How to rank in ChatGPT: what OpenAI documents, what nobody outside OpenAI knows, and what to do

There is no ranking in ChatGPT, there are two ways into an answer

When this article first went up, it promised patterns for "ranking" in ChatGPT and leaned on claims I cannot source. This version keeps the promise and drops the folklore: what OpenAI documents, what Microsoft has said about Bing, and what researchers have measured. Where nobody outside OpenAI knows, I say so.

Start with the word itself. In Google there is a list of results and a position on it. In ChatGPT there is no list. There is an answer, written fresh each time, and your business is either in it or not. It can get into that answer by two quite different routes.

The first route is memory. A language model is trained on a very large amount of text, and some of what it read stays with it. Ask a question with web search switched off, and the answer comes from that memory alone, with no links and no live sources.

The second route is search. ChatGPT can go to the web before it answers, fetch pages, and write the answer from what it found, with links to the sources. That route has a documented gate, which the rest of this article spends most of its time on, because it is the one you can actually do something about this month.

The diagram runs from the question at the top to the cited answer at the bottom. The first decision is the model's, not yours: whether to search at all. The second decision is yours: whether OpenAI's search crawler may read your site. The takeaway is that the memory route has no gate you can open this month, and the search route has exactly one.

This article is the ChatGPT specific one in a small family. How to write a passage that any answer engine can lift is in winning visibility in the age of AEO and GEO. How ChatGPT and Google's Gemini differ, side by side, is in ChatGPT vs Gemini. I link to both rather than repeat them.

Worked example: one question, both routes

Corrib Roastery is an invented coffee roaster in Galway. It sells beans online at corrib.example, mostly as a subscription at €9.50 a bag, and it is run by Aoife. Nobody real, every number made up.

A customer tells Aoife she found a rival roaster by asking ChatGPT "what is a good coffee subscription in Ireland for a cafetière". Aoife tries it herself, twice.

Step one: she asks with web search switched off. The answer names two large international brands and gives general advice about grind size. No links. That is the memory route, and Corrib, a roaster with 1,800 subscribers, is not famous enough to be in anyone's memory.

Step two: she asks again with search on. This time the answer cites five sources: two independent coffee review sites, a newspaper's gift guide, a roaster in Cork and a roaster in Limerick. Still no Corrib. That is the search route, and on this route she has something to check.

The memory route: what training can and cannot do for you

On the memory route, the honest answer is that there is very little a business can do quickly, and a great deal of advice pretends otherwise.

OpenAI documents one control. Its crawler GPTBot gathers content used, in OpenAI's words, "to make our generative AI foundation models more useful and safe", and "disallowing GPTBot indicates a site's content should not be used in training generative AI foundation models".¹ So you can say no to future training. You cannot say "please remember me", and nobody outside OpenAI can see which pages went into a given model, how much weight any one of them carried, or when the next model will be trained.

The old version of this article said that Reddit, Quora, Stack Overflow and Medium "are training data", and that posting answers there with links back to your site gives "semantic reinforcement". I have cut that. I could not find a primary source that states which forums are in which model, and the one study below that looked at where ChatGPT's search answers point found almost no social platforms among the sources at all.² Posting a link to your shop under every relevant question on a forum also has an older name, and it is not reinforcement.

What does put a business in a model's memory is the slow thing: being written about by other people, in many places, over years. That is reputation, not optimisation.

Worked example: Aoife's training decision

Aoife has to decide whether GPTBot may read her site. She writes down what each choice costs.

Allow it: her brewing guides may be used to train future models. She gets nothing measurable in return this year, and nothing in OpenAI's documentation promises she will.

Block it: her guides are not used for training. On OpenAI's own description, this is a separate switch from search, so it does not remove her from search answers.¹

She allows it, because her guides are marketing and she would rather a model learnt that Corrib exists. A firm with paid content might reasonably decide the other way. Either way it is a decision, written down with a date.

The search route has one documented gate

Here OpenAI is clear. Its documentation describes OAI-SearchBot as the crawler "used to surface websites in search results in ChatGPT's search features". Sites that block it "will not be shown in ChatGPT search answers", though they "can still appear as navigational links". OpenAI recommends allowing it, and says a change to robots.txt can take about 24 hours to take effect.¹

There is a third agent, ChatGPT-User, which fetches a page when a person in ChatGPT asks for it. OpenAI says it "is not used for crawling the web in an automatic fashion", that robots.txt rules "may not apply" because a person started the request, and that it "is not used to determine whether content may appear in Search".¹ So you will see it in your logs, and it tells you nothing about your search visibility.

