What Is LLMO? How to Get Cited by ChatGPT, Claude and Perplexity — Crawlers, Writing and Real Log Data

LLMO (Large Language Model Optimization) means preparing your site so that when AI assistants such as ChatGPT, Claude, Perplexity, and Gemini build an answer, they read your pages and show them as sources.

This article focuses on one question: what can a site owner actually do to get cited by AI? The most reliable clues are the crawler documentation that each AI company publishes. Which robot visits for what purpose, and what happens if you block it in robots.txt, is spelled out in each company's own pages. Below you will find that cheat sheet, the research on what kind of writing gets cited, and a measurement from this site's own access logs.

1. What Is LLMO?

Classic SEO aims to rank high in a list of search results. LLMO aims to be included in an AI's answer as a source or supporting evidence. Because the answer is prose rather than a list, what matters is less the rank of the whole page and more which passage of the page gets picked up and how it is rephrased.

LLMO does not replace SEO. ChatGPT and Perplexity search the web and read pages every time they answer, so being findable through search is itself the foundation of LLMO.

2. The Two Routes by Which AI Uses Your Pages

The single most important idea in LLMO is that AI uses web pages through two separate routes. Mixing them up leads to situations like "we blocked everything to avoid training, and now we never get cited either."

Route ① is finished once the model is built. Editing your page changes nothing until the next training run, and from the outside you cannot tell when, or in which version, your text is included. Route ② reads the current page each time a question is asked, so route ② is where a rewrite can take effect quickly. Most of what you can do in LLMO happens on route ②.

The AI companies split these two routes into separate robots. OpenAI states that a site can allow OAI-SearchBot to appear in search results while disallowing GPTBot to signal that its content should not be used for training (Overview of OpenAI Crawlers).

3. AI Crawler Cheat Sheet: What Happens If You Block Each One

The table below lines up each company's official description (checked against the original pages on October 1, 2026). The "If blocked" column summarizes each company's wording; see the linked source for exact conditions.

If you want to avoid training but still be cited, you can split robots.txt like this:

# Search and citation: allow
User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

# Training: disallow
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: Google-Extended
Disallow: /

Four easy-to-miss points follow from the original documents:

• Changes take time. OpenAI and Perplexity both say a robots.txt change can take about 24 hours to be reflected.
• Blocking by IP address is discouraged. Anthropic notes that IP blocking prevents it from reading robots.txt, so an opt-out may not work reliably.
• CDN and WAF settings are gates too. Perplexity publishes steps for allowing its robots in Cloudflare and AWS WAF. Even if robots.txt allows a robot, a bot-protection layer in front of it can stop the request.
• User-triggered fetchers are a separate case. OpenAI and Perplexity state that robots.txt may not apply to ChatGPT-User and Perplexity-User. Even if you think you have blocked them, a page can still be read when a user hands the AI its URL.

Which robots actually visited this site

This site (arte.itlibra.com) does not block any of the robots in the table above. Counting about three and a half days of access logs (September 26–29, 2026), OAI-SearchBot made 866 requests, PerplexityBot 1,086 and ClaudeBot 502. The user-triggered ChatGPT-User (1,275) and Claude-User (438) also visited. Meanwhile, no AI robot fetched the AI-oriented index file /llms.txt even once. The method and full table are in the measurement section of our llms.txt article (one site over three and a half days, so it cannot be generalized).

4. Writing That Gets Cited: What the Research Shows

There is little well-sourced research on "how to write so AI cites you." The most frequently referenced study is GEO: Generative Engine Optimization by researchers from Princeton University and elsewhere (accepted at KDD 2024). It calls AI systems that search and then generate an answer "generative engines" and compares methods for improving how visible a page is in their answers.

The paper's headline is that visibility can be raised by up to 40%, but keep three caveats in mind:

• The test environment. The results were measured on a generative engine the researchers built and on Perplexity as it was at the time. There is no guarantee the same numbers hold for today's ChatGPT or Claude.
• What "visibility" means. It is a score based on how many words of the answer are attributed to your source and where they appear, not clicks or revenue.
• Results differ by domain. The paper reports that adding statistics worked well for law and government topics, while adding quotations worked well for history and people.

The practical lesson is still clear: back up claims with sources and numbers, and say where a quotation comes from. That also makes your writing easier for readers to verify; it is not a trick just for AI.

5. LLMO Steps You Can Take Today

You do not need a large-scale rebuild. In order, the steps are:

Step 1: Check that the gates are open

• Check that robots.txt does not block the search robots listed in the table in section 3.
• Check that your CDN, WAF or hosting bot protection does not block AI robots wholesale.
• Check that your body text does not appear only after JavaScript runs (text present in the HTML is the most reliable).

Step 2: Confirm in your logs that they actually visit

• Count your server's access log by User-Agent. A one-line command for this is in our llms.txt article.
• If search robots rarely show up, suspect the gates first, then how easy your site is to discover (sitemap, internal links).

Step 3: Write passages that can be picked up

• State the conclusion in the first sentence of a paragraph, and make each paragraph make sense on its own (do not rely on "this" or "that" to point back to an earlier paragraph).
• Give every number its publisher and date. Replace "it is said that…" with "a survey published by X in year Y found…".
• Instead of summarizing other sites, include results you tested yourself, your own data and your on-the-ground judgment.

Step 4: Measure visits that come via AI

• OpenAI says ChatGPT automatically adds utm_source=chatgpt.com to referral links from its search results (Publishers and Developers FAQ). Filter for this value in your analytics to see visits from ChatGPT.
• For other AI services, look at the referrer in your analytics.
• Regularly ask ChatGPT, Claude, Perplexity and Gemini questions related to your business, and record whether you appear as a source. Answers change every time, so do not judge from a single result.

6. What You Can Skip, and What to Watch Out For

Adding llms.txt does not make you cited

llms.txt is a proposal for an AI-oriented index of a site (llmstxt.org). OpenAI, Anthropic and Google publish llms.txt files for their own developer documentation, but none of the crawler pages above says that their crawlers read other sites' llms.txt and use it for answers (checked October 1, 2026). Google states that Search does not use llms.txt and that having one neither helps nor hurts rankings. In this site's logs, AI robots fetched it zero times. Whether to add one is covered in our llms.txt article.

"Once in the training data, it works for a long time" cannot be verified

You cannot check from outside what went into training data, and even if your page was included, it is not shown as a source in answers. The claim that "once trained on, it works long-term" cannot be verified. If you want to act, act on route ②, where pages are read at answer time.

Answers change every time

The same question can produce different answers and different sources. You cannot manage this like "holding a top ranking" in SEO; it is more accurate to think of LLMO as raising the probability of being cited.

Do not water down content for AI

Mass-producing short, chopped-up versions of generic explanations to look "AI-friendly" gives AI no reason to cite you. In the GEO paper, what worked was substance: sources, quotations and numbers. Google also says there is no need to break content into tiny chunks for AI search (see our AI Overviews article).

FAQ

Related: AEO vs LLMO Differences / What Is AEO / What Is llms.txt