AI models bring brands into their answers through two doors: what they learned during training (their long-term memory of the web) and what they retrieve in real time when they search to answer a question. A brand can be famous in the training data, easy to retrieve on the live web, both, or neither. Getting cited means understanding which door you can open, and each door has different rules and different timescales.
This article goes under the hood. For the big picture, start with What is GEO? A practical guide.
Key takeaways
- Brands enter AI answers through two doors: training data (moves in months) and live retrieval (moves in weeks).
- Retrieval fails at three stages: not found (the AI search crawlers never indexed the page), not readable (JavaScript-only pages), or not quotable (no liftable answer).
- Models prefer unambiguous brands: the same name, category, and description everywhere is effectively a ranking factor.
- Open the retrieval door first: it is fast, measurable, and fully under your control.
Door one: what the model already "knows"
Large language models are trained on a snapshot of the web, books, and other text. During training, a brand that appears often (in articles, reviews, forums, documentation) becomes part of the model's parametric knowledge: the model can name it, describe it, and recommend it without looking anything up.
What determines your presence in this memory:
- Volume and breadth of mentions. Not just your own site: press, comparison articles, community discussions, directories. A brand described by many independent sources gets encoded more strongly.
- Consistency. If half the web calls you "a CRM" and the other half "a marketing suite", the model's picture blurs, and blurry entities get skipped when the model wants to sound confident.
- Time. Training snapshots lag reality by months. A brand that fixed its positioning last week is still described the old way until models retrain.
You influence this door slowly, through classic PR, reviews, and consistent brand messaging. It is the "offline notoriety" half of visibility: the part Huntair scores separately as Brand Knowledge, precisely because it moves on a different clock.
Door two: what the model retrieves live
When a question needs current or specific information, models search the web and read pages before answering (retrieval-augmented generation). This is where most citations with links come from, and it is the door you can open in weeks, not months.
The pipeline has three stages, and you can lose at each one:
Stage 1 (retrieval): does your page get found?
The model turns the user's question into search queries and pulls candidate pages from an index. Two hard facts follow:
- Index coverage decides eligibility. Each provider documents its own search crawler: OAI-SearchBot surfaces sites in ChatGPT's search results, Claude-SearchBot works on search quality for Claude, and PerplexityBot surfaces and links sites in Perplexity's results, which Perplexity states is not used to crawl content for foundation models. If one of them cannot fetch your pages, you are invisible in that assistant's answers, however good your content is. The user triggered fetchers are a separate case: ChatGPT-User, Claude-User and Perplexity-User fetch a page because a person asked for it. OpenAI writes that robots.txt rules may not apply to ChatGPT-User and Perplexity that Perplexity-User generally ignores them, while Anthropic states that its bots honor robots.txt directives.
- Ranking still matters. Retrieval favors pages that already rank well for the underlying queries. This is the concrete sense in which GEO builds on SEO.
Stage 2 (reading): can the crawler see your content?
The fetched page must be readable by the AI crawler, and most of them do not execute JavaScript. In a December 2024 analysis of roughly 1.3 billion monthly AI crawler fetches across their network, Vercel and MERJ found that none of the major AI crawlers rendered JavaScript, and that ChatGPT's and Claude's crawlers downloaded script files in 11.5% and 23.8% of their requests without executing them. Googlebot behaves differently: in the companion study of July 2024 it fully rendered 100% of the indexable HTML pages it fetched on the instrumented site, at a median of 10 seconds after the crawl. A page that renders client-side is therefore an empty document to the AI crawlers, which disqualifies entire websites from AI answers, including sites that rank well on Google. Server-side rendering or prerendering of public pages, plus a robots.txt that doesn't block AI search bots (OAI-SearchBot, Claude-SearchBot, PerplexityBot), are the entry ticket. Training bots such as GPTBot and ClaudeBot are a separate choice: blocking them concerns model training, not AI search results.
Stage 3 (selection): does the model quote you?
From the pages it read, the model composes an answer and picks which sources to cite. Content that wins here is content that is easy to quote:
- Direct answers early. A self-contained two-or-three-sentence answer at the top of a section can be lifted verbatim. Answers buried in paragraph seven rarely make it.
- Question-shaped structure. Headings that mirror real user questions help the model map "which section answers this?".
- Definitions and data. Crisp definitions, statistics, benchmarks, and original research are the most-cited content types: a number with a source beats an opinion every time.
- Clear authorship and freshness. Pages that state who wrote them, when, and on whose authority give the model reasons to trust the quote.
Why do models "prefer" safe, well-known answers?
Beyond the mechanics there is a bias worth knowing: models are tuned to avoid confidently wrong answers. When asked for recommendations, they lean toward brands that many sources agree about: widely described, consistently described, and clearly categorized. The practical consequence: being unambiguous is a ranking factor. Same name, same category, same one-line description everywhere. An AI model choosing between a crisply defined brand and a vague one picks the crisp one, because it can describe it without risk.
What can you actually influence, and how fast?
| Lever | Door | Timescale |
|---|---|---|
| Let the AI search crawlers and Googlebot index you | Retrieval | Days to weeks |
| Make pages readable without JS; allow AI search bots | Reading | Immediate once deployed |
| Answer-first content, question headings, FAQs | Selection | Weeks |
| Publish original data and definitions | Selection + training | Weeks, compounding |
| Consistent entity across the web | Selection + training | Months |
| PR, reviews, independent mentions | Training | Months–model updates |
The strategy writes itself: open the retrieval door first (it is fast and fully under your control), then invest in the slow accumulation of training-data presence.
FAQ
Do AI models have a ranking algorithm like Google?
Not in the same sense. Retrieval uses search-engine-style ranking to shortlist pages, but the final choice of what to cite happens inside the model as it composes the answer, favoring content that is readable, quotable, and consistent with what other sources say.
Why does ChatGPT cite my competitor but not me?
Usually one of three reasons, in order of likelihood: your pages aren't retrievable (blocked or never indexed by that assistant's search crawler, or unreadable without JavaScript), your content doesn't offer a liftable answer where theirs does, or the model's training data simply knows them better. The first two are fixable in weeks, and measurable, as shown in How to check if ChatGPT mentions your brand.
Does llms.txt influence citations?
Nothing in the documentation of Google, OpenAI, Anthropic or Perplexity says it does. llms.txt is an optional convention: a plain-text map of your site for AI systems, at /llms.txt. Google states that Google Search ignores it and that it neither helps nor harms rankings (Google Search Central). OpenAI, Anthropic and Perplexity document robots.txt as the way to manage their crawlers (OpenAI, Anthropic, Perplexity). Publish one if other tools you use need it, but work on retrievability and quotable content first.
Can I pay to be cited?
No mainstream model sells citations in organic answers today. Visibility is earned through the mechanics above, which is exactly why brands that invest early in GEO build a lead that is hard to buy back later.