ChatGPT × AI Search

Roughly 900 million people open ChatGPT in a given week, and when one of them asks who to hire or what to buy in your category, the engine names a ship or two and sails past the rest. ChatGPT SEO is the work of being one of the ships it names. The catch is that ChatGPT answers from two different brains, and only one of them is paying attention to what you publish today.

The short version

ChatGPT answers in two ways. Most of the time it recommends from memory: the trained model, which already encoded who mattered in your category when it was last built. Some of the time it browses live and pulls from pages it fetches in the moment. The two need different work, and most brands pour their effort into the one that touches the smaller share of answers. The same LLM SEO fundamentals sit underneath, but ChatGPT is the engine where the memory brain dominates.

900M

weekly active users on ChatGPT

Up from 400 million a year earlier. OpenAI, February 2026

84%

of AI citations point at earned media

Not a brand’s own pages. Muck Rack, May 2026

7.1%

conversion rate of ChatGPT referral visitors

Second only to paid search. Similarweb, 2026 GenAI Index

ChatGPT recommends from memory far more often than it browses, so the brand that wins is the one the trained model already knows, not the one with the freshest page.

The first brain is parametric. It is the trained model itself: a compressed reading of the public web, where brands that turned up again and again in credible places got encoded into the weights. You cannot inspect those weights, but you can read their proxies in domain rating, editorial standards, and how often serious publications wrote about you before the last training run. This brain has a cutoff and updates on a schedule OpenAI does not publish.

The second brain is retrieval. When ChatGPT browses, it fetches live pages, lifts claims, and can cite something published hours ago. The trouble is that most everyday ChatGPT use never triggers it. The default answer comes from memory, which means the parametric brain is the primary channel for organic brand recommendations, and the one almost nobody plans for. Retrieval is the tryout you can pass this week. The trained model is the roster you join the slow way. A tryout is not a roster spot.

SIGNAL 1

Editorial coverage the model already trusts

An article on a strong, real publication adds one credible node to the next training corpus. Five nodes start to build consensus. Each independent endorsement strengthens the statistical association ChatGPT forms around your name. Private blog networks (PBNs) and low-rating directories do not make it into quality training data, so volume there compounds into nothing.

SIGNAL 2

Quotable structure on pages you control

This is where you serve the retrieval brain. A bolded, 40 to 60 word direct answer near the top hands ChatGPT Browse something it can lift without guessing. Question-style headings with a one or two sentence answer follow the same logic, and a clean FAQ block helps both retrieval and structured-data parsing. Put your answer in paragraph nineteen and you are asking the engine to dig for a reason it does not have. Lead with it.

SIGNAL 3

Press releases that corroborate, not carry

A release through a wire service creates a permanent, citation-eligible record. Keep it in proportion: press releases account for only about 1% of all AI citations. They back up a story the editorial coverage is already telling. A brand running on releases alone, with no independent articles behind it, stays thin in ChatGPT no matter how many it pushes out.

Even within “AI,” the sources do not match. One large analysis of citations found Wikipedia made up nearly half of ChatGPT’s leading sources while Reddit made up nearly half of Perplexity’s. Two of the biggest answer engines, reading off entirely different shelves.

Engines lean on what the rest of the web says about you over what you say about yourself. For ChatGPT that pattern has teeth, because every credible publication that wrote about you before a training run is one more vote the model quietly counted. The memory brain is not counting links. It is counting who vouched for you in places worth reading.

Run your category queries in both standard and browsing mode; the gap between them tells you whether your problem is memory or retrieval. If you show up when ChatGPT browses but vanish in default mode, your pages are fine and your coverage is thin, so the memory brain never learned you. If you are missing in both, start with the coverage. The trend across three months means more than any single day, because models refresh and competitors move while you watch. Tools like Profound and Peec track how often ChatGPT names you and in what tone, over time.

Recommendation queries

“What is the best X?” names brands the most reliably, because that is exactly the call the model is built to make. For most companies these are the prize. Map the handful that match how your buyers actually decide.

Definition and comparison queries

“What is X?” often stays at the category level unless your space is commercialized. “X versus Y” names brands when both already have a foothold in the model. A “best X for a specific use case” gap is where you move first.