That is the whole published rule for the search route: let the search crawler in. Everything beyond it, how ChatGPT chooses which of the allowed pages to read and which to cite, is not documented. The robots.txt patterns for allowing search and refusing training are written out line by line in the AEO and GEO article, so I will not repeat them here.

Worked example: Aoife checks the gate

Step one: Aoife opens corrib.example/robots.txt. It was generated by a security plugin three years ago and contains a block that refuses every user agent whose name its author considered "AI", OAI-SearchBot among them.

Step two: she reads OpenAI's rule against it. Blocked from the search crawler means not shown in ChatGPT search answers.¹ So the five sources in her test could never have included her, however good her pages are.

Step three: she replaces the block with two explicit lines, OAI-SearchBot allowed and GPTBot allowed, as decided in the previous section.

Step four: two days later, her developer searches the server access log for requests whose user agent contains OAI-SearchBot. There are 37 of them, spread across the shop's guides and product pages. The gate is open, and she has evidence rather than hope.

Where the search results come from: the Bing question

The old article said, flatly, "ChatGPT uses Bing for live lookups", and built its first tactic on that. Here is what can actually be sourced, in date order.

In May 2023, Microsoft announced that "Bing is coming to ChatGPT as the default search experience", with answers grounded in web data and carrying citations.³ That was true and it was Microsoft's own statement.

Today, OpenAI's crawler documentation does not mention Bing or any other search provider.¹ Its developer guide to the web search tool in its own API does not name a provider either, and mentions an offline, cache only mode, which suggests OpenAI keeps fetched content of its own.⁴ And OpenAI runs its own search crawler, which is the gate in the previous section.

So the honest position for 2026 is this. ChatGPT search was built on Bing in 2023. OpenAI now fetches pages itself. Whether, and how much, it still draws on Bing's index is not stated in anything I could find from OpenAI or Microsoft. My sister article on AEO and GEO says the chat engines lean on Bing's index, and on today's evidence I would soften that to "did, and may still in part".

Is Bing then worth any time? Yes, for a reason that does not depend on ChatGPT. Bing runs its own AI answers in Microsoft Copilot and in Bing's summaries, and since February 2026 Bing Webmaster Tools has a report, in public preview, that counts how often your pages are cited as a source in those answers and shows the queries behind the citations.⁵ That is the only first party citation report I know of from any answer engine. Bing's own advice in the same post is to keep content fresh, notify it of changes through IndexNow, and keep headings, tables and sources clear.⁵ IndexNow has its own article.

To check whether Bing knows your pages, the useful number is a plain rate:

$$\text{coverage} = \frac{\text{pages indexed}}{\text{pages submitted}}$$

Here $\text{pages indexed}$ is the number of URLs from your sitemap that Bing Webmaster Tools reports as indexed, and $\text{pages submitted}$ is the number of URLs in the sitemap you gave it. It is events over opportunities, like any conversion rate (Foundations: ratios and rates).

Worked example: how much of the shop Bing knows

Step one: Aoife verifies corrib.example in Bing Webmaster Tools, which takes ten minutes, and submits the sitemap. It lists 140 URLs.

Step two: a week later the tool reports 62 of them indexed. Coverage is $62 / 140 = 0.443$, or 44.3%.

Step three: her developer looks at what is missing. Every coffee is listed three times, once per grind, under URLs that differ only by a ?grind= parameter, and none of them names a canonical version. The search engine sees 36 near identical pages and indexes a few at random.

Step four: he points each grind variant at one canonical product page and takes the variants out of the sitemap, which now lists 128 URLs. Four weeks later 121 are indexed. Coverage is $121 / 128 = 0.945$, or 94.5%.

Step five: Aoife opens the AI Performance report. It shows her first 11 citations in Copilot answers, nine of them on her guide to brewing with a cafetière. That is Bing's answer engine, not ChatGPT's, and she knows it. It is still the first hard evidence that a page of hers is being lifted into a generated answer.

What the research actually measured

Most of what is written about "ranking in ChatGPT" quotes studies second hand. Here are four, with what each one measured, because the detail changes what you can take from them.

