How ChatGPT Chooses Which Sources and Brands to Cite (2026)

ChatGPT picks sources in two very different ways. When it answers from memory, it draws on its training data and cites nothing. When it searches the web, it rewrites the question into several search queries, pulls results from search partners, reads a shortlist of pages and cites the ones that best support each part of its answer.

For brands, the second path is the one you can influence. A page has to be crawlable by OpenAI’s search bot, show up in the search results ChatGPT pulls, and match the sub-questions ChatGPT asks behind the scenes. Brands then get named when they appear often and consistently across the pages ChatGPT reads: “best X” lists, review sites, Reddit threads, Wikipedia, the press and their own product pages.

OpenAI says plainly that placement in ChatGPT search is not guaranteed. Nobody can promise a mention. What you can do is raise the odds, and the 2026 data below shows where those odds come from.

• The short answer: what decides a ChatGPT citation
• Training data vs live search: two ways ChatGPT knows your brand
• How ChatGPT search finds and cites pages, step by step
• How citations and brand links appear in answers
• OpenAI’s crawlers and robots.txt
• Which sites ChatGPT cites most in 2026
• Listicles, reviews and “best X” articles
• Freshness: does ChatGPT prefer new content?
• Which brand signals correlate with ChatGPT mentions
• What a brand can do: a 9-step plan
• Limits: what no one can guarantee
• FAQs
• Get your brand into the sources ChatGPT reads

The short answer: what decides a ChatGPT citation

Here is the whole process in one table. Each row is covered in detail below.

Training data vs live search: two ways ChatGPT knows your brand

ChatGPT can talk about your brand without searching at all. In that case it relies on what the model learned during training. There is no citation, no link and no way to see which pages shaped the answer.

OpenAI separates the two paths in its crawler setup. According to OpenAI’s crawler documentation, GPTBot collects content that may be used to train its foundation models. OAI-SearchBot is used to show websites in ChatGPT’s search features. The two settings are independent.

Live search is where citations come from. OpenAI’s help page on ChatGPT search says ChatGPT may search automatically when a question would benefit from current information, and users can also start a search manually.

Most marketing questions (“best CRM for startups”, “alternatives to X”) are the kind that trigger a search. That is why the rest of this guide focuses on the search path.

How ChatGPT search finds and cites pages, step by step

1. ChatGPT rewrites the prompt into search queries

OpenAI explains that ChatGPT search “sometimes partners with other search providers” and typically rewrites the user’s question into one or more targeted queries. After it reviews the first results, it may send more specific follow-up queries. It can also use approximate location and, if memory is on, saved memories when rewriting the query (OpenAI).

The scale of this is large. In an AirOps study published in March 2026 (15,000 prompts, 548,534 retrieved pages), ChatGPT generated two or more fan-out queries on 89.6% of searches. 95% of those fan-out queries had zero monthly search volume in traditional keyword tools.

For comparison prompts, the fan-out often splits the question apart. AirOps gives the example of “HubSpot vs Salesforce” turning into separate searches about pricing, features and reviews.

2. Search partners return candidate pages

OpenAI does not publish a full list of its search providers. Its help page links to the privacy statements of Microsoft and Shopify as providers that may process queries. In practice, this is why many SEOs check that a site is indexed well in Bing, not only Google.

Google rankings still matter, even if Google is not named as a provider. In the AirOps data, pages ranking #1 in Google were cited by ChatGPT 43.2% of the time. That was 3.5 times the citation rate of pages ranking outside Google’s top 20. This is a correlation: pages that rank well on Google usually rank well elsewhere too.

3. ChatGPT screens results before it reads them

An Ahrefs study of 1.4 million prompts (published April 2026) describes a gatekeeping layer. Each search result arrives with a title, URL, sometimes a snippet and an ID. ChatGPT uses that data to decide which pages to open.

Ahrefs found that cited pages had titles that were more semantically similar to the prompt and to the fan-out queries than non-cited pages. Search results with natural-language URL slugs had an 89.78% citation rate, compared with 81.11% for those without.

4. ChatGPT cites only part of what it retrieves

Both large studies agree that retrieval is not the same as citation. They disagree on the exact share, because they count retrieved pages differently.

Ahrefs also found that 88% of cited URLs came through ChatGPT’s general “search” channel. Reddit content pulled in through a separate channel made up 67.8% of the URLs that were retrieved but not cited. In other words, ChatGPT reads a lot of Reddit for context but cites it much less often than it reads it.

