How to get your business cited by ChatGPT

A page is more likely to be cited by ChatGPT when OAI-SearchBot can fetch it and the page states the relevant answer in plain text. Give the crawler access, then make the product, audience, price and alternatives easy to identify.

That is the short answer. The rest covers what to publish, where citations currently come from, how to check the technical details, and what the available data cannot prove.

A citation starts with access

A page cannot become a ChatGPT citation if the search crawler cannot fetch it. OpenAI says that “OAI-SearchBot is used to surface websites in search results in ChatGPT’s search features” in its documentation on bots and crawler controls. The same page recommends allowing OAI-SearchBot in robots.txt and allowing requests from its published IP ranges.

This is separate from GPTBot. OpenAI says the settings are independent, so you can allow OAI-SearchBot for search while disallowing GPTBot for training. That distinction matters if your team has blocked AI crawlers as a single category. OpenAI says sites opted out of OAI-SearchBot will not appear in ChatGPT search answers, although they can still appear as navigational links.

Check the actual production robots.txt, not the file in a staging branch. Check the response from the public hostname, too. OpenAI says it can take about 24 hours for its systems to adjust after a robots.txt update, so an immediate retest may tell you more about timing than visibility.

Access is necessary, not sufficient: it gets the crawler to the page, not to an answer worth citing.

Plain answers beat positioning

The main failure in our sample was not crawler access. It was the page failing to say, in ordinary words, what the business is, who it is for, what it costs, and what it is measured against. That is the practical requirement behind the whole exercise: a page needs sentences that answer a buyer's question without making a model infer the product from slogans or screenshots.

Write the answer first. Add the proof and detail after it. A useful product statement has a subject, audience, use case, price context and comparison point. Use your actual facts in a pattern such as: “[Product] is for [audience] who need [use case]. Plans start at [price]. It replaces [alternative] but does not do [X].”

Put the statement in visible HTML. Don't put the only version of the answer in an image, a client-side component or a hard-to-fetch transcript. Content that requires JavaScript can turn a clear answer into an absent one, so render the important copy in the initial page response where possible.

Metadata cannot replace the statement

Canonical tags, Open Graph data, titles, descriptions, sitemaps and schema remove ambiguity. They do not replace a direct explanation of the product. Keep the canonical URL stable, expose the important content without JavaScript, publish the sitemap and use schema where it describes the page.

Read the page as a buyer, not as its author. A page with valid schema and no clear statement gives a crawler correct labels attached to an answer that still doesn't exist. The homepage is a useful test because it carries the broadest claim, but a clear homepage still needs supporting pages for price, use cases and alternatives.

The technical checks are still worth doing. They catch pages that are difficult to fetch, poorly labelled or easy to misidentify. Treat them as foundations for useful content, not as a substitute for that content.

Citation share follows page type

Getting cited does not mean relying only on your own product page. Across CiteGraph's recorded sample, listicles and roundups made up 67.5% of citations, or 30,585 citations. Competitor product pages made up 24.4%, or 11,052 citations. Together, those two page types accounted for 91.9% of citations.

The product's own site accounted for 1.5%, or 663 citations. Documentation accounted for 1.1%, or 481 citations. Community threads accounted for 3.4%, or 1,534 citations. The sample also included social posts, review platforms, YouTube, news and Wikipedia, each with smaller shares.

That pattern changes the work. Improve your own pages so they can be understood and cited, then inspect the pages that already compare products in your category. A comparison page on your site may answer a buyer's question, but a roundup or independent product page may still be where ChatGPT finds the usable source.

Do not read the share as a promise that publishing a listicle will produce a citation. It describes where citations came from in this sample. It tells you where to inspect existing answers, not where to buy an outcome.

Buyer questions need their own pages

A product page answers “what is it”. Buyers also ask whether it fits a use case, what it costs, how it compares and what the alternatives are. Give each important question a page or a clearly labelled section, with one direct answer and the conditions that change it. Keep the wording concrete enough that a passage can stand alone when ChatGPT links to it.

State who should choose the product, who should not, what it replaces, what it does not do and where the price sits. A page that only describes features has left the buyer to do the comparison themselves. Run a comparison or use-case page through the free AI readiness check before publishing it, then use the free AI visibility checker to see whether it actually surfaces for the phrasing a buyer would use.

The tools answer different parts of the same problem: readiness asks whether the site can be used, while visibility asks whether it is being found. Use both on the pages that matter most, especially pages covering price, use cases, comparisons and alternatives. Keep the answer on the page rather than assuming a link to another section will supply the missing context.

A site audit shows the gap

Here is the worked example, using the readiness results rather than an invented company. Start with the audit size: 120 sites were checked. Then check access: 115 passed the AI crawler check and 5 failed. A blocked site has an immediate delivery problem, but access alone does not show that a site can answer a buyer's question.

Next, check the plain statement. 20 sites passed, 43 failed and 55 received a warning. Read that alongside the access result: most audited sites were available to the crawler, while the statement check produced the larger unresolved group. The useful diagnosis is therefore not “allow robots and stop”; it is “allow robots, then make the page say what the business is, who it serves, what it costs and what it is against”.

Now check the buyer-intent pages. 54 sites passed the comparison or alternatives check, 20 failed and 46 received a warning. The questions-answered check had 56 passes and 62 warnings, with no outright failures recorded. A site can therefore pass the access step, fail to make comparisons explicit and still leave many common questions only partly answered.

Finally, inspect the citation signal. The check for the site's own pages getting cited passed on 5 sites, failed on 10 and warned on 5. That check does not cover all 120 sites in the same way, so do not turn it into a market-wide citation rate. It is a diagnostic signal inside this audit, not proof that a technical pass creates a link.

The repair order follows the evidence: allow the crawler, write the plain statement, answer buyer questions, add comparison and alternatives pages, then test the resulting pages against live questions.

The sample cannot predict your site

We cannot tell you that any particular page will be cited, how many citations it will receive or how long a citation will remain. Our larger scan data records 45,309 citations across 7,260 AI answers, from 60 scans of 41 sites in 44 categories. Those scans asked 442 distinct buyer questions and found 7,624 distinct domains and 16,506 distinct pages.

That is a sample, not a census. The readiness sample is 120 sites; the citation sample is a separate set of scans. Neither is a census of the web, and we have not audited your site specifically.

We also cannot use the page-type shares to separate cause from correlation. Listicles may be cited often because they answer broad comparison questions, because those questions were common in the scans or because of other factors the data does not isolate. The data supports an inspection plan. It does not support a recipe with a guaranteed result.

The next question to ask is: “Which buyer questions do our pages fail to answer?” The answer lives in the free AI visibility checker and our weekly research on who owns AI answers.