And what's still a genuine black box, even to the people who built the model.
Search a while for "how to rank in ChatGPT" and you'll find confident lists of "7 ranking factors" or specific percentages — domain authority weighted at 40%, content quality at 35%, that sort of thing. Be skeptical of any number that specific. Nobody outside OpenAI or Anthropic has access to the actual weighting inside these systems, and the companies themselves haven't published one. What follows is a line between what's actually documented and what genuinely isn't knowable from the outside.
A useful and real distinction is between a model answering purely from what it learned during training, versus a model that's actively searching the live web before answering.
Answering from training (no browsing). In this mode, the model isn't looking anything up — it's generating a response from statistical patterns baked into its weights during training on a large text corpus. A brand shows up in an answer because it appeared, associated with that category, often enough and clearly enough across the training data that the model learned the association. This is why well-known, heavily-documented brands with a long public track record tend to come up more often than obscure ones — not because of some SEO trick, but because there's simply more text about them for the model to have learned from in the first place.
Answering with browsing/retrieval. When an assistant actively searches the web before responding — a mode that ChatGPT, Perplexity, and others support — it retrieves current pages and can cite them directly. In that mode, whether your page can be found and read at all (crawlable, indexed, not blocked) becomes a real, mechanical precondition, distinct from whether the model "likes" you.
These are genuinely different processes with different implications, and conflating them is one of the most common mistakes in AI-visibility advice. A brand can be recommended constantly from training-data memory while never once being cited with an actual link, or vice versa.
This follows directly from how training works, not from speculation: if a brand is discussed clearly, consistently, and by many independent sources across the web, it's more likely to be represented strongly in a model's training data than a brand that's barely written about anywhere. That's a real, mechanical tendency — but it is not a checklist. "Be well-known" is true and also not actionable advice on its own.
Here's where honesty matters more than confidence. Large language models are, by the admission of the companies that build them, still substantially opaque even to their own creators. Anthropic's own interpretability researchers have said plainly that these models are so large and complex that even the people who design them know relatively little about how they actually "think" moment to moment. Anthropic's mechanistic interpretability work has made real progress mapping which internal patterns of neuron activity correspond to which concepts — but that research is about understanding representations inside the model in a lab setting, not about giving outside marketers a lever to pull to get their brand mentioned more.
Concretely, nobody outside the provider can tell you, with real precision:
Any article that answers those four questions with specific confidence is guessing. Be wary of anyone selling you a precise mechanism for something the model providers themselves haven't published.
Work with what's documented, not what's speculative: make sure your brand is genuinely, clearly described across the web by sources other than yourself (real coverage, real reviews, real listings) so there's simply more accurate material for a model to have learned from or retrieved. Keep your own site crawlable and unambiguous for the retrieval-mode case. And don't chase a specific "ranking factor" nobody outside the lab can verify. Beyond that, the honest answer to "why did the model say what it said" is often: we don't fully know, and neither, in the moment-to-moment sense, does the company that built it.
Given how much of this is genuinely unknowable, the highest-value thing you can do isn't reverse-engineering a black box — it's measuring your actual current baseline, so you know whether there's a problem worth working on at all.