, because retrieval happens at query time. Corroboration is slower and accumulates — the authority loop de Oliveira (2026) describes runs in months, not days.
Does ChatGPT recommend based on reviews?
Review platforms function as corroborating sources, so they contribute. They are not a scoreboard the model reads and ranks; a strong profile helps establish that the entity is real and described consistently.
Why does ChatGPT recommend the same few companies?
Because AI answers show far lower response variety than classic search results (Aral, Li, and Zuo, 2026) and concentrate on large sites. The way around it is a narrower question, not a louder claim on the broad one.
Is being recommended the same as being mentioned?
No. A mention is your name appearing anywhere in an answer, including as background. A recommendation is being put forward as the choice. The mechanics of the first are covered in getting mentioned in ChatGPT.
What if the model describes us incorrectly?
Treat it as a corroboration failure rather than a content failure. The model assembled you from what it could find; the fix is making the correct description consistent across the sources it finds.
If you want to see which questions you are already eligible for and which you are losing at the selection step, the free scan runs the prompt set and the entity checks for you.
