ChatGPT has no ranking you can climb: it runs a few searches, pulls pages, scores passages against the question and writes one answer, so a page is either used as a source or left out. When leadership asks why ChatGPT names a competitor and never you, the answer sits in the pages ChatGPT used as sources.
Being named and being cited are two separate results. In our 2026 AI search visibility study, 39% of the brands AI engines mentioned never had their own website among the top 30 sources those engines cited for their category. Other people's pages put those brands in the answer.
My verdict: your own site is where you become eligible and where ChatGPT checks your proof, and most of what it reads about your category sits on other people's pages. You have two jobs, in this order: make your own key pages specific enough to cite, then get onto the lists, directories and threads ChatGPT reads.
The limit of what follows: OpenAI says ChatGPT "ranks search results using multiple factors" and names none of them (OpenAI Help Center). Everything past step 1 is inference from public tests and from the buyer questions we track every day, most of them about hiring an agency, the category we know best.
Below I show how ChatGPT picks what it cites, then four steps in order, each with a check. In this guide I cover:
How ChatGPT picks its sources
Why ChatGPT names your competitor
Step 1. Open your site to OpenAI's crawlers
Step 2. Put the answer where the crawler reads first
Step 3. Earn citations on the pages ChatGPT trusts
Step 4. Measure the change against a baseline
How ChatGPT picks its sources
A page reaches a ChatGPT answer only if it survives four decisions, and most pages are dropped after ChatGPT has already read them. Every answer with sources goes through the same four:
Reranking drops the most pages, because a page can be read in full and still never be cited.
Fan-out: ChatGPT rewrites the buyer's question into several shorter searches.
Retrieval: Pages come back for each search, from outside search providers and from OpenAI's own index.
Reranking: Each passage is scored against the original question, and most are dropped.
Generation: ChatGPT writes one answer and cites some of the passages that survived.
OpenAI describes the first decision itself: ChatGPT "typically rewrites your query into one or more targeted queries" and sends them to search providers. We track 30 English buyer questions on ChatGPT every day, and this is how one of them was rewritten on 2026-10-05:
Two searches found candidates, and four checked named agencies one by one, each with the words "case study". A page that matched none of those six searches never entered the pool.
AirOps measured the same drop at scale, following 548,534 pages that ChatGPT retrieved for 15,000 prompts:
Oshen Davidson, who led the study, reports that "ChatGPT left 85% of retrieved pages uncited" (AirOps). The same study found that pages ranked first on Google were cited 43.2% of the time, 3.5 times the rate of pages beyond Google's top 20, so ordinary Google rankings for the buyer's searches still raise your odds.
Some of those searches never reach Google at all. Tomek Rudzki of Peec AI found in ChatGPT's own server events that it now answers partly from Labrador, an index OpenAI built itself, and still pulls results from Google and Microsoft (Peec AI). Does ChatGPT use Bing or Google in 2026 covers what OpenAI documents and what outside tests show.
Why ChatGPT names your competitor
ChatGPT names a competitor when the pages it read for that question name the competitor and leave you out. Our study of 2,283,923 AI citations measured how small a part of those pages a business's own website is:
The typical business's own site took 2.0% of the citations about its category. It was cited somewhere in 90.6% of the 138 projects, usually as about the 10th most cited website on its own category's list.
The answer from the previous section shows the second route. ChatGPT read 68 pages for that question and cited 6:
All six cited pages sat on the agencies' own websites: service pages, a news post about one agency's own results and a blog post. The directories and the best-of list it read went uncited in this answer. My reading is that the lists supplied the names and the agencies' own pages supplied the proof.
So a competitor gets named through either route, or both: other pages mention it, and its own page gives ChatGPT something specific to cite. That is why the two jobs from the verdict come in that order, with step 2 for your own pages and step 3 for everyone else's. Why AI reads your site and recommends a competitor walks through the most common failure, a site that AI crawlers visit every day and never cite.
Reddit carries more weight on ChatGPT than on the other engines:
In our source study ChatGPT gives reddit.com 3.0% of its citations, five times Perplexity's share and more than it gives any other single website. Ahrefs' September 2026 count of the domains ChatGPT cites also puts Reddit first, at 16.8%, with Wikipedia second at 7.0%. Ahrefs counts US questions on every topic and we count buyer questions about specific categories, mostly in Lithuania, so the shares differ by design and both put Reddit on top.
Further reading: how ChatGPT decides which products to recommend, for shopping questions.
