How ChatGPT (and Other AI Assistants) Choose Which Businesses to Recommend

How does ChatGPT decide who to recommend?

When a buyer asks ChatGPT "who should I hire in Manhattan for a high-conflict divorce," ChatGPT does not run a Google-style search and return ten links. It generates an answer sentence by sentence, and at the point where a firm name needs to appear, it draws on some combination of three inputs: what it learned during training, what it just retrieved from the live web (if browsing is enabled), and its own trained preferences about which kinds of sources to cite for that question type.

That means there is no single ranking table inside ChatGPT that a business can climb. There is instead a probability distribution — the model has some non-zero chance of naming any of dozens of firms for a given query, weighted by how frequently and confidently those firms show up in evidence the model considers credible. Firms named more often across authoritative pages, structured listings, and reputable editorial coverage carry higher weight.

Different assistants make this call differently. Claude leans heavier on training-data reputation signals. Perplexity leans heavier on live retrieval and cites its sources inline. Gemini pulls Google's knowledge graph. GPT-4o balances training weight against browsing when available. Same question, four different weight-of-evidence calculations, four often-overlapping-but-not-identical answers.

Do AI assistants use Google rankings?

Indirectly, sometimes. None of the frontier assistants simply forward a Google SERP as their answer. What they do is treat pages that Google has surfaced as one input among many — a good Google ranking correlates with the page being in an assistant's training data or retrieval index, and that helps, but it is not sufficient and it is not the only path in.

Perplexity uses live web retrieval, so pages that are well-optimized for search and rank on the relevant queries have a real chance of being cited. Gemini, tied to Google infrastructure, benefits similarly from Google-visible authority. Claude and GPT-4o are more training-data-dependent for the base recommendation — they will cite a firm because they have seen it named repeatedly in reputable text, whether or not it is a top-three Google result on any given day.

The practical read: strong Google SEO helps but does not guarantee AI recommendations, and a firm can be recommended by ChatGPT without being the top Google result for the equivalent keyword. This is a big part of why generative engine optimization (GEO) exists as its own discipline. Our how to rank in ChatGPT guide breaks down the specific signals that matter across each of the four major assistants.

What signals make AI pick one business over another?

The signals that separate a recommended business from an ignored one cluster into a small number of categories. First: named-entity presence across authoritative third-party pages. Assistants cite businesses that appear in the sources they trust — legal directories, professional bodies, editorial reviews, association listings, high-quality news coverage. A firm that is only mentioned on its own site is much harder for a model to justify naming.

Second: content that describes the specific scenarios buyers ask about. AI buyers do not ask "nyc divorce lawyer" — they type full situations in first person. "My spouse hid assets in a business account and we are separating, who handles this in Manhattan." Firms whose sites describe those exact circumstances in plain language give assistants something concrete to retrieve. Firms whose sites only list practice areas do not.

Third: business resolvability. If the assistant needs to check that the recommendation is a real, addressable business, it wants to see consistent name-address-phone data, a functioning website, and cross-referenced listings. A canonical business identity that resolves cleanly across sources is easier to recommend than a directory-only ghost that looks like a scraping artifact.

Fourth: consistency of naming across sources. When "Aronson Mayefsky & Sloan," "Aronson, Mayefsky, LLP," and "Aronson Mayefsky and Sloan LLP" all appear across pages, an assistant that resolves them as one firm strengthens its signal. Sloppy naming across a business's own footprint fragments the evidence.

How to check if AI recommends your business

The naive method is to open ChatGPT and type the query. That is a fine sanity check for one prompt on one day on one assistant, but it is not measurement. A single query at temperature-1 gives you one draw from the model's distribution. Ask the same question five times in five sessions and you may get five different top-three lists.

Real measurement runs a designed prompt set across multiple assistants and multiple samples. Viclaro Atlas, for example, uses 141 designed prompts per vertical against four assistants (GPT-4o, Claude Sonnet 4.6, Gemini 2.5 Pro, Perplexity Sonar-Pro) at two temperatures each — 1,128 model calls per snapshot. Every citation is recorded, canonicalized against a firm registry, and rolled into share-of-voice numbers with confidence bands. That is the difference between "I asked ChatGPT once and my firm was mentioned" and "across a designed sample, my firm captures 8.2% of citations, is cited by 3 of 4 assistants, and has held that position for two consecutive monthly snapshots."

A free version of this measurement is available at the free AI visibility scan — one query, four assistants, side-by-side comparison. It answers the "am I on the list" question in about a minute. For continuous tracking against a category, Atlas leaderboards publishes ranked results across eight NYC verticals.

How AI recommendations compare across assistants

The per-assistant variance is often larger than intuition suggests. In the Viclaro NYC divorce dataset, Claude Sonnet cites Aronson Mayefsky 57 times across the prompt set; Gemini cites the same firm 11 times. Same firm, same 141 prompts, five-times variance between two frontier models. This is not an error condition — it is what happens when four different training pipelines make four different weight-of-evidence calls on the same market.

The practical implication is that "recommended by ChatGPT" is a narrower claim than it sounds. A business can be strongly cited by two assistants and effectively invisible on the other two, and that split is invisible if you only ever ask one model. Measuring across all four surfaces where the exposure actually lives.

The healthiest citation profile is high share plus high model coverage — a business named by 4 of 4 assistants at a defensible share number. A high share driven by one assistant is one training-data refresh away from disappearing. A moderate share evenly distributed across four is durable, and that durability is often more valuable than a top-one position on any single model.

The market signal to watch is when a category converges toward one answer across assistants. In the NYC IVF Atlas, RMA of New York captures 60.1% of AI mentions across the field — an effective canonicalization. When the assistants agree that strongly on one recommendation, the category has moved from "compete on visibility" to "unseat the canonical answer," which is a very different strategic problem.