How to Get Your Site Mentioned by ChatGPT and Claude

The short version, and the part most people get wrong

Most people treat this as a volume problem. More posts, more keywords, more headings. Citation does not work that way, because the model is not ranking your site against competitors in a list. It runs a search, reads a handful of results, and quotes whichever one answers the question it is holding. Being the fourth site to say the same general thing adds nothing. Being the only site that states the price, the exact error message or the actual sequence of steps gets you quoted.

Two different things people mean by “mentioned”

An assistant can name your site for one of two reasons, and they have almost nothing in common.

Retrieval is the one you can work on. The model decides it needs current information, runs a search, fetches pages, and answers with links. That loop happens at the moment of the question, so a page you published yesterday can appear in it. In Anthropic's own API, citations for this are not optional: the web search tool documentation states that citations are always enabled and that each one carries the source URL, the title and a short quoted span.

Training memory is the other one, and you cannot really work on it. If a model knows your brand without searching, that came from text absorbed during training, months or years back, from pages that were already widely repeated. A new site has no realistic path into that, and promises to the contrary are being sold rather than explained.

So the whole practical question is retrieval. When someone says they want to be mentioned by ChatGPT, what they want is to be among the pages fetched and quoted when it searches.

The gate nobody checks: which bots you are allowing

Both companies run several crawlers with different jobs, and they are governed separately. This is where sites quietly disqualify themselves, usually by copying a robots.txt file that blocks AI bots as a blanket policy and never revisiting it.

On OpenAI's side there are three worth knowing. OAI-SearchBot fetches pages so they can appear in ChatGPT's search features. GPTBot collects content for model training. ChatGPT-User is what visits a page when a person asks ChatGPT to look at a link. OpenAI's own bot documentation recommends allowing OAI-SearchBot and says plainly that sites opted out of it will not be shown in ChatGPT search answers. It also notes that ChatGPT-User is not used to decide whether content may appear in search, and that a robots.txt change can take around 24 hours to register.

Anthropic splits it the same way. Its crawler support article lists ClaudeBot for training, Claude-User for fetching pages in response to a user's question, and Claude-SearchBot for indexing content to improve search results. Blocking Claude-User, it says, prevents retrieval of your content in response to a user query and may reduce your visibility in user-directed web search.

The useful consequence: you can refuse training and still be citable. Allow the search and user agents, disallow the training crawler if that matters to you, and the two decisions do not interfere. Check your live robots.txt before doing anything else in this post, because none of the rest matters if the fetch fails.

The citation data is better news for a small site than it looks

Headline figures from this research get quoted as proof that only big publishers get cited, and they show close to the opposite. In a sample of roughly 700,000 ChatGPT conversations containing at least one web citation, collected from United States English-language users between October and December 2025, the AI-visibility vendor Profound found that the ten most-cited domains accounted for only 12% of all citations. Wikipedia, the single largest source, took about 5%. Reddit took about 3%.

Read that again as an opportunity rather than a ranking. Nearly nine citations in ten went to sites outside the top ten, which is a long tail of specific pages answering specific questions. Compare that with a search results page, where the top three organic positions take most of the clicks. The distribution of citations is flatter, and flat distributions are where small sites live.

The same research puts Wikipedia in roughly 1 in 6 conversations that cite anything, so the giants are not absent. They are just not crowding the field the way they do in search. Profound sells AI-visibility tooling, so treat the figures as vendor research rather than independent measurement, but the shape of the distribution matches what anyone watching their own referral logs sees.

What makes a page the obvious thing to quote

Assistants reach for pages that resolve an ambiguity. That is the whole selection criterion, and it explains why thin general content never gets picked no matter how well optimised it is.

Numbers you own are the strongest material. What you charge, how long something took, what a tool cost you last month, how many of a thing you have shipped. Nobody else can publish your numbers, so a question that needs them has exactly one answer on the web.

Procedures with real values beat procedures in the abstract. “Point your domain at your host” is in a thousand articles. The four records you actually entered, with the field names from the screen in front of you, is in one. The same goes for errors: the exact string a tool prints, with the fix that worked, is a page people and models both land on.

Dated specifics work because models get rewarded for currency. Claude's search results carry a page age field, and a page that says what changed in a named month reads as a live source rather than an undated essay. If you are starting from nothing on this, the fundamentals in our beginner guide to SEO for a new site still apply, because an assistant that cannot find you in a search index cannot cite you either.

What gets quoted is a sentence, not your page

This changes how you write a paragraph. In Anthropic's web search tool, the quoted span attached to a citation is capped at 150 characters. Whatever the model pulls out of your page, that is roughly the size of it: one claim, standing alone, without the two sentences of setup you put before it.

So state the answer in a single self-contained sentence, early, in the plainest words you have. Then explain it. Writing that builds slowly towards a conclusion reads well to a person and extracts badly to a model, because the extractable unit never appears. Hedged sentences fail the same way: a claim wrapped in three qualifiers cannot be lifted out intact.

That is as far as I will take the formatting question here, because page structure is its own subject and deserves its own treatment. The point for now is that quotability is a property of your sentences, not of your markup.

The mentions you do not control matter more than yours

Here is the uncomfortable part. A noticeable share of what assistants quote is other people discussing you, not you describing yourself, and that share is heaviest on exactly the platforms you do not own. Profound's cross-platform data has Reddit as the leading source for both Google's AI Overviews and Perplexity, at 2.2% and 6.6% of all citations respectively.

