--- title: "How ChatGPT Search works and which sites it takes its answers from: an analysis of 450 questions in Italian" description: "ChatGPT rewrites every question into targeted searches, hands them to Microsoft, Shopify, and an index of its own, and on the selection criterion says only “placement is not guaranteed.” Across 450 questions about cars, motorcycles, and motorsport asked in Italian it cited 25 sources per answer: Sky, government sites, and manufacturers at the top, almost no social platforms. What it takes to get in." url: https://www.thinkai.it/en/blog/chatgpt-search-how-it-works-sources/ date: 2026-10-16 modified: 2026-10-19 author: "Pino Barbiera" image: https://www.thinkai.it/wp-content/uploads/2026/10/ChatGPT-Search.jpg categories: ["English"] type: post lang: it --- # How ChatGPT Search works and which sites it takes its answers from: an analysis of 450 questions in Italian In brief - ChatGPT Search does not pass your question to a search engine as it is: it rewrites it into several targeted searches, sends them to outside engines (OpenAI names Microsoft and Shopify) and to an index of its own, then composes the answer and adds links to the sources. On how it picks them, OpenAI says only “multiple factors, placement not guaranteed.” - Across 450 questions about cars, motorcycles, and motorsport asked in Italian, ChatGPT cited an average of 25 sources per answer, against 11.6 for Perplexity and 5.2 for Google’s AI Overviews. At the top: Sky, government sites, and manufacturers’ sites. YouTube and Facebook, which dominate AI Overviews, rarely appear. - There is one technical condition for getting in: let the OAI-SearchBot crawler through. A site that blocks it disappears from the answers and survives only as a navigation link. ChatGPT has searched the web since October 31, 2024, and since February 5, 2025 it does so for everyone, including people without an account. It decides on its own when a question calls for a search and, when it searches, it lists the sources under the answer. How it chooses them is something OpenAI explains only in part. The rest can be read in the answers. We analyzed the answers to **450 questions about cars, motorcycles, and motorsport asked in Italian**, and the result is very different from Google’s. Index - [How a ChatGPT search works](https://www.thinkai.it/en/blog/chatgpt-search-how-it-works-sources/#How_a_ChatGPT_search_works) - [How it picks sources, according to OpenAI](https://www.thinkai.it/en/blog/chatgpt-search-how-it-works-sources/#How_it_picks_sources_according_to_OpenAI) - [Which sites it cites in Italian: 450 questions, 423 domains](https://www.thinkai.it/en/blog/chatgpt-search-how-it-works-sources/#Which_sites_it_cites_in_Italian_450_questions_423_domains) - [Products follow separate rules](https://www.thinkai.it/en/blog/chatgpt-search-how-it-works-sources/#Products_follow_separate_rules) - [What changes for a brand](https://www.thinkai.it/en/blog/chatgpt-search-how-it-works-sources/#What_changes_for_a_brand) - [Frequently asked questions](https://www.thinkai.it/en/blog/chatgpt-search-how-it-works-sources/#Frequently_asked_questions) ## How a ChatGPT search works The difference from a search engine starts at the beginning: ChatGPT **does not send your question as you typed it**. It rewrites it. OpenAI explains that the question is turned into [one or more targeted queries](https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt) sent to outside engines, and that after the first results more specific queries may go out to other engines. The example is OpenAI’s own: someone asking “what’s the latest on the development of drugs that target CCR8 for cancer” first generates the query “CCR8 immunotherapy drug development 2025,” then “CHS-114 conference 2025.” The outside engines have names. On the same page OpenAI links to the privacy policies of **Microsoft** and **Shopify**: the first for the web, the second for products. Added to these are the content that publishers under contract with OpenAI provide directly, and **OpenAI’s own index**: for companies with strict security requirements there is a mode that uses only OpenAI’s [indexed and cached web content](https://help.openai.com/en/articles/20001203-offline-web-search-for-chatgpt-workspaces), with coverage that varies from site to site, by language and by region. Two details affect the outcome. If memory is on, ChatGPT uses what it knows about the user to rewrite the search: for someone who has said they are vegan and live in San Francisco, a request for restaurants becomes “good vegan restaurants San Francisco.” The approximate location, inferred from the IP address, is passed on to the outside engines. The same question, asked by two different people, produces two different searches and therefore two different lists of sources. ## How it picks sources, according to OpenAI On the selection criterion the official statement is a single sentence: *“ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed.”