How to get cited by ChatGPT, Claude and Perplexity

Getting cited is not one thing you can do. It is a chain of four, and the chain breaks at the earliest link that fails. Working on the fourth while the first is broken is the commonest waste of effort in this whole subject, so here they are in the order they actually happen.

1. You are allowed in. The crawler that feeds that assistant is not refused by your robots.txt or your bot protection.
2. Your content arrives. The page returns real text to a plain HTTP request, with no browser rendering it.
3. A passage of it stands alone. The few hundred tokens the retrieval system hands the model make sense without the rest of the page.
4. It is worth citing. The passage states something specific enough to be quoted and attributed.

Only the last of those is what most people mean by "content quality", and it is the only one you cannot verify mechanically. The first three you can check today.

Link one: being allowed in

Two different doors, and sites close them for different reasons.

The first is robots.txt. The distinction that decides everything here is between a crawler that trains on you and one that cites you. Blocking the first is a legitimate business decision with no traffic cost. Blocking the second removes you from that assistant’s answers entirely. They are usually two lines in the same file, they are routinely written as one, and the site owner almost never learns which one they wrote.

The second door is bot protection, and this one is closed without any decision at all. A managed firewall or an anti-bot product classifies an unfamiliar user agent as a threat and serves it a challenge page. The assistant receives an interstitial instead of your article. Nothing is logged as a refusal on your side, because as far as your CDN is concerned it worked.

A further 72 of 292 (25%) returned something gated, degraded or partial. Together that is a substantial share of well-known public websites that an assistant simply cannot read, and in nearly every case nobody chose it.

Link two: your content arriving

Most AI crawlers do not run JavaScript. If your page is assembled in the visitor’s browser, what arrives is the shell: a <div id="root">, some script tags, and none of your argument. The page looks perfect to you and to Googlebot, which renders, and is empty to the thing you are trying to reach.

You can test this in one command, with no tools and no account:

Copycurl -sL -A "Mozilla/5.0 (compatible; test)" https://yourdomain/your-page \
  | sed 's/<script[^>]*>.*<\/script>//g' \
  | sed 's/<[^>]*>/ /g' | tr -s ' ' | head -c 2000

What that prints is approximately what an assistant gets. If your headline, your first paragraph and what you sell are not in it, that is the problem, and it is upstream of everything else in this article.

Link three: a passage that stands alone

This is the link almost nobody works on, because it is invisible from inside a browser. The model does not read your page. A retrieval system cuts it into passages of a few hundred tokens, matches one against the question, and hands over that fragment alone.

So a paragraph that opens "This means you can..." is unusable: whatever *this* was, it is in a different passage. A paragraph that opens "Warehouse robots from Acme cost from €40,000 and ship across the EU" is quotable on its own, and that is the entire difference.

Three habits from ordinary writing cause most of the damage, and all three are fixable without rewriting anything:

• Pronouns pointing backwards across a paragraph break. "It", "this", "they" and "the above" all resolve to something the model was not given. Name the subject again; it reads slightly repetitive to a person and is the difference between quotable and useless to everything else.
• Facts split from their subject. A heading that says "Pricing" and a paragraph that says "£29 a month" are one fact for a reader and two orphans for a retrieval system. Put the noun in the sentence.
• Navigation outweighing prose. If your page is a large menu, a hero and three short paragraphs, most of its passages will be chrome. Across 261 measured sites, 23 of 261 (9%) produced mostly navigation rather than content.

A quick test with no tooling: take any paragraph from your most important page, paste it into a document on its own, and read it as a stranger. If you cannot say what company it is about and what is being claimed, an assistant cannot either.

Link four: being worth citing

An assistant cites a source when quoting it is better than paraphrasing from the model’s own knowledge. That happens for specifics: a number, a date, a price, a limit, a named method, a measurement somebody made. It does not happen for consensus prose, because the model already contains the consensus and does not need you for it.

Which is why the advice that sounds most like content marketing is the one worth taking literally: publish something only you can publish. Your own pricing, your own limits, your own measurements, your own documented process. A page of general advice about your industry is competing with the model’s own training data and will lose.

What you cannot control, and how to sit with it

There is no submission form, no ranking, and no report from any assistant telling you how often you were used. You cannot buy a citation, and anybody selling you a guaranteed number of them is describing something they cannot observe, let alone control.

The one output-side check available to anyone is to ask the assistants themselves about your category and see who gets named. Do it in a fresh session with no history, ask it more than once, and treat it as an anecdote rather than a metric — the answer varies between runs and between users. It is still worth doing occasionally, because it is the only view of the far end you have.

Everything else in this article is the input side, and the input side is entirely yours. Most sites lose at link one or link two, which is good news: those are the two that take an afternoon rather than a strategy.

The order to do it in

1. Check whether a plain fetch of your most important page returns your actual content. Fix that before anything else if it does not.
2. Check which AI crawlers your robots.txt refuses, one at a time, and decide each deliberately: training and citing are different questions.
3. Check whether a bot wall is answering for you. This one is invisible from a browser and common.
4. Make sure your price, location, contact and what you sell exist in a machine-readable form and not only in the layout.
5. Reread your best page one paragraph at a time, asking of each whether it stands alone.

The first four are what a scan does automatically: one plain fetch with no browser, your robots.txt resolved against every named AI crawler individually, bot walls detected as bot walls rather than as content, and your page cut into passages the way a retrieval system would cut it.