How AI systems recommend businesses

There is no single mechanism behind “AI recommendations.” ChatGPT, Claude, Gemini, Perplexity, and Copilot are built by different companies on different data, and where they retrieve from the live web they use different sources and rank them differently. That is why they can disagree about the same business on the same day.

What is consistent enough to act on is the general shape of how any of them arrives at naming a business. Understanding that shape matters more than chasing the exact behaviour of one product, which changes without notice.

Two sources of knowledge, combined differently

Every assistant works from some mix of two things: knowledge baked into the model during training, and information retrieved at the moment you ask. Perplexity and Copilot lean heavily on live retrieval and tend to cite sources. ChatGPT and Gemini blend absorbed knowledge with retrieval depending on the product and question. Claude's answers likewise depend on whether the deployment has live browsing. The balance differs by system, version, and question — which is exactly why no one balance can be optimised for.

Retrieval: finding candidate businesses

Where a system retrieves, it first has to find plausible candidates. This is semantic matching: the question is interpreted by meaning, and businesses whose public information plainly covers that need become candidates. A business described in abstractions, or whose facts cannot be fetched, often never enters the candidate set at all.

Ranking: choosing among candidates

Among candidates, systems favour the businesses they can describe safely. Clarity, internal consistency, and corroboration from independent sources all reduce the risk of stating something wrong. A model generating a direct answer is exposed if it errs, so it gravitates toward the business it can characterise without hedging.

Where the systems genuinely differ

• Which live sources they trust and how heavily they weight them.
• Whether they cite sources (Perplexity and Copilot usually do; others often do not).
• How current their retrieval is, and how much they fall back on training data.
• How they handle conflicting sources, and how much randomness enters the result.

Because of this, a business well represented to one assistant can be poorly represented to another. Checking a single system tells you little about the others.

What survives the differences

Precisely because the systems vary and change, the durable strategy is the boring one: a clear identity, concrete services and areas, information consistent across sources, claims that can be checked independently, and content that answers real questions in retrievable text. Every plausible version of retrieval and ranking rewards those properties, so work spent on them is not wasted when a system updates.

None of it guarantees a recommendation from any specific assistant. It raises the odds by making your business the safe one to name.

For the ChatGPT-specific view, see how ChatGPT chooses which businesses to mention and how to get recommended by ChatGPT. Our process and services start by recording how each assistant currently describes you.