Getting cited by an AI assistant is not a ranking problem wearing a different hat. It is a supply problem: the model is writing a paragraph, and it needs a sentence it can stand behind and attribute. This is a playbook for supplying one.
How do I get my website cited by ChatGPT?
Three things have to be true at once. The page must be retrievable — indexed, crawlable, and not blocking OpenAI's bots. It must contain self-contained factual sentences that name your brand, because models quote claims rather than paraphrase pages. And the claim should be corroborated somewhere else on the open web, since retrieval rewards repeated co-occurrence.
Miss any one of the three and the other two do not compensate. The rest of this post is each in turn.
Step 1: Be retrievable
This is the step most often skipped, and the only one that can silently zero out everything else.
AI crawlers split into two kinds with genuinely different purposes, and conflating them is the common error:
- Retrieval crawlers —
OAI-SearchBot,Claude-SearchBot,PerplexityBot, and the live-fetch agentsChatGPT-User,Claude-UserandPerplexity-User. These fetch your page at answer time. Blocking them removes you from the candidate set entirely, no matter how good the page is. - Training crawlers —
GPTBot,ClaudeBot,Google-Extended,CCBot,Applebot-Extended. These relate to model training and grounding.
You can allow one group and block the other; they are independent decisions. What you should not do is block retrieval crawlers by accident while intending to opt out of training. Run your rules through the robots.txt Generator — it applies real precedence, where the longest matching path wins and Allow beats Disallow on a tie, which is where most hand-written files go wrong.
Step 2: Write extractable claims
A model quotes sentences, not pages. So the sentence is the unit of work.
An extractable claim is self-contained (it survives being removed from its paragraph), declarative (it asserts something checkable), and attributed (it names the subject rather than saying "we" or "this tool").
Compare:
- ❌ "We're committed to keeping your data private." — no fact, no subject, nothing to quote.
- ✅ "Quill Tools processes files entirely in the browser, so documents are never uploaded to a server." — a specific mechanism, a named subject, quotable as-is.
The second sentence can be dropped into a model's answer without modification. The first cannot appear anywhere.
The answer block
An answer block is a 40 to 60 word passage that answers one question completely, placed directly under a heading phrased as that question. It has to make sense with no surrounding context, because that is exactly how it will be extracted — lifted out of the page and dropped into someone else's paragraph.
Two rules make them work. Put the direct answer in the first sentence and the qualification second — an answer that opens with "it depends" gets truncated precisely where the meaning lives. And resolve every pronoun, because "it" has no antecedent once the block is travelling on its own.
Step 3: Be corroborated
Retrieval systems reward the co-occurrence of a brand and a topic across independent sources. A claim that exists on exactly one domain is a claim a cautious model may decline to attribute; the same claim echoed in documentation, a forum answer, a directory listing and a review is one it will state plainly.
This is the slowest part of the work and the least automatable. It is also why authority still matters here: higher-authority domains are cited disproportionately, for the ordinary reason that they are the safer thing to point at.
The multiplier: publish something only you have
Everything above is table stakes that competitors can copy. Original data is not.
A model has no reason to cite the fourth restatement of a known figure, but it must cite the source of a number that exists nowhere else. First-hand test results, measurements from your own systems, small surveys, and benchmarks you actually ran are irreplaceable — an aggregator can summarise them, but it cannot become them.
This does not require a research budget. Timing a task and reporting the result is original data. So is counting something in your own product.
Housekeeping that pays for itself
- Structured data —
FAQPage,HowToandOrganizationremove ambiguity about what the page contains and who is making the claim. Generate valid JSON-LD with the Schema Markup Generator. - llms.txt — a short curated summary at your domain root. Cheap, and it decides how you are described when a model consults it. See the llms.txt guide.
- Readability — dense, clause-heavy prose is harder to extract cleanly. The Readability Checker scores a draft on Flesch–Kincaid and related measures.
- Freshness — visible update dates and current framing matter for time-sensitive queries.
- Clean Markdown — many retrieval pipelines convert pages to Markdown before the model sees them. Sensible heading hierarchy survives that conversion; layout built out of styled divs does not. The Markdown ↔ HTML converter shows what your page reduces to.
Measuring it
There is no console for this, so build a baseline by hand. Write the ten to twenty questions a real visitor would ask. Ask each in ChatGPT with browsing, in Perplexity, and in Google's AI Mode. Record whether you were named, in what context, and — the valuable column — which sources were cited instead.
That list of substitutes is a direct statement of what the model currently treats as authoritative on your subject. Work through it. Re-run monthly.
Frequently Asked Questions
How long does it take to start appearing in AI answers?
Retrieval surfaces like Perplexity and ChatGPT search fetch live, so a changed page can be reflected within days of being re-crawled. Anything that depends on training data moves on the model release cycle instead — months. Prioritise the retrieval path; it is the part that responds.
Should I block AI crawlers to protect my content?
That is a legitimate choice, but make it deliberately and per-crawler. Blocking training bots while allowing retrieval bots keeps you eligible for citation without contributing to training corpora. Blocking everything removes you from AI answers entirely — which is a real cost, not a neutral default.
Does a high Google ranking guarantee AI citations?
No, but it is close to a prerequisite. Retrieval draws on conventional indexes, so ranking gets you into the candidate set. Whether you are quoted from there depends on whether the page contains a sentence worth quoting.
Start with the crawler check and the schema — both take minutes with the free SEO tools, and they are the two steps that gate everything else.