How ChatGPT Chooses Sources and How to Get Cited

ChatGPT search is the feature in ChatGPT that runs a live web search, reads the pages it retrieves and cites them in the answer, and ChatGPT uses it on about a third of queries: 34.5% in Semrush's February 2026 sample. There are two kinds of ChatGPT answer, and only one of them can mention your page. The first comes from the model's training data, with no search and no citations; it's what the model remembers, and it can be out of date or simply wrong about you. The second comes with search, and sources appear as linked citations beside the text. A person can also switch search on for a prompt. OpenAI publishes which crawler surfaces pages for ChatGPT search (OpenAI, bot documentation), but it doesn't publish the rule that decides when a prompt triggers a search, so the first thing to know is that it doesn't always happen.

No, ChatGPT does not search the web for every question; Semrush measured a web search on 34.5% of queries in February 2026, so roughly two in three answers come from the model alone and carry no citation. The practical consequence is that a page can only be cited on the prompts that trigger a search, which is why ChatGPT citation studies count "responses with a citation" rather than all responses. OpenAI hasn't published which prompts trigger one. What the studies cited on this page observe is that prompts asking for current, specific or comparative information tend to; a question about which provider to choose this year is far more likely to send ChatGPT to the web than a question about what a term means. So when you test your own brand later on this page, you'll turn search on deliberately. First, the crawler that does the surfacing.

OAI-SearchBot decides what ChatGPT can cite: OpenAI's bot documentation says it "is used to surface websites in search results in ChatGPT's search features", while GPTBot is the training crawler and ChatGPT-User fetches a page only when a user asks for it. A crawler is an automated fetcher of pages, and a user agent is the name it identifies itself with in robots.txt and in your server logs. Three OpenAI user agents visit sites, and most robots.txt advice online treats them as one. They aren't, and the table below is the distinction this page exists to make.

OpenAI says robots.txt changes take about 24 hours to register, so a fix made today is live tomorrow. What blocking the training crawler does and does not do deserves its own answer, because it's the most common mistake.

No, blocking GPTBot does not on its own stop ChatGPT from citing a site, because ChatGPT search surfaces pages through OAI-SearchBot; blocking OAI-SearchBot does remove a site from ChatGPT search results. One caveat comes from OpenAI's own documentation: OpenAI may reuse one crawl for both purposes, so a site that allows OAI-SearchBot and blocks GPTBot should expect its pages to be fetched once, and should read OpenAI's current wording before relying on the split.

Whichever you choose registers in about 24 hours. Where the search results themselves come from is next.

ChatGPT uses Bing as one of its search providers: Microsoft made Bing the default provider for ChatGPT browsing in May 2023, and OpenAI's VP of Engineering said in a November 2024 Reddit AMA that "we use a set of services and Bing is an important one". The evidence that this matters comes from Seer Interactive, which reported that 87% of SearchGPT citations matched Bing's top results in its 2024 to 2025 sample. Read that as a reported figure from one sample, not a law; OpenAI has since said it uses a set of services, and the number may well have moved. The practical reading hasn't: a page that Bing does not index has little route into a ChatGPT citation. If you've spent years thinking of Bing as the search engine nobody uses, this is the moment to open Bing Webmaster Tools. It's free, it takes an afternoon, and it's the part of ChatGPT work most sites skip. What happens after the search returns results is next.

ChatGPT selects sources by sending one or more searches to its providers, fetching the pages that come back, keeping the passages the model used to write the answer, and citing those pages, which comes to about 15 sources per answer (Semrush 2026 AI Visibility Index, 126M prompts). OpenAI publishes no ranking function for this step; everything after retrieval is inferred from citation studies, and anyone who tells you otherwise is guessing.

Fifteen sources per answer is more than Google's AI features cite, which means more room on the list, and also more competition for the eye. Which part of a page the passage tends to come from has a surprisingly precise answer.

