Anatomy of a ChatGPT search: how it decides which brands to recommend

When a customer asks ChatGPT for the “best” anything, they don’t get ten links to sort through — they get one confident shortlist of three or four brands. They’re in research mode, building a consideration set, long before any checkout. If you’re not on that list, you’re not in the running. We ran real shopping questions across running shoes, cold brew and business software, and took every answer apart to see exactly how a brand gets on it.

A ChatGPT answer isn’t one step — it’s a short pipeline, and each stage includes or excludes you. It decides to search, rewrites your customer’s question into its own, pulls up a stack of pages, leans on the few it trusts, names a shortlist, and — for some categories — shows products to buy.

It rewrites the question before it searches

The single most overlooked stage: ChatGPT rarely searches the words your customer typed. In four of five questions we captured, it reformulated first — adding the year, a format, or category jargon it supplied itself. Ask for “best marathon running shoes” and it searched “best marathon running shoes carbon plated,” injecting the deciding technical attribute into its own query.

“You’re no longer optimising for what your customer typed. You’re optimising for what ChatGPT typed on their behalf.”

It builds the answer from reviews, not your website

In every category, the shortlist was assembled from magazines, review sites, “best of” lists and Reddit — not brand sites. ChatGPT essentially recommends whoever those sources recommend, which flips the priority: the highest-impact work is getting into the round-ups it reads, not polishing your own pages. It’s closer to PR than SEO.

Whether you can win depends on your category

The shape of the shortlist isn’t the same everywhere. Business software has an entrenched default — HubSpot led ten of our twelve CRM answers — while cold brew is a free-for-all with a different winner almost every run. The first question before spending a dollar on AI visibility is which world you’re in: a default-dominated category where you win the specific intents the leader is weak on, or a fragmented one where the shortlist is genuinely up for grabs.

For some categories, there’s a buy box

Running shoes and cold brew both rendered products to buy, with prices and sellers; business software showed nothing. Where the buy box exists, the prices come from merchant feeds and ChatGPT picks which seller’s offer to surface — so the price a shopper sees is a feed-and-ranking outcome, not a fixed fact. And the shelf is idiosyncratic: shoes ran through specialty retailers and never showed Amazon, Walmart or Target; cold brew ran through the grocery aisle where those giants dominate.

Ask the same question twice, get a different answer

A ChatGPT answer behaves like a poll with a small sample, not a fixed ranking. Across our runs the brand list, the sources and even whether a buy box appeared all moved from run to run. A single check is a coin-flip — which quietly breaks the single-number dashboards much of the AI-visibility market sells. The honest way to track it is to sample each question many times and read the distribution.

“Share of the answer — and its sentiment — not position, is the number worth tracking.”

What to do about it

Work out if you’re the default or the challenger; win the reviews ChatGPT reads rather than your own site; own the one quality the category is actually judged on — even an unglamorous one like a free tier or joint protection; get clean feeds into the buy box where it exists; and measure like a statistician, sampling repeatedly. Taken together, this work is what’s now called generative engine optimization (GEO) — sometimes “AI SEO” or answer engine optimization (AEO): being the brand AI search recommends, not just a link it ranks. The category teardowns below take each of these apart, one vertical at a time.