How do you get cited by ChatGPT?

Only one of OpenAI's crawlers decides whether ChatGPT can cite you: OAI-SearchBot. OpenAI documents four user agents; three of them matter to a publisher (the fourth, OAI-AdsBot, only visits pages submitted as ads), and the crawler documentation is explicit about which one matters for search. OAI-SearchBot builds the index behind ChatGPT search; sites that opt out of it "will not be shown in ChatGPT search answers, though can still appear as navigational links". GPTBot crawls for model training and has no bearing on whether you are cited. ChatGPT-User fetches a page when a user's action calls for it, but OpenAI says it "is not used to determine whether content may appear in Search" and that robots.txt rules may not apply to it at all.

So the earlier advice, ours included, that you must allow both the search agent and the user agent was wrong: the decision rests on OAI-SearchBot alone.

The robots.txt file is not the whole gate. OpenAI also asks that your host or CDN (the service in front of your site that filters traffic) allow requests from the IP ranges it publishes for OAI-SearchBot, which is where sites with a clean robots.txt still lose: an anti-bot rule at the CDN blocks the crawler before it ever reads the file. After you change robots.txt, OpenAI says its systems take roughly 24 hours to adjust, so checking the same evening proves nothing.

Those two conditions are the whole eligibility list on OpenAI's search help page, and its publisher FAQ names nothing else a public website must do; there is no submission form (more on that in the FAQ below). The publisher FAQ does add a detail worth knowing if you block pages on purpose: when OpenAI learns a disallowed URL from a third-party search provider or from crawling other pages, it may still show the link and title in ChatGPT Atlas, OpenAI's browser. Full exclusion needs a noindex meta tag, and the crawler must be allowed to read the page to see it. For the complete list of AI crawlers and ready-made configurations, see our guide on robots.txt for AI crawlers.

ChatGPT's search retrieval works on the HTML your server returns: it extracts text from that response and does not wait for a client-side framework to fetch data and render it. OpenAI does not publish how much JavaScript its search crawler executes, and in practice content that appears only after the browser has run your JavaScript is routinely read as an empty page. Treat this as a working constraint, not a documented rule, and test it rather than trust it.

The test for ChatGPT citations costs nothing: open the page with JavaScript disabled, or request it with curl, and check that the sentence you want quoted is in the markup. While you are there, check the rest of what a crawler sees: a 200 status without a redirect chain, no noindex on pages you want cited, and one canonical host, meaning a single hostname (with or without www, https only) that every internal link and every external profile points to. A brand split across two hosts is two half-brands to a retrieval system.

Every guide on this site is rendered on the server and can be read with JavaScript disabled. This gives crawlers access to the content before the work on individual pages, independent coverage and brand information described below.

A page ChatGPT quotes has one buyer question as its heading, a 40–60-word answer with a checkable fact directly beneath it, and the detail after. To see why, start with how ChatGPT chooses sources: it searches rather than recalls. OpenAI's help documentation says that when ChatGPT hands a question to a search provider it typically rewrites it into one or more targeted queries, may send further, more specific queries after reading the first results, and sometimes partners with other search providers. It also says plainly that ChatGPT ranks results "using multiple factors" and that placement is not guaranteed.

The rewriting step is what your page has to survive. OpenAI's own example is a broad question about a drug programme being turned into a query naming the specific compound and the year: the model is looking for a specific fact, and the page that states it plainly beats the best general article on the topic. Our September 14, 2026 fintech run below shows the same preference from the other side: the pages ChatGPT cited were the ones that owned the fact, not the ones that discussed the subject.

Google ranking is not the filter people assume it is. Ahrefs looked at the 1,000 pages ChatGPT cited most often (October 2025) and found that 28.3% of them had no organic keywords in Ahrefs' index. Of the 564 pages with a determinable date, 76.4% had been updated within the previous month. Two conclusions for your pages: you do not need to rank first in Google to be quoted, and the pages ChatGPT cites tend to carry a recent, visible date next to specific claims.

To get ChatGPT to cite a page of your website, the pattern that survives extraction is simple: a heading phrased as the question a buyer actually asks; immediately under it, a compact answer of 40–60 words that names the subject ("a UK business account at Starling", not "it" or "this option") and states the fact with a number, a date or a named source; then the detail. One page answers one question; a page that answers six is rarely retrieved for any of them.

