How to rank in ChatGPT: how it picks sources, what gets cited, and a 12-point checklist.

1. How ChatGPT search picks and cites sources.

ChatGPT answers in two modes. Most prompts are answered from what the model learned in training, with no web lookup and usually no links. The rest trigger a web search, and the answer comes back with sources. Semrush analyzed more than a billion lines of US clickstream data and found that ChatGPT ran a search on 34.5% of prompts in February 2026, down from 46% in late 2024. When people say SEO for ChatGPT, or ChatGPT search optimization, they mostly mean that second mode: the answers with links in them.

When ChatGPT does search, four things happen in order.

1. It decides to search. ChatGPT makes the call from the prompt, or the user turns search on. Questions about prices, products and recent events are typical triggers.
2. It rewrites the prompt. ChatGPT's help pages say it may turn a prompt into one or more targeted search queries. Promptwatch, which logs the live interface, measured about 1.8 searches per answer after August 8, 2026. These rewritten queries, often called fan-out queries, are what your page has to match. The words the user typed are not.
3. It retrieves candidate pages. The queries go to ChatGPT's own index and to outside search providers, and the model reads what comes back.
4. It writes the answer and cites a few. Since a model update on March 4, 2026, Promptwatch counts 4.7 to 4.9 sources per answer, down from 6.4 before it.

The cut between step three and step four is severe. In a July 2026 test by SEO consultant Suganthan Mohanadasan, ChatGPT retrieved 3,554 pages across 57 conversations and cited 110 of them, about 3.1%. Being retrieved gets you into the room. Being quoted is the job.

So there is no list of ten blue links to climb. ChatGPT's help page says results are ranked using multiple factors meant to surface relevant, reliable information, and adds: “Placement is not guaranteed.” In practice, to rank in ChatGPT means two things: your page is one of the roughly five sources under the answer, and for commercial prompts your brand is one of the names in it. This guide is about earning both. If you came here to use ChatGPT as a tool for keyword lists, briefs and drafts, that is our ChatGPT for SEO guide. The wider discipline, across every assistant, is LLM SEO.

2. Does ChatGPT use Google or Bing?

Both show up, and neither is the whole story. ChatGPT search runs on an index its maker builds with its own crawler, OAI-SearchBot, plus results from outside search providers. Its help pages say search uses third-party providers and content supplied directly by partners, and they name Bing as one of the providers. Google is the undocumented part. Here is what is on the record.

• April 2025. At Google's antitrust remedies trial, ChatGPT's head of product Nick Turley testified that the company asked to use Google's search API in July 2024 and was turned down that August. He also said ChatGPT was years away from answering 80% of queries with its own search technology.
• August 2025. The Information reported that ChatGPT was partly using Google results scraped by a company called SerpApi, for current topics such as news and sports.
• July 2026. A federal judge dismissed Google's copyright-law suit against SerpApi, with leave to amend.
• October 2026. Peec AI published an analysis of the data ChatGPT's interface receives with each answer. Results carried a source label: an in-house index, plus outside suppliers. Peec counted at least eight outside sources, including Google results, Bing, Yelp and TripAdvisor.

What matters more than the plumbing is how little the output resembles a Google results page. Ahrefs ran 15,000 long-tail queries in 2025 and found that about 8% of the pages ChatGPT cited were in Google's top 10 for the same query. The figure for Bing's top 10 was also about 8%. A second Ahrefs study, of 3,311 short keywords, found that 10% of the pages ChatGPT cited were in Google's top 10, but 31.8% of the domains it cited were. The cause is step two above: ChatGPT searches its own rewritten queries, so it is reading different results pages from the one you would check by hand.

Read that both ways. A number 1 Google ranking does not earn the citation. And you cannot skip search engines either, because every route into ChatGPT's candidate set runs through an index: its own, Bing's, or Google's. Be crawlable and indexed in all three, then compete on the page. Google's own AI answers work from a different pipeline, which we cover in the Google AI Mode guide, and the AI search engines roundup compares the assistants side by side.

3. Let the right ChatGPT crawlers in.

ChatGPT's maker runs four separate agents, and they do four separate jobs. Mixing them up is how sites disappear from ChatGPT by accident.