ChatGPT cites the top of a page most: 44.2% of citations come from the first 30% of a page (Growth Memo via Search Engine Land, 18,012 ChatGPT citations, 2026). The same study found that cited pages read at a Flesch-Kincaid grade of 16 against 19.1 for pages that were not cited. Both are above a general audience's reading level, so the honest reading is "cited pages are clearer than their competitors", not "write simply". That finding is why this page, and every guide here, puts the direct answer, the key takeaways and the main visual in the first 30%: it's the part that gets read, by machines and people alike. Which domains receive those citations is next.

ChatGPT cites Wikipedia most of all, at 7.8% of about 680 million citations between August 2024 and June 2025 (Profound), followed by Reddit at 1.8%, Forbes at 1.1% and G2 at 1.1%, and .com domains take 80.41% of its citations. Before you despair at competing with Wikipedia, notice how flat the list is below the top entry. The long tail of ordinary sites holds most of the citations, and that tail is where a specific page about a specific question lives.

Page types tell a different story from domains. Surfer's September 2026 analysis of ChatGPT's top 50 cited domains put clutch.co second and g2.com third, with review and directory sites at 23.6% of top-50 citations. And here is a conflict to weigh rather than hide: a DerivateX study of B2B SaaS prompts found G2 and Capterra at zero citations in its sample. The difference is scope, all-category top-50 domains against one vertical's prompts, and the lesson is to test your own category with the prompt panel further down rather than assume either figure. The Profound numbers also predate a shift, which the chart above shows.

Reddit and Wikipedia citations fell in ChatGPT during the second half of 2025, and OpenAI has not said why: Semrush (230,000+ prompts, July to October 2025) measured Reddit falling from about 60% to about 10% of ChatGPT responses with a citation, and Wikipedia from about 55% to under 20%, by mid-September 2025, while PRNewswire, Forbes and Medium gained. Two measurements that look like they contradict each other don't. Profound's 7.8% is Wikipedia's share of all citations across eleven months; Semrush's 55% is the share of responses that cited Wikipedia at all. Different units, both true; the direction of Semrush's series is the point. Ahrefs added another series in November 2026: OpenAI's partner publishers were 7.5% of ChatGPT's top-200 citations in March 2026, 15.8% in July and 12.5% in late September. The source mix moves within months. That's why any "most-cited domains" list needs a date next to it, and why your own panel beats a published list. Two page attributes hold up across all of these shifts.

Yes, ChatGPT favors original research and recently updated pages: Growth Memo's research index finds that primary research is rare but far more citation-dense than ordinary pages when it appears, and 73% of ChatGPT-cited pages were updated within the last year (Seer Interactive, 47K citations, 2026). Ahrefs adds the URL-age finding: AI-cited URLs average 1,064 days old against 1,432 for organic results, so cited pages are both younger as URLs and more recently updated. Seer's reading, in paraphrase, is that the freshness being rewarded comes from updating, not only from publishing new pages. And "original research" is within reach of a company of any size: a survey of your own customers, your own aggregated data, or a test you ran, with the method published alongside the result. The research doesn't need to be large. It needs to be yours, and dated. A citation is not the same as being named, though, and the next distinction is the one that confuses most dashboards.

A ChatGPT mention is a brand's name appearing in the answer text, while a ChatGPT citation is a linked source the answer was drawn from, and only about 20% of ChatGPT brand mentions include a clickable link (Topify, 2026). If a client or a colleague has told you "ChatGPT recommends us" and your analytics show nothing from chatgpt.com, this is why. Both of them are right.

A brand can have many mentions and no citations. It shows up as ChatGPT recommending you while no visit ever arrives, which is a good problem and a confusing one. Only the citation produces the chatgpt.com referral that analytics can see; the mention has to be counted another way, and the prompt panel further down does that. How to measure AI Overview, AI Mode and ChatGPT traffic covers the referral side. How much of all this tracks Google is next.

ChatGPT citations overlap with Google rankings far less than with Bing results: only about 12% of ChatGPT, Gemini and Copilot citations ranked in the Google top 10 (Ahrefs, 15,000 prompts, August 2025), while Seer Interactive reported 87% of SearchGPT citations matching Bing's top results. Note that the Ahrefs figure pools three engines, so it is a ceiling for ChatGPT alone rather than an exact share.