What gets a retrieved page cited? AirOps saw a citation rate of 20.1% for pages whose title overlapped with the query by 50% or more, against 9.3% for pages with less than 10% overlap. Easier-to-read pages were also more common among cited sources.

5. The same prompt can give different sources

ChatGPT’s sources shift between modes and over time. In a Semrush analysis with Kevin Indig (100 prompts across 20 buyer journeys, reported by Search Engine Land), only 25.6% of cited domains overlapped between minimal and high reasoning for the same prompts. High reasoning ran 1,130 web searches across the test set, against 245 for minimal reasoning.

The mix of sources changed too. With high reasoning, Reddit’s citation share fell from 15% to 7%, and official documentation and support pages rose from 12.4% to 17.5%. A brand that shows up in quick answers can disappear in deeper research answers, and the reverse.

How citations and brand links appear in answers

According to OpenAI, responses that use web search “may include citations”. Users can click a citation to open the source, hover over it on desktop to preview it, or open Sources to see cited sources and other relevant links (OpenAI). OpenAI also warns that citations can be incomplete, outdated or wrong.

The May 2026 branded link change

There is now a difference between a citation (a link to the page that supports a claim) and a brand link (a link on the brand name itself). Profound reports that on May 7, 2026 ChatGPT began turning brand names in answers into links to the brand’s homepage far more often (Profound).

• The share of ChatGPT responses containing a URL went from about 4.5% to 20–24% on the same day.
• Homepages went from about 4% to 24% of OpenAI referrals across the sites Profound monitors.
• B2B software and SaaS brands saw the largest gain, with daily OpenAI referrals up more than 200% on their earlier baseline. E-commerce stayed roughly flat.

OpenAI did not announce this change, so treat these as one vendor’s measurements. The practical point holds anyway: being named in an answer can now send traffic, not just awareness. Your homepage has to work for visitors who arrive knowing only what ChatGPT said about you.

OpenAI’s crawlers and robots.txt

This is the one part of the process you fully control. OpenAI lists four user agents in its crawler documentation.

Key facts from OpenAI’s documentation:

• Sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links.
• You can allow OAI-SearchBot and block GPTBot. That keeps you eligible for search answers while signalling you do not want your content used for training.
• Changes to robots.txt can take about 24 hours to take effect for search.
• OpenAI also asks site owners to make sure their host or CDN allows traffic from its published OAI-SearchBot IP addresses. A firewall that blocks the bot cancels out a correct robots.txt.

Many sites have never made this decision. Our study of AI bots in SaaS robots.txt files found that 73% of top SaaS sites name no AI bot at all. You can check your own file with the free robots.txt AI bot checker on our tools page.

What about llms.txt? OpenAI’s crawler documentation does not mention it. Adoption is growing anyway: 74% of 47 leading B2B SaaS sites serve an llms.txt file. Treat it as optional housekeeping, not a ranking lever.

Which sites ChatGPT cites most in 2026

The most-cited domains tell you where ChatGPT looks when it builds an answer. If your brand is absent from those places, it is less likely to be named.

The freshest public snapshot is Ahrefs’ list of the 50 most-cited websites in ChatGPT, based on broad US queries across all topics in September 2026.

Three things stand out:

• Reddit and Wikipedia lead, but the lead moves. A Semrush study of 230,000 prompts (July to October 2025) found ChatGPT cited Reddit in close to 60% of responses in early August 2025, then around 10% by mid-September. Wikipedia fell from roughly 55% of responses to under 20% in the same period.
• Review and comparison sites are a big share. Consumer Reports, TechRadar, Trustpilot, Tom’s Guide and similar sites sit in the top 50.
• The list is consumer-heavy. It covers all topics, so B2B sources look smaller than they are in a SaaS category. In Profound’s analysis of 680 million citations (August 2024 to June 2025), G2 took 1.1% of all ChatGPT citations, the same as Forbes.

The lesson for a SaaS brand: find the sources ChatGPT cites for your prompts. A general top-50 list is a starting point, not a target list.

Listicles, reviews and “best X” articles

“Best X” articles are the most direct route to a brand mention, because they already have the shape of the answer ChatGPT is building: a shortlist of named options with reasons. They are also the format that changed most in 2026.

