What Sources Does ChatGPT Cite? What the Data Says for Tech Companies
Wikipedia and Reddit lead the studies, but every ranking has a date, a sample and a seller. Here’s what the data shows, and how to read it before you act on it.
ChatGPT sources change by study, by month and by prompt. Wikipedia and Reddit lead in the named studies, but shares shift fast, so read any list of ChatGPT sources as a snapshot.

You ask ChatGPT a question about your own category. It answers, with a few small source links tucked under the text. None of them is yours.
So you go looking for the list. Which sites does ChatGPT actually trust? Somebody must have counted.
Somebody has. Several vendors have, and they don’t agree with each other. It makes sense why that’s irritating, because a list of sources feels like something you could simply act on, the way a keyword list feels like something you could simply write to.
But a ranking of ChatGPT sources isn’t a to-do list. It’s a measurement, and every measurement has a date, a sample, a unit and a seller standing behind it, whether or not the chart on the page tells you so.
Read it as a photo and you’ll learn something. Read it as a plan and you’ll build for last season’s answer. That’s the trap. The work that comes after the reading is what our generative engine optimization services are for.
The Short Version
Wikipedia and Reddit lead the named ChatGPT studies, but their share of responses fell by tens of points within about six weeks in 2025, so a ranking is only as good as its date. The studies also count different things, so their numbers can’t be lined up side by side, and in one sample of 29 technology brands, owned websites were cited more than twice as often as earned media. OpenAI documents three crawlers, and the one for ChatGPT search is OAI-SearchBot. The steps to act on all of it are in ChatGPT SEO for tech companies.
What Counts as a ChatGPT Source?
In these studies, a ChatGPT source is a domain that ChatGPT links to or references in an answer built with web search. Each study sets its own counting rules, so the same domain can look big in one report and modest in another.
Here’s how the four studies used on this page counted. Read the last column twice.
How Each Study Counted
| Study | When, where and which platforms | Sample | What it measures |
|---|---|---|---|
| Semrush, published November 10, 2025 | July 14 to October 12, 2025. ChatGPT search, Google AI Mode and Perplexity. Country not stated. | 230K prompts, over 100M citations | The share of prompt responses that cited a domain |
| Similarweb, published April 14, 2026 | January and February 2026. United States. ChatGPT in web browsing mode, and Google AI Mode. | “Nearly 600,000 citation events” | A domain’s share of all citation events |
| Ahrefs, published December 4, 2025 | Date range and country not stated. ChatGPT. | 750 prompts in software, products and agencies, and 10,000+ third-party URLs | The share of page types among the source links |
| MarTech, published February 27, 2026 | Run period not stated. ChatGPT, Perplexity, Grok and Google AI Mode/Gemini, blended. English. | 1,000+ prompts, 29 technology brands | A brand’s visibility, meaning how often it’s cited across a set of prompts |
A domain “cited in 60% of responses” and a domain with “12% of citation events” aren’t the same measurement. Different units. Different answers.
Semrush’s page uses “share of responses” and “citation share” in the same breath and never defines a citation. That’s not a scandal. It’s a reason to read the footnotes.
ChatGPT Sources by the Numbers: What Two Large Studies Found
Two studies cover ChatGPT at real scale. Both come from companies that sell tools for tracking AI visibility.
Semrush: July to October 2025
Semrush tracked 230K prompts every week from July 14 to October 12, 2025, across ChatGPT search, Google AI Mode and Perplexity, for more than 100M citations in its three-month study. On ChatGPT, Reddit appeared in close to 60% of prompt responses in early August, then fell to around 10% by mid-September. Wikipedia dropped from roughly 55% of responses to under 20%. Even after the fall, those two were still the domains ChatGPT cited more than any others.
Semrush’s summary also puts Reddit and LinkedIn among the top five cited domains across the three platforms, and says Google AI Mode cited LinkedIn in nearly 15% of its responses. We cover that side separately, in how LinkedIn posts end up in ChatGPT answers.
Big sample. Real limits, and they’re stated on the page. Semrush says the data doesn’t tell it exactly why the shifts happened, and the page doesn’t state a country or define a citation.
Similarweb: January and February 2026
Similarweb looked at U.S. prompts in January and February 2026 and described nearly 600,000 citation events across ChatGPT in web browsing mode and Google AI Mode, a window the page itself limits to the United States and to ChatGPT’s browsing mode. In ChatGPT, wikipedia.org took 13.15% of citation events and reddit.com took 11.97%. openai.com came third, at 6.21%. In AI Mode, fandom.com led at 7.16%.
