How to get cited by ChatGPT and other AI assistants
Citation depends on three things you control: whether the assistant may fetch you, whether your answer can be cleanly bounded, and whether there is anything to attribute it to.
By Sapun Lamichhane · Arcetis
Being cited by an assistant is not a ranking you climb. It is a set of conditions that either hold or do not, and each is testable. Nothing here requires guessing at a model's preferences.
Condition one: it has to be allowed to fetch you
OpenAI operates three crawlers and they do different jobs. GPTBot collects training data. OAI-SearchBot indexes for search. ChatGPT-User fetches your page live, at the moment somebody asks a question your page might answer.
That last one matters most here, and it is the one blanket AI blocks catch by accident. A site that decided in 2023 to keep its content out of model training frequently blocked all three, and is now absent from live answers about its own business — a completely different decision from the one it thought it was making.
Anthropic splits the same way with ClaudeBot, Claude-User and Claude-SearchBot. Perplexity uses PerplexityBot and Perplexity-User. Google's AI surfaces are governed by Google-Extended, which — and this trips up nearly everyone — has no effect whatsoever on ordinary Google ranking.
Condition two: your answer has to be liftable
An assistant composing a response has to find a claim, decide where it starts and ends, and judge whether it survives being removed from its surroundings. Content that fails that test does not get cited even when it is the best answer on the internet.
What makes a passage liftable is structural rather than stylistic:
- A heading phrased as the question being answered, so the match is explicit rather than inferred from context.
- The answer in the first sentence beneath it. Building up to a conclusion across four paragraphs makes the conclusion unliftable.
- Self-containment. A passage that depends on 'as described above' cannot travel.
- Specifics — numbers, dates, named standards. Assistants prefer citable precision to confident vagueness, and so do readers.
- FAQ or HowTo markup where the content genuinely is questions or steps, which removes the parsing problem entirely.
Condition three: there has to be something to attribute
A citation needs a source. A page with no named author, no date and no identifiable organisation behind it is a page an assistant can use and cannot credit — and given the choice between two equally good answers, it will cite the one it can name.
- Put a real named author on anything substantive, with a page saying who they are.
- Show a publication date and a modification date, visibly and in your structured data.
- Add Organization structured data so the entity behind the site is machine-readable.
- Keep your facts consistent across pages. Contradictions make a source unreliable, and unreliable sources get dropped.
What does not work
Worth stating plainly, because the market is already full of it:
- Keyword stuffing for AI. There is no keyword density to game; retrieval is semantic.
- Writing prompts or instructions into your page addressed to models. They are content like any other text and confer nothing.
- llms.txt as a fix. It grants no access, and no major operator has documented honouring it. Fix robots.txt instead.
- Publishing volume. A hundred thin pages are a hundred pages nothing can bound a claim within.
A checklist
- 1Resolve your robots.txt per token and confirm OAI-SearchBot and ChatGPT-User are permitted if you want live citation.
- 2Fetch your page with JavaScript disabled and confirm your content is actually in the response.
- 3Rewrite your best pages so each section's heading states a question and the answer follows immediately.
- 4Add a named author and visible dates.
- 5Add FAQ or HowTo markup where it genuinely applies.
Our answer engine checker tests the extractability half of this, and the AI crawler checker tests the permission half. Both are free.