What the 2026 data says

• Listicles are a large share of citations. Seer Interactive analysed 2 million ChatGPT citations from November 2025 to February 2026. URLs containing “best” or “top” made up 16.6% of them.
• Self-serving lists are losing ground. In the same data, listicle citations fell 30% from December 2025 to January 2026, faster than the 22.7% drop in total citations. 13 of 16 industries saw declines. Wikipedia, Reddit and review sites such as G2 and Capterra gained share.
• Brand product pages are rising. According to David Konitzny’s analysis of the GPT-5.6 rollout in August 2026, summarized by Peec, listicles fell from 15.47% to 8.19% of retrieved pages on transactional prompts. Product pages led at 16.39%, usually on the brand’s own domain, driven by fan-out searches that use the site: operator.

Which lists still get cited

Seer looked at 78 listicles that gained citations while the rest declined. They shared four traits:

1. Fresh. Recently updated pages grew. Older lists that had been cited for years dropped.
2. Comprehensive. Growing pages listed 10 to 20 options. Declining ones often listed 5 to 8 with thin comparisons.
3. Backed by evidence. They used outside data such as G2 ratings or a disclosed methodology.
4. Structured. Every one used numbered lists, often with features, pros and cons, pricing and “best for” sections.

So the goal is not “get on any list”. It is to be included, with accurate detail, in independent, well-maintained comparisons that rank for the queries ChatGPT runs. That is what our brand mentions service focuses on.

Reddit, Wikipedia and review sites

These three source types play different roles:

• Reddit is retrieved constantly and cited less often than it is read, according to Ahrefs. It shapes ChatGPT’s sense of what real users recommend. Seer suggests checking Reddit threads in your category to see which brands users mention and whether your own messaging matches.
• Wikipedia is a top-two domain in every dataset above. Most brands do not qualify for an article, and Wikipedia’s conflict-of-interest rules discourage brands from editing their own pages. Focus on the press coverage that would make you notable instead.
• Review sites (G2, Capterra, Trustpilot, Consumer Reports) feed both citations and the “external validation” that the best listicles quote. Up-to-date profiles with recent reviews are basic hygiene.

Freshness: does ChatGPT prefer new content?

Yes, but freshness is not enough on its own. Ahrefs reports that in an earlier study of 17 million citations, ChatGPT cited URLs that were 458 days newer than Google’s organic results. In its 1.4-million-prompt study, the median cited page was still about 500 days old (Ahrefs).

Within one set of retrieved results, the newest pages were often the ones left out. Relevance did the heavy lifting. Freshness acted as a tie-breaker, most clearly for news queries, where cited pages were younger than non-cited ones.

In practice:

• Update comparison and pricing pages when facts change, not just the date.
• Expect freshness to matter most for “best X in 2026”, news and pricing prompts.
• Do not expect a new page to beat a relevant, established one just because it is new.

Which brand signals correlate with ChatGPT mentions

Citations are about pages. Mentions are about brands. Ahrefs studied 75,000 brands (December 2025) to see which signals correlate with being mentioned in ChatGPT, Google AI Mode and AI Overviews.

Ahrefs stresses that correlation is not causation. Still, the pattern is consistent: brands that are talked about widely, in articles, videos and forums, get named more. Link counts and content volume on their own show weak relationships.

Ahrefs also found that ChatGPT correlates less with classic authority metrics than Google’s AI features do. That may make ChatGPT a more open door for smaller brands. The AirOps data points the same way: about 74% of citations went to sites with a Domain Authority under 80.

This is why digital PR and earned mentions sit at the centre of AI visibility work. Links still help pages rank in the search results ChatGPT pulls from, so link building supports the retrieval step. But the brand mention itself is what gets you named.

What a brand can do: a 9-step plan

1. Allow OAI-SearchBot. Check robots.txt and your CDN or firewall rules. Decide separately whether to allow GPTBot.
2. Get indexed beyond Google. Confirm your key pages are indexed in Bing, since Microsoft is one of the search providers OpenAI names.
3. Build a prompt list. Write 20 to 50 prompts a buyer would ask: “best [category] for [use case]”, “[competitor] alternatives”, “[your brand] vs [competitor]”, “[your brand] pricing”.
4. Record who gets cited. Run the prompts (logged out or with memory off, in more than one mode) and note the cited URLs and named brands. Repeat monthly, because results move.
5. Fix your own pages. Make product, pricing, feature and comparison pages clear, current and readable without JavaScript. Use titles and URLs that match how people ask.
6. Cover the fan-out questions. Answer pricing, integrations, security and “who it is for” questions directly on your site. Fan-out queries often target exactly these.
7. Get into the lists ChatGPT already cites. Pitch the independent “best X” articles and reviews from step 4 with accurate, useful information. Prefer lists that are maintained and show their method.
8. Earn press and expert mentions. Original data gives journalists a reason to name you. Data-driven content and digital PR campaigns create mentions on sites ChatGPT trusts.
9. Keep third-party profiles consistent. Update review profiles, LinkedIn, directories and YouTube descriptions so your name, category and positioning match everywhere.

