Get cited by ChatGPT.

The systems behind the answer.

ChatGPT is not a single model answering from a single source. It's a stack of retrieval, reasoning, and generation layers — each of which has its own optimization implications.

GPT-5 (the base model). Trained on a closed corpus that OpenAI has not fully disclosed but which includes Common Crawl, books, code, licensed datasets, and ChatGPT conversation data. Knowledge cutoff is updated periodically — content published after the cutoff isn't memorized until the next training cycle, which is exactly why retrieval exists.

ChatGPT Search. OpenAI's own search product, launched in late 2024 and now the default behind most queries that need live information. Combines an OpenAI-built search index, Bing's web index, and a reranker that decides which results to ground the answer in. The cited sources you see in the answer come from this layer.

Browsing / Tool use. When a user pastes a URL or an agent invokes a tool that fetches a specific page, ChatGPT issues a request from the ChatGPT-User user agent and reads the page in real time. This is where modern agentic workflows live.

Memory + Custom Instructions. Per-user memory adjusts answers to long-running context, but doesn't affect cross-user citation patterns. Worth noting only because power users with strong custom instructions sometimes see citation patterns that look very different from the average user's experience.

Why being talked about wins.

Third-party citations are the #1 lever in ChatGPT GEO, and the gap between sites that have them and sites that don't is the largest gap in any GEO program. ChatGPT's training corpus is a snapshot of how the internet talks about the world. If your brand is mentioned in editorial articles, listed in roundup posts, profiled by analysts, and cited in industry reports, GPT-5 has learned you exist. If it's not, GPT-5 hasn't.

This is meaningfully different from classic SEO authority. A traditional backlink earns PageRank weight. A citation in a Forbes roundup earns LLM presence — the model not only knows you exist, it knows what context to associate you with, what category you compete in, and which competitors you sit alongside. That contextual association is what causes ChatGPT to recommend you when a user asks an open-ended question.

The competitor-set problem. When a user asks ChatGPT "what are the best [X] tools?" the model returns the brands it most strongly associates with that category. Brands not in that association set don't appear, regardless of their site quality. The work of GEO is making sure your brand is in the association set — and the only durable way to do that is to be mentioned, in context, across enough authoritative sources that the model encodes the association as fact.

Letting ChatGPT actually read your site.

ChatGPT sees your site through three different user agents, and getting the technical layer wrong on any one of them silently caps your visibility on that mode. We audit all three on every ChatGPT engagement.

GPTBot. OpenAI's training crawler. Allowing GPTBot in robots.txt is what gets your content into the next training cycle and, by extension, into closed-book recall. Sites that block GPTBot vanish from training-time memory. Many sites block it by default through Cloudflare or similar — it's the most common silent error we surface.

OAI-SearchBot. The crawler that builds OpenAI's own search index used by ChatGPT Search. Different bot, different policy decision. Blocking OAI-SearchBot makes you ineligible for ChatGPT Search citations even if GPTBot is allowed.

ChatGPT-User. The live user-agent used when ChatGPT (or a user-driven agent) fetches a specific URL during a conversation. Blocking ChatGPT-User breaks every "summarize this page for me" workflow and, increasingly, every agent-driven workflow.

llms.txt. The emerging convention for telling LLM crawlers what your site covers, where the canonical content lives, and how to interpret your structure. We deploy llms.txt on every GEO engagement — it's low-cost, growing in influence, and rapidly becoming a baseline expectation.

Schema. JSON-LD remains the cleanest signal you can give every LLM about what a page is. Article, FAQPage, Product, Organization, BreadcrumbList — all foundational. Schema doesn't just lift Bing rich results; it directly improves the reranker's confidence in what to cite.