Notice who’s third. ChatGPT’s maker.
One more limit. The page’s table counts imply a bigger total than the “nearly 600,000” it states, so treat its percentages as approximate. And Similarweb sells AI search tools too.
Why a Source Ranking Is a Photo, Not a Map
Line the two studies up by date and you can see the movement. In summer 2025, ChatGPT cited Reddit in close to 60% of responses. By mid-September, around 10%.
Similarweb, looking at January and February 2026 in the United States, still has Reddit second, at 11.97% of citation events, though its unit is different from Semrush’s, so the two numbers can’t be subtracted from each other. The direction still tells you something. Nothing held still.
Semrush says its data doesn’t explain the shift, so any article that tells you exactly why is guessing.
This is where teams go wrong. A chart says Reddit is huge, so the next quarter goes into Reddit.
Then the chart changes. More motion. No momentum.
A source ranking is a photo, not a map. It shows where things stood on one day, in one sample, for the questions someone chose to ask, and if your plan only works while one chart stays still, it isn’t a plan.
The better use of a ranking is as a question. Does this look like my category? Which brings us to the sample closest to a technology company.
What ChatGPT Sources Look Like for Tech Companies
Neither large study says it sampled technology buyers’ questions. The closest published sample is smaller, and it comes with its own flags.
The MarTech Study of 29 Technology Brands
Mike Maynard, managing director of Napier, a PR firm that specializes in the technology sector, ran more than 1,000 prompts for 29 technology brands through a tracking tool called Brandi and published the results in MarTech on February 27, 2026. The brands ranged from semiconductors and electronic components to software and logistics, predominantly enterprise and midsize.
Only 21% of the brands appeared in more than 25% of the prompt answers that mattered to them. One-third appeared in under 5%. Brands whose own domain sat in the top 10 cited domains averaged 25% visibility, and brands outside it averaged 7.6%. And more than twice as many owned websites as earned media sites were cited.
What the Study Reported for Four Well-Known Domains
| Domain | What it reported across the 29 brands |
|---|---|
| YouTube | Among the top 25 cited domains for about 75% of brands |
| In the top 25 for 37% of brands | |
| Wikipedia | Not in the top 25 for 60% of brands, and not cited in any answer for 30% |
| Not in the top 25 for more than four-fifths (83%) of brands |
Read that against the two big studies. Wikipedia and Reddit lead the broad counts. In this sample of technology brands, neither made the top 25 for a clear majority of them.
That’s worth pausing on. The ranking everyone quotes and the sample closest to you point different ways.

Now the flags. A tool wrote the prompts, from topics each brand said it wanted to rank for, so real buyers didn’t.
The four engines are blended, so it isn’t ChatGPT alone, the run period isn’t stated, the prompts were in English, and the sample leans enterprise and midsize, which is a narrower window than the headline numbers suggest. MarTech is owned by Semrush.
And the author’s own reading of the Reddit result, that it reflects the nature of these brands, isn’t tested on the page. Twenty-nine brands is a small room. It’s still the nearest room we’ve got.
Ahrefs on “Best” Lists
Ahrefs asked a different question. In a December 4, 2025 study, its author ran 750 top-of-the-funnel prompts across software, products and agencies and classified the pages that came back as source links.
“Best X” blog lists made up 43.8% of all page types. A self-promotional list still showed up in more than a third of ChatGPT responses in the software category when its publisher was the one recommended. And 35% of the cited best lists sat on low-authority domains.
Ahrefs sells the tools behind the study, and its author concludes that agencies and software companies should publish self-promotional lists. That’s the publisher’s view. The study counted appearances. It didn’t measure buyer trust or leads, it isn’t a study of managed IT buyers, and the author notes that the terms people type into Google may not match what they type into a chat.
What OpenAI Documents, and What It Leaves Out
OpenAI’s crawler overview names three user agents. They do three different jobs.