Ranking in Google is not the route into ChatGPT; being indexed and ranked in Bing, and being the kind of page the model fetches, is. Google's own engines select from the Google index with query fan-out, and how Google AI Overviews and AI Mode choose their sources holds that mechanism in full. From here on, this page is about what to change.

To get cited by ChatGPT, a business changes seven things: it allows OAI-SearchBot, gets indexed in Bing, opens every page with the answer, serves server-rendered HTML, updates pages instead of only adding them, lists on the directories ChatGPT already cites, and publishes original research. Three of the seven raise the most questions and get their own sections below. The other four close here, with the evidence already on this page and a note on whose job each one is, because half of these are a developer's afternoon and the other half are a writer's month.

The first two changes are the gate, so they come first.

Yes, allow OAI-SearchBot and get indexed in Bing, because OAI-SearchBot is the crawler OpenAI names for ChatGPT search and Bing is the search provider OpenAI calls "an important one".

Whether the page can be read once it is fetched is the next question.

Yes, serve server-rendered HTML, because no major AI crawler renders JavaScript (Vercel and MERJ), so content that appears only after a script runs is invisible to OAI-SearchBot. The check is simple and most people have never done it: view the page source, not the rendered page in your browser's inspector, and confirm the answer paragraph, the headings and the tables are in it. Prices, reviews and specifications loaded by script are the common gaps; the page looks complete to a person and empty to a crawler. Two sentences on llms.txt, since it comes up: Ahrefs found 97% of llms.txt files are never read by any crawler, and Google says it ignores the file. It is not a substitute for crawlable HTML, and this page doesn't recommend relying on it. The off-site question is next.

Yes, directory listings on sites such as Clutch and G2 help with ChatGPT citations in most categories, because clutch.co and g2.com were the second and third most-cited domains in ChatGPT's top 50 (Surfer, September 2026) and review and directory sites took 23.6% of those citations, with one conflicting study to weigh. That study, from DerivateX, found G2 and Capterra at zero citations in its B2B SaaS sample. So run your category's own prompts before paying for a listing, and treat a directory profile as a mention source first and a citation source second. For a listing to be useful, it needs a complete profile, the same name and description as your site, and reviews with dates. Services businesses tend to find Clutch is the one that matters; software companies, G2 and Capterra. How to know whether any of it worked is next.

You check whether ChatGPT cites your brand by running a fixed panel of buyer prompts in ChatGPT with search on, recording every mention and every cited URL, and repeating the same panel on a schedule.

Keep the mention and citation columns separate; the 20% figure above is the reason. AI visibility monitoring tools such as Profound, Peec, Otterly and the Semrush AI toolkit automate the same panel, and they earn their place once the manual version has shown you what matters in your category. What the work costs is next.

Getting cited by ChatGPT costs the same as other AI search work at an agency, because it shares the pages, the updates and the mentions with SEO; the ChatGPT-specific parts are crawler access, Bing indexing and the prompt panel. Agencies price it as a monthly retainer or an add-on to one, with the access checks done inside a one-off audit.

No agency can promise a citation, and a flat result on the panel is still a result. How Noorscale runs it is next.

Noorscale gets clients cited by ChatGPT with the same program it runs for Google: crawler and Bing access first, answer-first pages on a semantic topical map, directory and third-party mentions, and a monthly prompt panel that records mentions and citations separately. The method comes from the founder's training in semantic SEO (a Semantic SEO Expert Certificate recognized by Koray Tuğberk GÜBÜR, March 2024). AI search optimization is where the program is scoped and where you can see what a month of it contains. Noorscale's published client results are search results, not ChatGPT figures, so none is quoted here; when there is a ChatGPT result with a control behind it, it will be on this page with its date.

This week, do two things that cost nothing. Open your robots.txt and look for OAI-SearchBot, by name or under a wildcard. Then open ChatGPT, switch search on, type the question your best customer asked before they found you, and write down who is cited and who is merely named. If you're in neither column, you now know which of the seven changes comes first.