The paper that gave this field its name is GEO by Pranjal Aggarwal and colleagues, presented at KDD in 2024.⁶ Its engine was a research setup: the top five Google results for each query, handed to gpt3.5-turbo to write the answer. The authors also repeated the tests on Perplexity, where the best edits lifted visibility by up to 37%.⁶ They did not test ChatGPT. What they found, adding statistics, quotations and cited sources helps and keyword stuffing hurts, is summarised with the figures in the AEO and GEO article. It is a good guide to writing. It is not a measurement of ChatGPT.

The study closest to ChatGPT is by Mahe Chen and colleagues at the University of Toronto, published as a preprint in September 2025.² They sent 1,000 ranking questions such as "rank the best smartphones" through OpenAI's search model gpt-4o-search-preview via the API, and through Google, in August 2025, and sorted every cited domain into three kinds: brand (the maker's own site), social (forums and social networks) and earned (independent publishers and reviewers). They ran the comparison in two countries, the United States and Canada, and report no figures for Ireland or the UK, so the chart below shows the American ones. The Canadian figures point the same way: for automotive questions, 69.1% of the OpenAI model's sources were independent publishers, against 40.6% of Google's.²

Along the bottom are two product categories, each with Google's sources and the OpenAI search model's sources. Up the side is the share of cited sources that were independent publishers. The takeaway: for the same questions, the OpenAI model leaned far more heavily on third party publishers than Google did, and in the same study it cited social platforms almost never, close to zero in every category and country it reported.² The authors stress that the data is from one month, that their three way classification is their own and somewhat subjective, and that the API model is not the consumer ChatGPT app.² I take the direction from it, not the decimals.

The third study is older and sobering. Nelson Liu, Tianyi Zhang and Percy Liang audited four generative search engines, including Bing Chat, in 2023. On average only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to.⁷ Being cited is not the same as being represented correctly. Read what the answer says about you, not just whether your link is there.

The fourth kind of study is a warning. Aounon Kumar and Himabindu Lakkaraju showed that a crafted string of text on a product page could push that product to the top of a language model's recommendations,⁸ and Samuel Pfrommer and colleagues showed that hidden instructions on a page could promote low ranked products in conversational search engines, with attacks that carried over to Perplexity.⁹ Both are adversarial research, the equivalent of hidden text in 2005. I mention them so you recognise the pitch when someone sells it. I would not put it on a client's site.

Worked example: where Aoife's answers come from

Step one: Aoife writes 25 questions her customers actually ask, from her own enquiry inbox, such as "best coffee for a cafetière", "is a coffee subscription worth it" and "how long do roasted beans stay fresh".

Step two: she asks each in ChatGPT with search on, in a fresh chat, and copies every cited domain into a spreadsheet. Across the 25 answers there are 58 citations.

Step three: she counts by domain. Three independent coffee review sites account for 31 of the 58. The share is

$$\frac{31}{58} \times 100 = 53.4\%$$

which uses the plain percentage, a part over the whole, times 100, with $31$ the citations of the three review sites and $58$ all citations (Foundations: percentages).

Step four: she draws the conclusion the Toronto study points to. More than half of the answers in her niche lean on three independent publishers. None of them has ever reviewed Corrib. Being reviewed by one of them, honestly, by sending beans and letting them say what they think, is probably worth more to her ChatGPT visibility than anything she can change on her own site. That is ordinary press work, and it was before anyone said GEO.

What nobody outside OpenAI knows

This is the section most articles on the topic leave out, so here it is plainly. As of November 2026, none of the following is documented by OpenAI, and anyone who states them as fact is guessing.

• Which search providers ChatGPT search uses today, and in what mix with OpenAI's own crawl.
• How ChatGPT decides whether to search for a given question, and how many searches it runs.
• How it chooses which of the fetched pages to read, and which to cite.
• Whether structured data such as schema.org markup is read or weighted at all.
• Whether a person's past chats or saved memories change which sources they are shown.
• How much the answer from memory and the answer from search are blended in one reply.

There is no ChatGPT equivalent of Google Search Console, and no report of how often your site is cited. The old version of this article ended on "there will never be a dashboard", and that is the one line from it I still agree with, for ChatGPT at least. Bing, as above, now has one for its own answers.⁵

What follows from not knowing is method. You cannot reverse engineer a system that is not published, but you can measure what comes out of it, consistently, and change one thing at a time.