OpenAI’s Three Crawlers
| Crawler | What OpenAI says it does | What that means for your site |
|---|---|---|
| OAI-SearchBot | Surfaces websites in search results in ChatGPT’s search features | Sites opted out of it will not be shown in ChatGPT search answers, though they can still appear as navigational links. OpenAI recommends allowing it in robots.txt, along with its published IP ranges. |
| GPTBot | Crawls content that may be used in training generative AI foundation models | Disallowing it signals that your content shouldn’t be used for training. |
| ChatGPT-User | May visit a page when a user asks ChatGPT or a custom GPT a question | Not used to crawl automatically and not used to decide whether content may appear in Search. Because the visits are user-initiated, robots.txt rules may not apply. |
The settings are independent. OpenAI’s own example is a site that allows OAI-SearchBot so it can appear in search results while disallowing GPTBot, which is the setting that governs training and not search. And a robots.txt change can take about 24 hours to reach search.
Here’s what the page leaves out. It says nothing about how sources are ranked or chosen. That silence matters. It’s why every list in this post comes from outside measurement, from companies running prompts and counting what comes back. For the wider picture of what each assistant documents, we wrote up LLM optimization for tech companies.
How to Read the Next ChatGPT Sources Study
Another study lands next month. Run it through six questions before you repeat a number from it.
- Who’s selling? Every dataset above comes from a company that sells search or AI-visibility tools, or from a PR firm. That doesn’t void the numbers. It means the advice at the bottom of the page is the publisher’s.
- What counted as a citation? A link in the answer, or any mention of the domain? Some pages never say.
- Responses or events? Those are different numbers.
- When? January and February isn’t November.
- Where, and on which platform? One country and one platform is a narrow window. A blend of four engines is a different one.
- Who wrote the prompts, and can you see the sample size? A tool, a team or real buyers. A stated count, or a headline percentage with nothing behind it.
My working rule, and not a published standard, is to trust a number more when the page gives its date, country, platform, sample and unit, and less when it hands you a rank without them.
What to Do With a Source List
Use it to ask questions. Which sources does ChatGPT use for the prompts your own buyers type, and are you in them? That’s a test you run yourself, and more than once, because one answer is a sample of one. Our guide to prompt research covers where those prompts come from.
The steps sit elsewhere, so this page doesn’t repeat them. Check the door and publish what a buyer’s question breaks into in our ChatGPT SEO guide. Earn the mentions you can’t write yourself in off-site brand mentions. And before you assume forums matter for your category, read Reddit and forum answers for MSPs.
Then do it again. Quarterly is my working rule, not a standard. The sources move.
The better question isn’t which site ChatGPT cites. It’s which sources it reaches for when your buyer asks. You can’t read that off a chart, so start by asking it. If you’d like a second set of eyes, book a growth strategy session.
What Owners Ask About ChatGPT Sources
So which site does ChatGPT lean on the heaviest?
Wikipedia or Reddit, depending on the study and the month. Similarweb’s January and February 2026 U.S. sample gave Wikipedia 13.15% of ChatGPT citation events and Reddit 11.97%.
Semrush’s 2025 data had the same two on top even after their steep fall. Different units. Different dates. The order can flip.
Does ChatGPT cite a company’s own website, or only other people’s?
Both. In MarTech’s sample of 29 technology brands, more than twice as many owned websites as earned media sites were cited.
Ahrefs found a publisher’s own “best” list in more than a third of ChatGPT software responses when that publisher was recommended. Small samples, built by tool vendors. Treat both as directional.
If I block GPTBot, do I disappear from ChatGPT answers?
Not by that setting alone. OpenAI says each crawler setting is independent. GPTBot relates to training, and OAI-SearchBot is the one used to surface sites in ChatGPT search.
Check robots.txt for both. Allow a day for a change to reach search.
Why does every study give a different answer?
Different units, dates and samples. Semrush reports the share of responses that cited a domain, Similarweb reports a domain’s share of all citation events, and MarTech reports a brand’s visibility across four blended engines.
Put those side by side and you’re comparing three different things.
Can I apply to be one of ChatGPT’s sources?
Not through anything OpenAI’s crawler page describes. It covers crawler access, not a submission process or paid placement.
What it does say is that sites opted out of OAI-SearchBot won’t be shown in ChatGPT search answers, so access is the part you control.
How often should I recheck the sources for my category?
Quarterly, in my working view. That’s a rule of thumb, not a standard.
The 2025 swings show how fast the picture can change, so rerun the same prompts each time and log who gets named.
Find out which sources ChatGPT uses for your category
Talk to us about testing the prompts your buyers actually type, so you know which sources the assistants reach for and whether you’re among them.
Book a Growth Strategy Call