How to get cited by ChatGPT: what OpenAI says and what you control

How does ChatGPT decide what to cite?

ChatGPT only cites when it searches, and it does not search for every question. OpenAI’s help page “Searching the web with ChatGPT” says it may search automatically when a question would benefit from current information. An answer from the model’s memory has no sources behind it to link to, so the first thing to understand is that your business can reach a person through ChatGPT in three separate ways.

When ChatGPT does search, OpenAI says it typically rewrites the question into one or more targeted queries sent to search providers. Its example is a researcher’s broad question about a drug class becoming a narrower search string. Its help page links the privacy policies of Microsoft and Shopify as providers, and adds that the approximate location from your IP address can shape local results.

Then OpenAI gives the only sentence it publishes on ranking: “ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed.” It does not list the factors, so the exact formula is not public.

What does OpenAI actually require?

Two things, both in its own documentation:

• Allow OAI-SearchBot in robots.txt. OpenAI says sites opted out of OAI-SearchBot “will not be shown in ChatGPT search answers, though can still appear as navigational links”. A change to robots.txt can take about 24 hours to take effect.
• Let its IP addresses through. OpenAI says to confirm that your host or content delivery network allows traffic from its published searchbot IP addresses, listed at openai.com/searchbot.json.

The second one is where sites fail without knowing it. robots.txt is a file you can read in a browser, but a firewall rule only shows up to the bot. When we tested the MWY checker on 36 business websites in seven countries on November 3, 2026 (a hand-picked sample, not a market figure), 4 refused the checker itself. Of the 32 we could read, 1 blocked an AI search bot in robots.txt, while 4 refused a request sent under the names of AI bots. That second number is only a signal, because the real bots arrive from their own addresses, but it shows where to look first.

GPTBot is a separate question. It collects material that may be used to train models, and OpenAI says each setting is independent: you can allow OAI-SearchBot and block GPTBot, and you stay eligible for search. Blocking GPTBot on purpose is a reasonable choice, and it does not take you out of ChatGPT search. Our guide to AI crawlers lists these bots with the exact robots.txt lines.

How do you check that ChatGPT can reach your site?

Do these in order. The whole check takes about half an hour:

• Read your robots.txt. Open yourdomain.com/robots.txt. Look for “Disallow: /” under “User-agent: *” or under OAI-SearchBot. Lists of “AI bots to block” copied from elsewhere often include the search bots by accident.
• Check the firewall settings. In Cloudflare, look at the settings for AI bots and AI Crawl Control. With another host, ask support whether OAI-SearchBot and ChatGPT-User are challenged or blocked.
• Search your server logs for the bot’s name. Requests from OAI-SearchBot with normal “200” responses are the best evidence you can get. Rows of 403 or 429 point to a block.
• Check your Bing listing. OpenAI links Microsoft as a search provider, so make sure your main pages are indexed in Bing as well. Search Bing for “site:yourdomain.com”, and set up Bing Webmaster Tools if the result is thin.
• Read the raw HTML. View the page source of your main pages and look for your prices and services in the text. If they appear only after scripts run, a bot that does not run scripts may never see them.

What makes a page worth citing?

OpenAI publishes no list of what earns a citation, so this part rests on indirect evidence. Three sources point the same way.

The GEO research paper, posted in 2023 and accepted at KDD 2024, rewrote pages in a test environment and found that adding quotations, statistics and source references raised their use in AI answers, while keyword stuffing did not help. Its tests were on a research engine built on GPT-3.5 and partly on Perplexity, not on today’s ChatGPT, so read it as a direction and not a guarantee.

Microsoft’s guidance from October 2025 says assistants break content into pieces and pick the ones that stand alone. It recommends question headings, lists and tables, and warns against answers tucked into tabs or expandable menus.

Google’s guidance from May 2026 asks for “non-commodity” content, meaning first-hand experience and specific knowledge rather than a restatement of what others have written. That is about Google, but a page only one company could have written is a better source for anyone.

In practice, a citable page tends to do these things:

• Answers the question in its first sentences. Say the answer, then explain it.
• States specifics a summary can use. “Crowns take two visits, about two weeks apart” can be quoted. “We pride ourselves on quality” cannot.
• Publishes the number people ask about. Price ranges, lead times, service areas, minimum orders.
• Says who wrote it and when. An author and a date help a reader decide whether to trust it, and OpenAI tells users to check when a source was published or updated.
• Covers one thing properly. A page that fully answers one question gives a clearer match than a page that touches ten.

