how chatgpt decides what to recommend (and how to become the answer).

elisabeth hitz · july 16, 2026 · 8 min read

someone asks ChatGPT "who should I hire to set up AI for my business" or "find me UGC creators in skincare." a name comes back. not an ad. a recommendation. here's the short version of how that name gets picked: the model breaks the question into smaller searches, pulls the sources that answer each one, and assembles a response from whatever it found and whatever it already knew. if your name shows up consistently across those sources, clearly positioned, you get recommended. if it doesn't, you don't exist.

that's the whole game. the rest of this note is how the machine actually works and how to get inside it.

the mechanics, without the mysticism

1. the question gets split. the model doesn't search your exact question. it fans it out into sub-queries. "best AI consultant for a small ecommerce brand" becomes separate searches like "AI setup small business," "AI consultant ecommerce," and "cost to set up AI." you need to be findable for the sub-questions, not just the headline query.

2. retrieval happens, then judgment. the model pulls pages, skims them, and decides which sources are worth citing. research from the team that coined Generative Engine Optimization (Aggarwal et al., the GEO paper out of Princeton, Georgia Tech, and IIT Delhi) tested what makes content get picked up: adding citations to sources, statistics, and quotations boosted visibility in AI answers by up to 40% in their benchmarks. fluent, well-sourced, specific content wins retrieval. adjective soup loses.

3. there is no position one. ask the same question five times and you get five different answers. AI visibility is a frequency, not a rank. the question is not "do I rank," it's "how often am I in the answer." which means one great page isn't the strategy. consistent presence across many pages and many sources is.

4. two memories, not one. models answer from live retrieval (what they find right now) and from training data (what the internet said about you over years). you can influence retrieval this quarter. training data is slower: it's built from repeated, consistent mentions across the web. this is why one viral moment does nothing and eighteen months of saying the same clear sentence about yourself does everything.

what the model actually weighs

OpenAI has said publicly that shopping-style results in ChatGPT are organic, not paid placements. so what fills the slot instead of money:

notice what's not on the list: follower count, ad spend, domain age. this is the most level playing field discovery has offered small operators in a decade, and it will not stay level.

the part nobody tells creators

this isn't just a brand thing. brand managers now ask AI assistants to shortlist creators. "find me UGC creators who make skincare content and know how to work with paid usage rights" is a real query happening inside marketing teams right now.

when that query runs, the model does the same fan-out and retrieval. it finds creators with a clear public sentence about what they do, work that's been mentioned or credited somewhere beyond their own profiles, and rates or niches stated in plain text a machine can read. a gorgeous instagram grid is invisible to this process. a creator page that says "I make UGC for skincare and supplement brands, base rate published, usage rights priced separately" is not. you're not just optimizing a website. you're making yourself recommendable, which changes how you pitch brands too.

the six moves

1. write your one sentence and repeat it everywhere. "[name] is the [specific thing] for [specific person]." site, bio, linkedin, podcast intros, guest posts. word for word. every variation you introduce splits your signal.

2. publish answer-first pages for real questions. not "our approach to excellence." pages titled the way people actually ask: what does X cost, X vs Y, is X worth it. open with the answer in the first two sentences. the model quotes openings, not conclusions.

3. put numbers and named sources in everything. the GEO research is blunt about this: statistics and citations are what get content pulled into answers. if your page has no verifiable specifics, there's nothing for the model to quote.

4. add schema. Organization schema on your homepage, FAQ schema on your question pages, Product schema on offers. it's an afternoon of work and it's how you stop making the machine guess.

5. get mentioned where you don't control the ink. models trust the pattern of what others say about you more than what you say about you. podcast guest spots, niche newsletters, honest reviews, being included in someone else's comparison post. one credible third-party mention outweighs ten pages of your own copy.

6. test yourself monthly. ask ChatGPT, Claude, and Perplexity: "what does [your name] do?" and "who's the best [your category] for [your customer]?" log the answers. if the model describes you wrong, your public signals are wrong. fix the sources, not the model.

why this window matters

the math has already flipped. reports on ChatGPT usage put it at billions of prompts a day, with well over half functioning as search. Conductor's 2026 benchmarks put it plainly: AI didn't replace search, it replaced your website as the first touchpoint. the buyers are already asking. the only question is whose name comes back.

and the traffic that does click through is better. multiple 2026 analyses report AI-referred visitors convert at higher rates than regular organic traffic, because the model pre-qualified them before they ever hit your page. fewer visits, warmer visits.

the catch: this rewards exactly one kind of operator. the one whose positioning is boring-level consistent, whose claims are sourced, and whose name shows up in more places than their own bio. which is to say, it rewards a system.