You cannot manufacture that honestly, and attempting it is both against those platforms' rules and easy to spot. What you can do is be genuinely present where your subject is discussed, under your own name, answering questions without pitching. Over a year that produces a trail of real mentions. Over a week it produces nothing, which is why it gets skipped.

Corroboration is the second-order version. When a model finds the same claim about you on your site and on two places that are not your site, the claim gets used. When it exists only on your own marketing page, it tends to get described as a claim rather than repeated as a fact. Getting written about anywhere real, including small newsletters and niche forums, is worth more here than another page on your own domain. Our notes on getting traffic to a new site cover the same ground from the visitor side.

How to check whether any of it is working

Start with your server or host logs and look for the user agents by name: OAI-SearchBot, Claude-User, Claude-SearchBot, GPTBot. Their presence tells you the fetch is permitted and happening, which is the only part of this you can confirm directly. Their absence after a few weeks is the signal to re-read your robots.txt.

Then ask the questions yourself. Open a fresh conversation, ask the thing your page answers, and watch whether a search runs at all. Often it will not. In the same Profound sample, about 18% of ChatGPT conversations triggered at least one web search, and the rate of citations was highest on the first turn of a conversation and fell steadily after it. Most answers never touch the live web, so most of the time there is nothing for you to win.

Referral traffic is the weakest signal, and worth setting expectations about. Assistant referrals show up in analytics as a trickle rather than a channel, often single visits, because a cited page is read inside the answer rather than clicked. Judge this work by whether you are being fetched and quoted, not by a traffic line that may never move much.

Things sold as the answer that are not

No submission form exists. Neither company takes a request to be included in answers, and anyone offering to submit your site to ChatGPT is selling a fiction. Paid placement in an organic citation is not a product either.

Schema markup on its own does not do it. Structured data helps a machine parse what is already there and cannot make an unremarkable page worth quoting. The same applies to publishing volume: thirty general posts are easier to ignore than three specific ones.

Rewriting old posts to sound more confident does nothing, and is quietly harmful if it strips out the qualifications that made them accurate. Models are not scoring your tone. They are looking for a claim they can attach a URL to without being wrong.

One more, since it comes up with AI-built sites: there is no penalty here for having used AI to build the page. That question is separate and we went through the evidence in whether Google penalises AI-generated websites. What matters to a citation is whether the page says something true and specific, whoever typed it.

What a mention is actually worth

Modest, and worth doing anyway. One citation in one answer to one person is not a traffic event, and treating it as the new SEO oversells it badly.

What accumulates is different. Being the page an assistant reaches for on a narrow question means you are the default source on that question for the subset of people who ask it that way. That is a small, durable position, and it is available right now in a way that ranking first for a competitive search term is not.

The work is also not wasted if the citation never arrives, which is the real argument for it. Allowing the right crawlers, stating answers plainly, publishing your own numbers and being present where your subject is discussed are all things that pay off with human readers too. That is not true of most tactics sold under the AI-search banner.

If you do one thing after reading this, open your robots.txt and check that OAI-SearchBot, Claude-User and Claude-SearchBot are not blocked. That single file decides whether any of the rest is possible, and a surprising number of sites built in the last two years are blocking all of them by inheritance from a template nobody read.

After that, pick the one question you can answer better than anyone because you have done the thing, and answer it in a sentence someone could lift out whole. Then keep the page accurate. Accuracy is the only durable advantage in this, since a model that quotes you wrongly once has no reason to come back, and a page that is right about a narrow thing keeps being the right page to quote. Keeping it that way is a maintenance habit more than a writing one, which we laid out in keeping an AI-built site maintainable.

Related: What Is AEO and Does a Small Site Need It?

Common questions

How do I get my website cited by ChatGPT?

Allow OAI-SearchBot in your robots.txt, since OpenAI states that sites opted out of it will not appear in ChatGPT search answers, then publish pages that state a specific claim plainly enough to quote in a sentence. Citation happens when ChatGPT runs a live search, so the page needs to be both fetchable and the most direct answer to a narrow question.

Is there a way to submit my site to ChatGPT or Claude?

No. Neither OpenAI nor Anthropic offers a submission form, an index to join, or paid placement in an organic citation. The only controls a site owner has are the robots.txt directives that allow or block each named crawler, plus the content of the pages themselves. Anyone selling submission to an AI assistant is selling something that does not exist.

Should I block AI crawlers on my site?

It depends which ones, and the decisions are independent. Blocking GPTBot or ClaudeBot keeps your content out of model training. Blocking OAI-SearchBot, Claude-User or Claude-SearchBot keeps you out of the answers those assistants give when they search the web. Most small sites that want the mentions should allow the search agents and can still refuse the training crawlers.

How long does it take to get mentioned by an AI assistant?

There is no reliable timeline, but the retrieval path is fast once access is in place: a page that is indexed and fetchable can be quoted within days, because the search happens at the moment of the question. OpenAI notes that a robots.txt change can take around 24 hours to register for search. Getting into a model's training memory is a different matter and is not something a new site can plan for.

Does being cited by ChatGPT actually bring traffic?

Not much, usually. Assistant referrals tend to arrive as a trickle of single visits rather than a traffic channel, because the cited page is often read inside the answer instead of being clicked. The value is positional rather than volumetric: you become the default source on a narrow question. Judge the work by whether crawlers are fetching you and answers are quoting you.

Do I need schema markup to get cited by AI?

It helps a machine parse a page that already says something worth quoting, and it does nothing for one that does not. Structured data is a formatting aid rather than a cause of citation. Spend the first hour on crawler access and on writing one plainly stated, self-contained claim per section, then add markup afterwards if you want it.