Worked example: the same question five times

Step one: Aoife asks "what is a good coffee subscription in Ireland for a cafetière" five times, each in a fresh chat with search on, on the same afternoon.

Step two: she lists the cited domains. Across the five answers there are 11 different domains. Two appear in all five. Four appear only once.

Step three: the lesson. One test tells her almost nothing, because the sources move between identical runs. The Toronto study saw sources move when questions were reworded, and far more when they were asked in another language.² A measurement needs many questions, asked the same way, on a schedule.

Measuring your place in the answers

Measurement here is plain counting, and the counting has to be honest about small numbers. The method in detail, with 30 questions across two tools, is in the ChatGPT vs Gemini article. Here is the short version for ChatGPT alone.

Keep a fixed list of questions. Ask each one in ChatGPT with search on, in a fresh chat, once a month. Count the answers that cite your site. The number you track is

$$r = \frac{c}{q}$$

where $r$ is the citation rate for the month, $c$ is the number of questions whose answer cited your site at least once, and $q$ is the number of questions asked, which never changes (Foundations: ratios and rates).

Because $q$ is small, the rate wobbles even when nothing has changed. The standard error tells you how much:

$$SE = \sqrt{\frac{r(1 - r)}{q}}$$

Here $SE$ is the standard error of the rate, $r$ the rate you measured and $q$ the number of questions. A month on month change smaller than about two standard errors is the kind chance produces all the time (Foundations: sampling).

Two more numbers belong next to it. First, whether the answer is right about you, because citations are often attached to sentences they do not support.⁷ Second, the visits. In the analytics I look after, visits from ChatGPT arrive tagged with utm_source=chatgpt.com, which is how I count them. Visits only show the people who clicked, never the answers where you were named and nobody did.

Worked example: Aoife's six months

Aoife's list has the 25 questions from earlier. Her robots.txt fix went live at the end of February, the Bing work in March, and in April a review site in her niche published a review of her subscription.

Along the bottom are the months, February to July. Up the side is how many of the 25 answers cited Corrib. The takeaway: nothing before the gate opened, a small step after it, and a steady climb after the review appeared. Here is whether the climb is real.

Step one: rates. March is $3 / 25 = 0.12$, July is $9 / 25 = 0.36$.

Step two: the standard error of each month. March: $\sqrt{0.12 \times 0.88 / 25} = \sqrt{0.004224} = 0.065$. July: $\sqrt{0.36 \times 0.64 / 25} = \sqrt{0.009216} = 0.096$.

Step three: the standard error of the difference combines both: $\sqrt{0.004224 + 0.009216} = \sqrt{0.01344} = 0.116$. The rise is $0.36 - 0.12 = 0.24$, which is $0.24 / 0.116 = 2.07$ standard errors.

Step four: read it. Just over two standard errors, five months in a row moving the same way, and a timeline that matches the work. That is grand as a reason to keep going, and not proof of which change did it. February's zero, on the other hand, needs no statistics: blocked is blocked.

Step five: the visits. Sessions tagged utm_source=chatgpt.com went from four in March to 23 in July. Small, and real, and seven of the July sessions became subscriptions at €9.50 a bag.

The old advice, checked against the sources

The old article listed five tactics. Here is each one, kept, corrected or cut.

"Optimise for Bing, not just Google." Partly kept. The claim that ChatGPT uses Bing was Microsoft's own statement in 2023³ and is not confirmed by anything current from OpenAI. Being indexed in Bing is still low cost insurance, and it feeds Copilot, which now reports citations.⁵ I also cut the line about my own clients' pages appearing "in GPT suggestions within weeks" after a Bing restructure, because I cannot show you the before and after, and you should not take it on trust.

"Target long tail, high intent questions." Kept, reframed. Writing one page that answers one real question plainly is good writing, and it is what the GEO research rewarded.⁶ How to do it is in the AEO and GEO article.

"Use FAQ schema, it trains language models." Cut. OpenAI does not document any use of structured data in ChatGPT search, and on Google, FAQ rich results were shown only for "well known, authoritative government and health websites" from August 2023.¹⁰ Since 7 May 2026 they do not show at all.¹¹ Accurate markup that describes the page and the business is still worth having, for the reasons in the sister article. It is not a lever on ChatGPT that anyone has shown.

