How AI assistants pick which firm to cite

When a prospect asks ChatGPT "who's a good fee-only advisor for physicians in Austin," the model doesn't reach into a ranked list and read off the top three. It builds an answer. Understanding those steps is the whole game, because a firm can be excellent and still fall out at a specific, fixable stage.

The four steps between a question and a name

Modern assistants — ChatGPT, Claude, Gemini, and Perplexity all differ in detail but share the shape — move through roughly four stages before they name anyone:

Query fan-out (splitting your one question into several web searches). The assistant rarely searches your exact words. It rewrites the prospect's question into several search queries: "fee-only financial advisor physicians Austin," "best RIA doctors Texas," "flat-fee planner medical professionals." Each one pulls a different slice of the web. More →
Retrieval (going and fetching pages to read). Those queries hit a search index or a live web tool. The engine pulls back a few dozen candidate documents — directory pages, review aggregators, editorial "best of" lists, community threads, and firm websites. More →
Grounding (only trusting names it can back up with those pages). The model reads the fetched documents and pulls out the firm names that appear in relevant, on-topic passages. A name that shows up in three sources is on much firmer ground than one that appears in a single page the firm wrote about itself. More →
Consensus (picking the few names most sources agree on). The model composes a short answer — usually three to six names — favoring firms it saw confirmed across independent sources. It stops early. There is no tenth result.

Where firms actually fall out

Most firms assume they lose at step 2, retrieval — "the AI can't find my website." Usually that's not it. Here is where the losses really cluster:

Stage	What going wrong looks like	Why it happens
Fan-out	Your pages match your brand name but not the category phrasing	Site copy is written about the firm ("Our approach"), not the question ("fee-only planning for physicians")
Retrieval	Your own site is found, but nothing else is	You're absent from the directories, review platforms, and lists the fan-out queries surface
Grounding	You appear once, in a source that's clearly self-authored	The model discounts a firm that only its own site vouches for
Consensus	You're mentioned but not among the three it commits to	Competitors are corroborated across more independent sources

The pattern that matters: independent corroboration beats self-description. A firm quoted in one industry directory and one editorial list clears grounding and consensus more reliably than a firm with a beautiful website and nothing else. This is the opposite of classic SEO, where a single strong page can win — and it's why search rankings don't transfer.

A worked example

Take a real diagnostic shape. We ask one prospect question — "best fee-only financial advisor for physicians in Austin" — across all four engines and record whether the firm at bluffcreek.com is named:

Engine	Named the firm?	What it cited instead
ChatGPT	No	Two firms from a regional advisor directory
Claude	No	A NAPFA listing + one editorial "best of" article
Gemini	Yes	Found the firm's own physician-focused landing page
Perplexity	No	Three firms with dense review-platform presence

Read the "cited instead" column and the diagnosis writes itself. The firm's own page was strong enough for Gemini to ground on directly — so retrieval isn't broken. What's missing is third-party presence: the directory, the NAPFA-style listing, the editorial mention, the review density. Three of four engines wanted a corroborating source and the firm wasn't in any of them. The fix isn't "rewrite the website." It's "get onto the surfaces the fan-out queries surface." (For which surfaces, and in what order, see the citation-surfaces guide.)

Why the four engines disagree — and why that's useful

In the example above, one engine named the firm and three didn't. That split is normal, not noise. Perplexity leans hard on live web retrieval and review-heavy sources; Gemini blends Google's index; ChatGPT and Claude weight their training-time knowledge differently depending on whether web tools fire. A firm can be an "A" in one engine and an "F" in another for the same question.

The practical consequence: a one-engine spot check is misleading in both directions. Ask only ChatGPT and you might declare victory that Perplexity would puncture — or panic over a gap that Gemini already covers. Measuring across all four is the only way to see the real shape, which is exactly why the diagnostic captures 48 answers — 12 questions across four engines — rather than a handful.

What this means for what you build

If assistants reward independent corroboration around specific questions, the work is clear:

Write pages shaped like the question, not like a brochure. A page titled "Fee-only financial planning for physicians" matches more fan-out queries than "Our Services."
Get corroborated. A profile on the two or three directories your category's engines actually cite does more than a fifth blog post.
Remove ambiguity. Schema markup and a clean llms.txt help the grounding step read your firm correctly — necessary, not sufficient.
Measure the whole board. One engine is an anecdote; four engines across a dozen questions is a diagnosis.

See exactly where your firm falls out — free.

We run 2 of your prospects' real questions across the major engines and email you the exact answers, with the "who wins instead" column. Usually within one business day. No payment, no call.

Get my free scan Full diagnostic — $490

Next: The AI visibility audit checklist → · How the score is calculated →