Step 1. Open your site to OpenAI's crawlers
Every later step is wasted on a site OpenAI's crawler cannot reach, which is why this one comes first. The work runs in four steps, each with a check:
Record your starting numbers before you change anything, as step 4 describes, so the fixes below have something to be measured against. If you are not sure buyers in your category ask AI assistants at all, where to start with GEO settles that first.
OpenAI's crawler documentation gives one instruction for search:
"Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers", and OpenAI recommends "allowing OAI-SearchBot in your site's robots.txt file and allowing requests from our published IP ranges" (OpenAI). Its help page adds the host side: confirm "that the website host or content delivery network allows traffic from OpenAI's published searchbot IP addresses." After a robots.txt change, OpenAI says its systems can take about 24 hours to adjust.
Sidenote: GPTBot is OpenAI's training crawler, and OpenAI says each setting is independent. Blocking GPTBot keeps your pages out of model training and leaves search alone, and ChatGPT-User, which fetches a page when a user's request needs it, "is not used to determine whether content may appear in Search."
Hand your developer these checks, in this order:
Read the file with
curl -s https://yourdomain.com/robots.txtand look forDisallow: /underUser-agent: *or under any OpenAI agent.Request the homepage as OpenAI's crawler with
curl -A "OAI-SearchBot" -o /dev/null -s -w "%{http_code}\n" https://yourdomain.com/and expect a 200.Open your CDN's security log, filtered to verified bots, and look for blocked or challenged requests from OpenAI.
Allow the other engines' crawlers in the same file, PerplexityBot and Googlebot included, as Perplexity's crawler documentation and our list of 51 AI crawlers describe.
What good looks like in robots.txt:
GPTBot is in the example because most businesses we work with allow training. Leave that block out if you do not want your pages used to train OpenAI's models. A plain User-agent: * followed by Allow: / also passes, and the explicit lines are there so a later plugin cannot quietly narrow it.
Sidenote: a spoofed user agent tests robots.txt and simple user-agent rules only. Rules that fire on IP ranges or on a CDN's verified-bot category show up only in the site's own security log.
We have seen that gap on a real site. On one client site we track, robots.txt carried our Allow lines and the curl test passed, while Cloudflare still blocked the crawler, and the site's security log showed the block. In Cloudflare, AI Crawl Control lists which AI services reach the site, and Bot Fight Mode is worth checking first, since it challenges visitors it suspects are automated.
Step 2. Put the answer where the crawler reads first
Write each key page so its first lines answer a buyer's question, because the reranker scores passages against that question. The H1 is the question people type, the first sentence answers it, and every table is restated in a sentence, since text extractors lose answers that live only in cells. Our guide to AEO covers this answer shape in more depth.
Here is the difference on the first line of a software company's page:
Before, a typical product page: "Smart accounting software for ambitious businesses."
After, a page written for one buyer question: "[Your product] against Xero and QuickBooks for a UK company of 10 to 50 staff: payroll, VAT filing and price per user, compared in one table."
Peec AI tested why the second kind of line wins. Metehan Yesilyurt scored two passages for the question "what are the best aeo tools" with a set of open reranker models:
On BGE-reranker-large, the product description scored 0.20% and the shortlist 99.8% (Peec AI). Three of the four models in the chart agreed, and the fourth rated both passages above 99%. These are open models, and ChatGPT's own reranker was not tested.
Your own pages earn citations when they carry something specific. In the answer from the first section, every cited page was a service or results page on the agency's own site. A page that states a result, a price or a named method gives ChatGPT a passage it can quote, and our content service writes these pages for the buyer questions a client tracks.
Schema comes after the copy. Add Organization markup on the homepage and BlogPosting and FAQPage markup on articles, and expect little from it in ChatGPT.
Sidenote: Ahrefs tracked 1,885 pages that added JSON-LD against 4,000 matched controls and measured ChatGPT citations up 2.2%, small enough to be random noise (Ahrefs). Google says "Structured data isn't required for generative AI search" and still recommends it for rich results (Google Search Central).
Step 3. Earn citations on the pages ChatGPT trusts
The pages ChatGPT trusts for your category are the ones that keep coming back in its answers, and a citation audit finds them:
Write 10 to 15 questions a buyer asks about your category, in the buyer's own words.
Ask each one on ChatGPT, Gemini and Perplexity in a logged-out window.
Record every cited domain, question by question.
Pick the three to five domains that repeat, and work on those first.
For English questions about hiring an agency, our own tracking shows which kinds of pages carry ChatGPT's citations:
Best-of lists took 24.7% of the citations across the 1,000 most cited pages, and the agencies' own service pages took 39.6%. In this category the lists are the third-party pages to work on first, and your service page is the one step 2 has already fixed. Your audit will show the mix for your category.