Our guide to answer engine optimization goes further into how to write these passages.

Which questions can your site realistically win?

Not all questions are open to a small website, and it saves time to know which ones are.

Questions about your own category with a specific angle are the most winnable, because few pages answer them well. Broad “best of” questions tend to be answered from comparison pages, so the practical route is getting covered by those pages honestly, rather than publishing your own.

Reservations show how partner data can matter more than your own site: ChatGPT’s help page says restaurant availability may come from OpenTable, Resy or Yelp. For a restaurant, being bookable on one of these may do more than any page you write.

What is a waste of time?

Some popular advice has little behind it, and some can backfire.

• Writing llms.txt to get cited. OpenAI’s crawler documentation does not mention it. It takes an hour and may help other tools, as our llms.txt guide explains, but it should not be your plan.
• Publishing one page per question variation. ChatGPT rewrites questions anyway, so an exact match buys little. Google calls the same tactic scaled content abuse when it is done to manipulate its results.
• Buying or faking mentions. Google’s guide lists inauthentic mentions among the things that are not as helpful as they seem, and fake praise does nothing for a customer who reads it.
• Keyword stuffing. The GEO paper tested it and found no gain.
• Treating ChatGPT ads as citations. OpenAI runs ads on ChatGPT and has a separate bot, OAI-AdsBot, to check landing pages. An ad is a paid placement and not a source in an answer.

How do you find out whether you are being cited?

There is no report from OpenAI that tells you this, so you have to look yourself.

• Write 10 to 20 real questions. Use the words customers use, including your city where it matters. Include questions about your category that do not name you.
• Ask them in ChatGPT and read the sources. Record whether you are named, whether you are linked, and who is cited instead. A single answer proves little, because answers vary between tries.
• Repeat monthly with the same list. A trend over several rounds is far more useful than one result.
• Check the facts. OpenAI itself warns that citations “can be incomplete, outdated, or incorrect”. If an answer gets your hours or prices wrong, look at which source it cited and correct that page.
• Read your analytics. Look for referral visits from chatgpt.com, keeping in mind that many visits from app users can arrive without a clear source.

Our guide to checking AI visibility sets out this routine in detail.

What if ChatGPT gets your business wrong?

Wrong hours, an old address or a service you stopped offering can do more harm than being left out. OpenAI’s search and crawler pages describe no way for a business to correct an answer, so the correction has to happen at the source.

• Find the source. Open the citations in the answer. OpenAI itself tells users to open a cited source and check that it supports the answer.
• Fix what you control first. If the source is your own site, change it, and look for old pages too. A price in a blog post from years ago is still a page that can be read.
• Then fix the listings. Directories, review sites and map listings with old details. Keep the name, address and phone the same everywhere.
• Ask again with a push. OpenAI’s advice to users is to ask ChatGPT to search again using a specific source, date or location. Try naming your site in the question. If that returns the right fact but the plain question does not, the page can be read, and the issue is selection rather than access.
• Watch for answers with no sources. An answer with no citations may come from the model’s memory. A fix on your site may take a long time to reach it, perhaps not until a later model, and OpenAI publishes no timing.

A worked example: a dental practice

An illustrative case: a two-dentist practice in a mid-sized US city, invented here to show the order of work.

The owner asks ChatGPT “how much does a crown cost in [city]” and sees two insurance sites and a competitor cited. The practice’s own site says only “contact us for pricing”. Its robots.txt is fine, but its host’s security plugin blocks unknown bots, so OAI-SearchBot gets a 403.

The owner has the plugin set to allow verified search bots, then adds a page section with a real price range, what changes the price, how many visits it takes and what insurance typically covers, signed by the dentist with a date. Six weeks later they ask again. The answer is still not guaranteed to change, but the reason the practice could not be cited has been removed, and they now have a page that a person can use on its own.

What the MWY checker can tell you about citations

The MWY checker tests the conditions on your side of a citation. It checks robots.txt for OAI-SearchBot and the other AI search bots, asks your site for a page under their names to spot firewall refusals, reads how much text is in the raw HTML, and looks at business schema, contact details and whether pages answer real questions.

It cannot tell you why ChatGPT chose another source for a given question, and it does not promise that you will be cited. OpenAI’s own words apply: placement is not guaranteed. What the checker can show is whether something on your site is standing in the way.