GEO for ChatGPT

ARCHITECTURE

ChatGPT is not a single model answering from a single source. It is a stack of retrieval, reasoning, and generation layers, and each one carries its own optimization implications.

GPT-5 (the base model). The current OpenAI frontier family powering ChatGPT, first released in August 2025 and since refreshed with incremental point releases inside the same series. When ChatGPT answers without reaching for a tool, it answers closed-book, drawing only on what the model learned during training. That training memory is bounded by a knowledge cutoff, so anything published after that date is invisible to closed-book recall until the next training cycle, which is exactly why the retrieval layers below exist. For GEO, closed-book visibility depends on appearing, in context, across the open web OpenAI trains on, often enough that the model has encoded your brand at all.

ChatGPT Search. OpenAI's own search product, launched in late 2024 and now the default behind most queries that need current information. It grounds the answer in live results drawn from OpenAI's own search index plus Bing, then attributes its claims with inline citations to the pages it used. Those cited sources, not the training corpus, are what controls citations for anything recent. To be cited here, your page has to be in the indexes ChatGPT Search pulls from and be the most authoritative, extractable answer for the query.

Browsing / Tool use. When a user pastes a URL or an agent invokes a tool that fetches a specific page, ChatGPT issues a live request and reads the page in real time. OpenAI runs three distinct crawlers with separate roles, and blocking each one carries a different consequence: GPTBot gathers the training corpus that feeds closed-book recall, OAI-SearchBot builds the ChatGPT Search index and explicitly does not train the model, and ChatGPT-User performs these user and agent-triggered live fetches. The three are governed by separate robots.txt decisions, so a site can be visible on one surface and invisible on another without ever realizing it. For GEO, the implication is that ChatGPT visibility spans three surfaces, each won differently: closed-book training recall, ChatGPT Search citations, and live browse fetches.

Memory + Custom Instructions. Personalization that tailors answers to a single account, drawing on remembered context and the user's own standing instructions. It shapes how one person's conversation reads, but it is scoped to that account and does not change what other users see or how the model cites sources across the wider population. Power users with strong custom instructions can experience citation patterns that look unlike the average user's, which is why audit query sets are run from clean accounts. For GEO, the takeaway is that personalization is downstream noise, not a lever: cross-user citation patterns are decided by the three surfaces above, not by any one user's settings.

THIRD-PARTY CITATIONS

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.

TECHNICAL & CRAWL

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.