The crawler documentation is explicit about the first one: “Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links.” GPTBot is the training crawler, and blocking it does not remove you from search. ChatGPT-User is not a crawler at all. It fetches a page when a person's prompt calls for it, and the documentation says it plays no part in deciding what appears in search. A robots.txt that keeps you eligible looks like this:

# ChatGPT search: required to be cited
User-agent: OAI-SearchBot
Allow: /

# Pages a person asks ChatGPT to open
User-agent: ChatGPT-User
Allow: /

# Model training: your call, does not affect search
User-agent: GPTBot
Allow: /

Sitemap: https://www.example.com/sitemap.xml

Whether to let GPTBot in is a real choice. Blocking it keeps your pages out of future training. Allowing it lets your pages be part of what future models learn, which matters because most answers come from the model's memory with no search at all.

Three more checks sit outside robots.txt.

• Your CDN and firewall. ChatGPT's help page says the host or content delivery network has to allow traffic from the search crawler's published IP addresses. A “block AI bots” switch in a CDN or security plugin can block OAI-SearchBot along with the training crawlers. Check your server logs: OAI-SearchBot requests should return 200, not 403.
• The 24-hour lag. The documentation says a robots.txt change can take about 24 hours to register. Do not test five minutes after you deploy.
• Text in the HTML. Keep answers, prices and specs in the server-rendered HTML. Content that only appears after JavaScript runs, or that sits in an image or a PDF, is a risk you do not need to take.

Claude runs the same split of training, search and user-triggered bots, and our ClaudeBot guide has the matching robots.txt. If you want to audit crawl access properly, work through the crawl checks in our technical SEO audit checklist, because the faults that keep Googlebot out (blocked paths, broken rendering, server errors) usually keep ChatGPT's crawler out too.

4. How to get cited by ChatGPT: what makes a page citable.

Once your page is in the candidate set, the model decides what to quote. Nobody outside the company has the formula. What we do have is evidence about the pages that get picked, and it points at six properties.

• The answer is in the first two sentences. The model is looking for a passage it can lift to answer one specific rewritten query. A heading phrased like the question, with the answer directly under it, is the easiest passage to lift. That is the core of answer engine optimization.
• The facts are specific and sourced. The GEO benchmark presented at the KDD conference in 2024 tested content edits against generative engines. Adding statistics, quotations and cited sources raised a page's visibility in generated answers by up to 40%. Keyword stuffing was among the weakest tactics tested. The generative engine optimization guide walks through that study.
• The page is current. The set of pages ChatGPT cites turns over fast. Green Flag Digital tracked the 215 finance pages ChatGPT cited most in October 2025. By September 2026, 107 of them had no citations at all. Show an updated date and keep the numbers behind it true.
• It is the right type of page. In Promptwatch's July 2026 data, product pages made up 32.8% of ChatGPT citations. Listicles were 9.7%, news articles 5.2% and how-to guides 4.1%. A blog-only plan leaves the largest category untouched: your product, pricing and comparison pages need the same care as your articles.
• It answers one question fully. Because ChatGPT searches rewritten sub-queries, a page that settles one narrow question beats a long page that brushes ten. Give pricing, each use case and each comparison its own URL.
• It uses names, not descriptions. Write the product, company and category in full, so a sentence still makes sense when it is quoted alone. “Our platform” means nothing outside your site.

None of this is exotic. It describes a good reference page. The hard part is volume: a page like that for every question your buyers ask, with a source behind every claim, kept current month after month.

5. Be the brand ChatGPT searches for by name.

The biggest change in ChatGPT search in 2026 is where it looks. On August 8, Promptwatch saw the share of ChatGPT searches using a site: operator (a search limited to one domain) jump from 0.37% to 16.8% overnight. Within six days Reddit's share of ChatGPT citations fell 86%, from 3.83% to 0.52%. Search analyst Lily Ray, reviewing several datasets that month, pointed out the word “official” turning up in those searches. ChatGPT now goes to the brand's own site, or a government site, for a fact, where it used to take a forum's word for it.

Mohanadasan's test shows what that is worth. Brands that ChatGPT named in its own search query were cited 68.9% of the time. Pages that were merely retrieved were cited 2.1% of the time. Much of the decision is made before the search runs: if the model already ties your brand to the category, it searches for you by name and then cites you. Two jobs follow from that.

• Be the official source on yourself. Pricing, plans, limits, integrations, refund terms and what you do not do, each stated plainly in HTML on your own domain, on a page whose title says what it is. If ChatGPT searches your domain for a price and finds “contact sales”, the only number left to quote is the one in someone else's review.
• Get named where the model learns your category. A model's sense of who belongs in a category comes from the pages that describe the category: roundups, comparisons, reviews and directories. Ask ChatGPT the category question, open the sources under the answer, and work on earning a real place on those pages. Use one spelling of your brand name everywhere, so every mention counts toward the same entity.

