Quick Answer: To get cited by ChatGPT, make sure OpenAI’s crawlers are allowed in robots.txt, get the page indexed by mainstream search engines, write the answer in the first two sentences under a question-shaped heading, expose the question-and-answer pairs in FAQPage JSON-LD, and then test the actual prompts your buyers use so you can see which pages get quoted and which get skipped.
How does ChatGPT decide what to cite?
When ChatGPT answers a question that needs current information, it runs a retrieval step first: it turns the conversation into one or more searches, pulls back a small set of candidate pages, reads them, and writes an answer that cites a few of the sources it actually used. OpenAI has not published the ranking logic, and anyone selling you a fixed formula is guessing.
What is documented is the plumbing. OpenAI operates separate crawlers for separate jobs, and they respect robots.txt by user agent. Retrieval also depends on your page being discoverable through conventional web indexes in the first place. So the levers you control are access, discoverability, and how easy your page is to quote — not the model’s internal scoring.
Step 1: Let the right crawlers in
Check robots.txt before anything else, because this step is binary. A blanket disallow, or an allowlist written years ago for traditional search bots, will quietly exclude you from every AI answer regardless of how good your content is. The agents worth naming explicitly are OpenAI’s GPTBot, OAI-SearchBot and ChatGPT-User, plus PerplexityBot, ClaudeBot and Google-Extended if you want coverage beyond one assistant.
These serve different purposes, and you may legitimately want different answers for each. Blocking the training crawler while allowing the search crawler is a coherent position: you keep your text out of model training but stay eligible to be retrieved and cited. Decide deliberately rather than by default, then verify the file actually serves what you think it does — a robots.txt behind a redirect chain or a firewall rule is a common silent failure.
Step 2: Be findable in more than one index
Retrieval starts from a web index, and Google is not the only one in play. Sites that are strong in Google and effectively absent from other engines are a recurring pattern behind "we rank well but never get cited". Submit and validate your sitemap in the webmaster tools of the major engines, confirm your key pages return a healthy indexed status, and fix the usual suspects: canonical tags pointing at the wrong URL, parameters splitting one page into five, and pagination that hides half your library.
Step 3: Write so a sentence can be lifted
A citation happens when a model finds a passage that answers the user’s question and can stand alone. That has concrete implications for how you write.
- Headings are questions. Use the phrasing a person would type or say, not a keyword fragment.
- The answer comes first. One or two sentences directly under the heading, before background, before caveats.
- Self-contained sentences. No "as mentioned above", no pronouns whose antecedent is three paragraphs back.
- Specifics over adjectives. A decision rule, a named tradeoff or a concrete step gets quoted; "industry-leading solution" never does.
- Short blocks. Paragraphs of two to four sentences, plus lists, steps and comparison tables that survive extraction.
A useful test: copy any paragraph out of the page and paste it into a blank document. If it still answers a question on its own, it is quotable. If it only makes sense in sequence, rewrite it.
Step 4: Mirror the content in structured data
Structured data does not make a weak page authoritative, but it removes ambiguity. Article or BlogPosting schema tells a machine what the page is, who wrote it, and when it was last updated. FAQPage schema hands over clean question-and-answer pairs instead of asking a parser to infer them from markup. Organization schema, kept consistent across the site, helps a model resolve your brand to a single entity.
Two rules keep this honest. The schema must match the visible page — questions and answers that appear only in JSON-LD are a liability, not a shortcut. And it must validate: a single malformed block can invalidate everything a parser sees on that page.
Step 5: Give assistants a map with llms.txt
An llms.txt file at your site root is a short Markdown summary of what the site is and which pages matter most, with one-line descriptions. It is a convention rather than a mandated standard, so treat it as cheap insurance rather than a ranking factor. The work of writing it is still worth doing: forcing yourself to name your ten most important pages in one line each usually exposes navigation and positioning problems you can fix on the site itself.
Step 6: Test the prompts, not the keywords
You can't manage what you never look at, and there's no rank tracker for generated answers. Build a small prompt set instead — the twenty to fifty questions your buyers genuinely ask, phrased conversationally — and run it on a schedule. For each prompt, record whether you were named, linked, or absent, and which sources were cited instead.
That competitor list is the most useful output. Read the pages that beat you and look at their shape, not their word count: where the answer sits, how the headings are phrased, whether there is a table, how specific the claims are. Then fix one page, wait for a recrawl, and re-run the same prompts. Because generated answers vary between runs, judge results on the trend across many prompts rather than a single response.
What commonly blocks citation?
- Crawler blocks in robots.txt, a WAF, or a bot-management rule nobody remembers enabling.
- Client-side rendering that leaves the main text out of the initial HTML response.
- Gated content — a model cannot cite what sits behind an email form.
- Buried answers that arrive in paragraph six after a long narrative introduction.
- Undated pages with no author, which lose to a dated, attributed alternative when sources disagree.
- Thin duplicates, where five near-identical pages split whatever authority you have.