"Build thematic content clusters." Kept as site structure, because linked guides help readers. Cut as a ChatGPT claim, because I found no source showing ChatGPT rewards it.

"Get mentioned in Reddit, Quora and Stack Overflow, they are training data." Cut, and replaced with its honest cousin. The study closest to ChatGPT found its search model cited independent publishers heavily and social platforms almost never.² Be worth reviewing, and get reviewed. Do not seed links.

Worked example: Aoife's monthly hour count

Aoife has four hours a month for this, and she spends them like this.

Step one, 15 minutes: check the robots.txt and the access log for OAI-SearchBot, so the gate never quietly closes again after a plugin update.

Step two, 15 minutes: check Bing coverage and the AI Performance report.

Step three, two hours: improve one guide that answers one of her 25 questions, with the question as the heading and the answer in the first sentence.

Step four, 30 minutes: write to one independent reviewer in her niche, offering beans and no conditions.

Step five, one hour: ask the 25 questions, fill in the spreadsheet, and work out $r$ and its standard error.

Four hours, no tool subscription, and every step either opens a documented gate or measures the result.

What I would do on Monday

If I sat down with your site tomorrow, the order would be short.

1. Read the robots.txt. Allow OAI-SearchBot unless you have a reason not to. Decide on GPTBot on purpose and write the decision down.
2. Verify the site in Bing Webmaster Tools, submit the sitemap, and get coverage above 90% by fixing duplicates rather than by submitting more.
3. Find out which independent publishers ChatGPT cites in your niche, and earn a place with one of them.
4. Write 25 questions your customers ask, and count citations monthly. Treat a single month's move under about two standard errors as noise.
5. Read what the answers say about you, not only whether you are linked.

ChatGPT is a black box with one documented gate and a measurable output. That is enough to work with.

Footnotes

1. OpenAI, Overview of OpenAI crawlers (GPTBot, OAI-SearchBot, ChatGPT-User), OpenAI developer documentation, accessed 7 November 2026, https://developers.openai.com/api/docs/bots

2. Mahe Chen, Xiaoxuan Wang, Kaiwen Chen and Nick Koudas, Generative Engine Optimization: How to Dominate AI Search, University of Toronto, arXiv preprint 2509.08919 (not peer reviewed), 10 September 2025, https://arxiv.org/abs/2509.08919

3. Frank X. Shaw, Microsoft Build brings AI tools to the forefront for developers, Microsoft, 23 May 2023, https://blogs.microsoft.com/blog/2023/05/23/microsoft-build-brings-ai-tools-to-the-forefront-for-developers/

4. OpenAI, Web search (tools guide for the Responses API), OpenAI developer documentation, accessed 7 November 2026, https://developers.openai.com/api/docs/guides/tools-web-search

5. Krishna Madhavan, Meenaz Merchant, Fabrice Canel and Saral Nigam, Introducing AI Performance in Bing Webmaster Tools Public Preview, Bing Webmaster Blog, Microsoft, 10 February 2026, https://blogs.bing.com/webmaster/February-2026/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview

6. Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, GEO: Generative Engine Optimization, Proceedings of KDD 2024, arXiv 2311.09735, 2024, https://arxiv.org/abs/2311.09735

7. Nelson F. Liu, Tianyi Zhang and Percy Liang, Evaluating Verifiability in Generative Search Engines, Findings of EMNLP 2023, arXiv 2304.09848, https://arxiv.org/abs/2304.09848

8. Aounon Kumar and Himabindu Lakkaraju, Manipulating Large Language Models to Increase Product Visibility, arXiv preprint 2404.07981, 2024, https://arxiv.org/abs/2404.07981

9. Samuel Pfrommer, Yatong Bai, Tanmay Gautam and Somayeh Sojoudi, Ranking Manipulation for Conversational Search Engines, EMNLP 2024, arXiv 2406.03589, https://arxiv.org/abs/2406.03589

10. Google Search Central, Changes to HowTo and FAQ rich results, 8 August 2023, https://developers.google.com/search/blog/2023/08/howto-faq-changes

11. Google Search Central, Latest Google Search documentation updates, entries of 8 May 2026 (Deprecating the FAQ rich result feature) and 15 June 2026 (Removing documentation for the FAQ rich result feature), https://developers.google.com/search/updates#faq-deprecation