Then earn a place on the pages that repeat:
Best-of lists: Find the author's name and email on the page, then send the facts that would make your entry accurate, prices included, with a named contact.
Directories and review sites: Complete every field of the profile, since these pages are cited for being structured and complete.
Reddit: Answer the threads ChatGPT already cites as a person with a name, the limits of your product included, because a sales pitch gets removed and a useful answer stays up.
Placements take time, because other people publish on their own schedule. The free route is the four audit steps above in a spreadsheet. If you would rather have it run across weeks of tracked answers, we run a free citation audit, and if you hand this work to an outside firm, our list of GEO agencies in Europe shows what to check before signing.
Step 4. Measure the change against a baseline
A screenshot of ChatGPT naming you proves one answer on one day. Visibility has moved when the share of tracked buyer questions that name you rises against a baseline measured before the work started, read weekly and per engine.
The setup we run for clients through our GEO service:
Prompts: 50 buyer questions, written the way buyers type them and never naming the brand.
Engines: ChatGPT, Gemini and Google AI Overviews, each read on its own.
Baseline: Measured before the first change ships, so the fixes from step 1 read against it.
Free instruments: A GA4 channel that groups referrals from chatgpt.com, perplexity.ai and gemini.google.com, read next to organic search, plus Microsoft Clarity for what those visitors do.
Expect a low starting point. The median European brand in our study appears in 5% of AI answers about its own category, so a brand that reaches 10 to 15% has doubled or tripled and still misses most answers. The engines also disagree with each other:
A brand's best engine shows it a median 5.1x more often than its worst, and almost a third of the brands we could see were missing from at least one engine. A ChatGPT-only check can mislead you. Our guide to prompt tracking explains how to write the question set, and best AI visibility tools in 2026 compares the tools that run it.
This is what a weekly read from a measured baseline looks like:
When Outcraft, our B2B SaaS case study, came to us in May 2026, its robots.txt blocked GPTBot, ClaudeBot and Google-Extended, and it was named in 0.25% of the answers we tracked in its first week. Opening the crawlers in month one made the site eligible, the content that followed moved the number, and 14 weeks later 11.9% of the week's answers named it, roughly one in eight.
Sidenote: llms.txt is the last file to ship and the first to skip. Ahrefs found that 97% of the llms.txt files it studied received zero requests in May 2026 (Ahrefs), and Google's guide says you don't need "AI text files" to appear in Google Search.
Further reading: our llms.txt guide.
FAQ
Can you rank on ChatGPT?
No. ChatGPT writes one answer and cites a handful of pages, so there is no list with a position to hold. The nearest equivalent is the share of tracked buyer questions whose answer names your brand, read weekly on each engine. Crawler access, a passage that answers the question and a presence on the pages ChatGPT trusts are what move it.
Does the same work apply to Gemini and Perplexity?
Mostly, with different crawlers and different sources. Perplexity crawls with PerplexityBot and Google's AI features draw on Google's search index, so the robots.txt and firewall checks in step 1 cover them in the same pass. The cited sources differ: in our source study, ChatGPT, Google AI Overviews and Perplexity shared a median of 2 of their top 10 cited domains for the same questions, so run the citation audit on each engine.
How long does it take to show up in ChatGPT?
The crawler fix is fastest: OpenAI says its systems take about 24 hours to adjust after a robots.txt change. Content and third-party citations take weeks to months, because other people publish on their own schedule. In our B2B SaaS case, visibility went from 0.25% to 11.9% of tracked answers in 14 weeks, so read your own numbers weekly against the baseline.
Can I pay to appear in ChatGPT?
Yes, as a labelled ad. OpenAI began testing ads in the United States in February 2026 and said "Ads do not influence the answers ChatGPT gives you" (TechCrunch), and ads reached 31 European countries on 2026-08-24, as ChatGPT ads in Lithuania explains. In March 2026 OpenAI also let merchants use their own checkout and share product feeds with ChatGPT (CNBC). Which brands ChatGPT recommends in the answer itself stays unpaid, and our ChatGPT ads service covers the paid side.
Final thoughts
You can't rank on ChatGPT, because there is no list. You can make your site eligible, put the answer where the crawler reads first and get named on the pages ChatGPT already trusts, in that order, and measure each step against a baseline.
This week, record a baseline for ten buyer questions, then hand your developer the checks in step 1, the CDN's security log included. If any of them shows a block, nothing else on this page can work until it is fixed.
If you run the checks and want a second pair of eyes on the result, send it to me on LinkedIn.