Our own run data shows what this looks like from the model's side. On September 14, 2026, we put three fintech buyer prompts to ChatGPT in signed-out sessions (a UK business account, a neobank for freelancers in Spain, the cheapest UK-to-India transfer), three repeats each, nine answers in total. The providers' own pages were cited in all nine, and the pages the model picked were not "About us": Starling's and Monzo's published service-quality survey results, and N26's account terms for self-employed customers in Spain. Pages with a checkable fact for a specific segment. To see which of your pages ChatGPT quotes today, run the signed-out test protocol from our guide on checking your AI visibility for free.

Commercial prompts go to third-party pages because a model recommending a vendor wants corroboration: independent sources that describe the vendor the same way the vendor describes itself. For a question like "which agency" or "which tool", your own site is a claim; a directory, a review platform or a comparison article is evidence. Ahrefs quantified this for ChatGPT in December 2025 across 750 "best X" prompts and 26,283 cited source URLs: "best X" blog lists were 43.8% of all page types cited.

The same September 14 run adds a nuance that matters for planning. Across the 24 signed-out ChatGPT answers on GEO-buyer prompts, the type of question decided the type of source. On prompts about what happens if GEO fails after 90 days, or what guarantees agencies offer (six answers), the sources were agency websites: 18 different domains, and only one of them, citant.ai on the guarantee prompt, appeared in all three repeats. On prompts about doing the work in-house versus hiring (12 answers), agencies vanished: developers.google.com was cited in 12 of 12 answers, alongside OpenAI's help pages and arXiv, and agency sites in none. So a commercial prompt is won on third-party pages, while a how-to prompt is won by a page that reads like documentation: sourced, dated, specific, and free of sales copy.

Local services add a layer of their own. When a prompt names a service and a city, ChatGPT can put a map of local businesses above the answer and name the firms on its cards. In our study of 72 answers to lawyer searches, run on 22 September 2026, it did so in 11 of 12 first-run answers, while the pages it cited were mostly directories and bar associations: firm websites made up 8 of its 71 links. For a local business, the listing belongs to the third-party layer, and it is the first thing a client sees.

This is the slowest layer and the one you control least. Getting into the directories and comparison pages that are already being cited for your category, keeping review profiles current, and being described with the same name, the same one-line positioning and the same facts everywhere you appear all take quarters, not weeks. Paying for placements in "best of" lists is a separate decision with its own trade-offs, which we work through in our guide on whether to buy listicle placements. If a competitor is already named for your prompts and you are not, the diagnostic order is in our guide on why ChatGPT doesn't recommend your company.

If your goal is to get your business recommended on ChatGPT, distinguish that from getting an article cited. A citation can support an explanation without endorsing the company that published it. Test questions describing a customer need without naming your brand, then record recommendations and source links separately. For a business that is still establishing its offer, our guide to GEO for new businesses explains where to start.

Before a model can recommend you, it has to resolve you as an entity: one organization with one name, one description and a set of independent places where it also appears. Organization markup in JSON-LD (a block of structured data on the page that tells machines who owns the site) on the canonical host, an identical name and one-line description on every profile, and a sameAs list that points at real external profiles (a company registry, LinkedIn, Crunchbase, Clutch, a Wikidata item if you qualify) all reduce ambiguity. A sameAs list that names only your own homepage adds nothing; the point is to link the entity to places you do not control.

What this looks like from the retrieval side: on September 15, 2026, we re-ran the four-question set from our Hindi GEO agency case, signed out, two repeats each. GET-GEO.AI was named in eight of eight answers, and on the one prompt we had lost on September 3 it came back at #4 and #5, with the case page itself cited as the source. Across the eight answers the model drew on the homepage, the case page and two different guides: it was drawing on the organization, not on one lucky URL. One assistant, one day, our own prompts: these are the first reproducible citations, not a stable position, and the next re-measurement is scheduled rather than assumed.

These five types of work have different timescales: about a day for access changes, weeks to a quarter for consistent brand information once profiles exist, and several quarters for independent coverage. Our timeline guide explains the estimates. The table below sets out the actions and checks.

If you want the first layer checked and the first ten prompts run for you, the free AI visibility audit does exactly that: ten prompts in one language on which you are invisible today, the current answers, a crawler-access check, and the specific edits that would change the picture.