How ChatGPT picks its sources

ChatGPT is the strangest engine we track, because it’s really two engines wearing one interface. Sometimes it browses the web and cites live sources. Sometimes it answers purely from training memory and cites nothing. Which mode you get depends on the question, and the difference decides whether your GEO work shows up in days or in the next training run.

Mode one: browsing

Ask something time-sensitive or commercially specific, “best payroll software for restaurants 2026,” “how much does a kitchen remodel cost in Denver”, and ChatGPT typically searches (Bing’s index, plus its own crawling), reads a handful of pages, and answers with source pills. This mode behaves like a faster, pickier version of Perplexity: your page can be published Tuesday and cited Friday.

From our checks, browsing-mode citations favor pages that are easy to read at speed: answer up top, clear claims, no interstitial walls. ChatGPT reads fewer sources per answer than Perplexity, often three to five, so the bar per citation is higher. One practical implication people miss: being indexed by Bing matters again. We’ve seen sites with sloppy Bing coverage simply not exist in ChatGPT’s browsing results while ranking fine on Google. Check Bing Webmaster Tools; it takes ten minutes and occasionally explains everything.

Mode two: memory

Ask something evergreen, “what’s a good CRM for a small agency”, and ChatGPT often skips browsing and answers from what it learned in training. No links, just names. This is where brand mentions live, and they were baked in months ago, from the open web as it existed during data collection: review sites, comparison posts, directories, Reddit, documentation, press.

You can’t edit a model’s memory, but you can influence the next one. Models train on the corroborated web, so the work is old-fashioned: get named on lists people actually maintain, in category roundups, on G2/Capterra-style platforms, in real coverage. If your brand is absent from every comparison page in your category, no model has a reason to know you exist. That’s a digital PR problem wearing an AI costume.

What moves the needle, ranked

1. Win the comparison surface. When ChatGPT browses a “best X for Y” question, it reads comparison content, yours or somebody’s. Publish honest comparisons including competitors, with a real table and real criteria. Engines cite these constantly because they map exactly onto the question shape.
2. Be specific enough to quote. Pricing on the site, named integrations, deployment timelines. Every concrete fact is a potential extraction; vagueness is invisible.
3. Fix Bing indexing. Unglamorous, occasionally decisive.
4. Seed the corroboration layer. Profiles, directories, third-party reviews that repeat your core claims in their words.
5. Keep priority pages fresh. Browsing mode shares the freshness bias the other engines show.

The patience layer

Browsing-mode wins arrive on content timelines, weeks. Memory-mode wins arrive on training timelines, quarters. Plan both, report them separately, and don’t let anyone judge the second program on the first program’s clock. The rest of the engine-by-engine picture is in the GEO working guide.