This is where ChatGPT work overlaps with ordinary authority building, and where it parts ways with it. Our piece on how GEO differs from SEO draws that line, and the AI search optimization page covers the same levers across ChatGPT, Perplexity, Gemini and Google's AI features.

6. The how to rank in ChatGPT checklist.

Twelve checks, in the order to do them. Access comes first because nothing else counts until the crawler can read you. This is ChatGPT search optimization reduced to what you can verify. For the classic search basics underneath it, use our SEO checklist.

7. How to measure whether it is working.

You can measure most of this with what you already have. Three signals cover it.

• Referral traffic. ChatGPT's publisher FAQ says it adds utm_source=chatgpt.com to the links it sends. Filter your analytics by that source. In Semrush's clickstream data, referrals from ChatGPT to websites grew 206% between January 2025 and January 2026.
• Server logs. Count requests from OAI-SearchBot, which tell you the index is reading you, and from ChatGPT-User, which tell you a person's prompt led ChatGPT to open your page. ChatGPT-User hits per URL are the closest thing you have to an impression count.
• A fixed prompt set. Write 20 prompts your buyers would really ask. Run them every month with search on, and record two things for each: whether your brand is named in the answer, and whether your page is in the sources. Answers vary from run to run, so follow the trend across months. The AI visibility tools we reviewed automate exactly this, and the ChatGPT SEO tools roundup separates the tools that track ChatGPT citations from the ones that use ChatGPT to do SEO work.

Expect the numbers to move under you. In the Green Flag Digital study, Wikipedia and Reddit went from 18% of the top 500 citations to none in eleven months, while pages from banks and card issuers rose from 2% to 22%. A page that is cited today holds that place only while it stays the best current answer. The wider picture of how fast AI search is moving is in our AI SEO statistics.

A worked example: 30 days from invisible to cited.

Take a fictional company, Fernhill, that sells scheduling software to physiotherapy clinics. The company, its site (on the reserved .example domain) and every number below are invented to show the method. The mechanics match the documentation and studies above.

Fernhill has two problems. Ask ChatGPT for the best scheduling software for a physiotherapy clinic and it names three competitors. Ask what Fernhill costs and it quotes a price from a two-year-old review.

The fixes ran in the same order as the checklist.

1. Day 1, access. Allowed OAI-SearchBot and ChatGPT-User in robots.txt, exempted the published crawler IP addresses from the CDN rule, and confirmed 200 responses in the logs two days later.
2. Day 3, the official facts. Rebuilt the pricing page as plain HTML: three plans, the price per practitioner, what is not included, and an updated date.
3. Days 5 to 25, the pages. Published eight pages, one per buyer question: the category page, four use cases and three comparisons. Each opens with a two-sentence answer.
4. All month, the name. One spelling, Fernhill, everywhere. They also asked the authors of the two roundups ChatGPT cited for the category prompt to review the product. One added it.

Here is the opening of the category page before and after.

What worked is not clever. The access fix made Fernhill eligible. The pricing page made it the official source on itself. The eight pages gave ChatGPT's rewritten queries something to match. The category prompt moved slowest, because being named among the best depends on what the rest of the web says, and only one roundup out of two changed. The pages took most of the month, and Fernhill's list of buyer questions has forty more on it.

Common mistakes when trying to rank in ChatGPT.

1. Blocking the search crawler by accident. People block AI bots to keep their content out of training, and the rule takes OAI-SearchBot with it. Block GPTBot if you want to. Leave OAI-SearchBot and ChatGPT-User alone, and check the CDN as well as robots.txt.
2. Treating a Google ranking as a ChatGPT ranking. It feels safe because the page already wins. But only about 8% of the pages ChatGPT cites are in Google's top 10 for the same query. Check what ChatGPT cites for your prompts before assuming you are covered.
3. Optimizing for the prompt, not the searches. The prompt is long and conversational. The searches ChatGPT runs from it are short and specific. Cover the sub-questions (price, comparison, fit for a use case) with their own pages, and stop tuning one page to one long prompt.
4. Hiding your own facts. Pricing behind “contact sales”, specs in a PDF, plan limits loaded by JavaScript. Each one hands the answer to a third party that may be out of date.
5. Publishing an llms.txt and calling it done. It is quick to ship and easy to tick off, which is why it gets sold as the fix. No answer engine has said it reads the file, as our llms.txt guide lays out. Access and pages are what move citations.
6. Checking once. One good answer on a Tuesday proves little. Answers vary from run to run, and the set of cited pages turns over within months. Track a fixed prompt set over time.