Limits: what no one can guarantee

Be careful with anyone who promises ChatGPT placements. The facts do not support it.

• OpenAI does not guarantee placement. Its help page says ChatGPT ranks results “using multiple factors” and that “placement is not guaranteed”.
• Answers vary. Location, memory, model and reasoning mode change the queries ChatGPT runs and the sources it cites. Only a quarter of cited domains overlapped between reasoning modes in the Semrush study.
• Source preferences shift without notice. Reddit and Wikipedia citations dropped sharply in September 2025. Listicles dropped in early 2026. Brand links jumped in May 2026. None of these changes were announced in advance.
• Training data is slow. Work that affects how the model itself describes you only shows up after future model updates.
• Most studies are correlational. They show what cited pages have in common, not what causes a citation.

A sensible goal is to raise the share of relevant prompts where your brand appears, track it over months and compare it with competitors. A single screenshot proves little in either direction.

FAQs

Does ChatGPT use Google or Bing for search?

OpenAI says ChatGPT search partners with third-party search providers and links to Microsoft’s and Shopify’s privacy policies. It does not publish a full provider list. Google rankings still correlate with ChatGPT citations, so it makes sense to be indexed and ranking well in both Google and Bing.

If I block GPTBot, will ChatGPT stop mentioning my brand?

No. GPTBot controls whether content may be used for model training. OAI-SearchBot controls whether your site can appear in ChatGPT search answers. You can block GPTBot and still allow OAI-SearchBot. Blocking OAI-SearchBot keeps your pages out of search answers, except as possible navigational links.

Can I pay to be recommended by ChatGPT?

ChatGPT now carries ads, and OpenAI uses a separate bot, OAI-AdsBot, to check ad landing pages. An ad is not the same as being named in the answer. We found no verified way to buy a place in ChatGPT’s organic answers. Mentions come from what ChatGPT finds across the web.

How long does it take for ChatGPT to cite a new page?

There is no fixed timeline. OpenAI says robots.txt changes can take about 24 hours to reach its search systems. After that, a page still needs to rank for the queries ChatGPT runs and answer them better than competing pages. Being crawlable makes a page eligible, not cited.

Why does ChatGPT give different brands for the same question?

Each run can rewrite the question into different searches and pull different results. Location, memory and the reasoning mode also change the sources. In a Semrush test with Kevin Indig, only 25.6% of cited domains overlapped between quick and deeper reasoning answers to the same prompts. Track trends over many runs.

Does llms.txt help my brand appear in ChatGPT?

OpenAI’s crawler documentation does not mention llms.txt, and we found no evidence that ChatGPT uses it to choose sources. It does no harm and many SaaS sites already serve one. The proven basics are allowing OAI-SearchBot, ranking for relevant queries and earning mentions on sites ChatGPT cites.

How can I track whether ChatGPT mentions my brand?

Run a fixed list of buyer prompts every month and log the cited URLs and named brands. AI visibility tools from Ahrefs, Semrush, Profound and others automate this at scale. Also watch referral traffic from chatgpt.com in your analytics, especially since brand names started linking to homepages in May 2026.

Get your brand into the sources ChatGPT reads

ChatGPT names the brands it keeps finding in trusted lists, reviews and press coverage. Create & Grow places brands in relevant “best X” articles, reviews and press pieces, and runs digital PR campaigns that earn mentions on sites AI assistants cite. Brand mention plans start at €1,500 a month.

We cannot guarantee that ChatGPT will mention you, and nobody honestly can. We can show you which sources it cites for your category and get you into the ones that matter. Contact us to start with a short audit of your prompts.

More guides in this series on how AI search picks sources: Google AI Overviews, Gemini, Perplexity, Claude, Copilot.