* Multiple factors, placement not guaranteed. Google, by comparison, has published guidelines hundreds of pages long for its quality raters for years. What OpenAI does describe precisely is the **entry condition**. The search crawler is called OAI-SearchBot and is distinct from GPTBot, which collects content for training: a site can block GPTBot and allow OAI-SearchBot, and sites that block OAI-SearchBot “[will not be shown in ChatGPT search answers](https://developers.openai.com/api/docs/bots), though can still appear as navigational links.” A change to robots.txt takes about 24 hours to register. A third agent, ChatGPT-User, opens a page when the user asks for it and has no bearing on presence in search. On quotidianomotori.com, the online motoring daily we run, OAI-SearchBot is the most active bot in the Cloudflare logs: 12,370 requests for the Markdown versions of pages in four months, 42% of all search bots. Then there are the publisher agreements. At launch OpenAI listed fourteen groups it had consulted, from **Associated Press** to **Le Monde**, from **Axel Springer** to **GEDI**, and in January 2025 it counted “[nearly 20, with more than 160 news outlets](https://openai.com/index/partnering-with-axios-expands-openai-work-with-the-news-industry/) in more than 20 languages,” with agreements that let the search “feature select summaries and excerpts with clear citations and direct links.” On the same page OpenAI adds that any site can appear, with or without an agreement. The two channels coexist: partners bring their text inside the answer, other sites come in as cited sources. ## Which sites it cites in Italian: 450 questions, 423 domains Between September 16 and October 16, 2026, we asked the same 450 questions about cars, motorcycles, and motorsport, in Italian, to ChatGPT, Perplexity, and Google’s AI Overviews. ChatGPT listed **11,251 sources, 25 per answer**, against 11.6 for Perplexity and 5.2 for AI Overviews, across 423 distinct domains. The difference from AI Overviews, which on the same questions use [five sources per answer](https://www.thinkai.it/en/blog/ai-overviews-what-they-are-where-answers-come-from/) and draw on Facebook 42% of the time and YouTube 41%, lies in the kind of sources even before the number. The domains ChatGPT uses most often, motorsport included, with the share of answers in which they appear: **sky.it** 32.6%, with its TV listings, sports, and news subdomains; **gov.it** sites 31.9%, from the ACI motoring authority to the ENAC aviation authority to the Ministry of Transport; **gazzetta.it** 26.3%; **aci.it** 18.9%; formula1.com 14.8%; oasport.it 14.2%; [quotidianomotori.com](https://www.quotidianomotori.com/) 13.7%, in seventh place; motogp.com 13.5%; **honda.it** 12.6%; ansa.it 12.1%. Next come three trade publications between 7% and 11%, Reuters at 10.1%, European Union sites at 6.7%, the Italian driver’s portal at 5.6%, then the Official Gazette, the Normattiva law database, and the Chamber of Deputies. Social platforms, which dominate AI Overviews, rarely appear on ChatGPT: **YouTube in 3.8% of answers**, Reddit in 2%, Facebook in 1.1%. Wikipedia in 2.9%. In their place are the **manufacturers’ sites**: Honda, Stellantis at 6.7%, then SYM, Kymco, Yamaha, Piaggio, Mazda, and BYD. On technical questions in Italian, ChatGPT prefers the official source to the conversation: the site of whoever builds the product, the body that writes the rule, the outlet that publishes the race schedule. In the United States the picture is different: a company that sells monitoring services counted, across 30 million citations, [Wikipedia, Reddit, Forbes, TechRadar, and LinkedIn](https://peec.ai/blog/top-domains-cited-by-ai-search-analysis-based-on-30m-sources) at the top on ChatGPT, but on questions in English and across every industry. The last data point concerns **traffic**. ChatGPT puts the links at the bottom of the answer and people open them: of the 1,443 human visits that reached Quotidiano Motori from AI assistants in four months, [1,072 came from ChatGPT](https://www.thinkai.it/en/blog/llms-txt-what-it-is-who-reads-it/), 74%, more than all the other assistants combined. ## Products follow separate rules When the question is about a purchase, ChatGPT shows product cards with images, prices, and links to stores. Here OpenAI is more explicit: products “are [selected independently by ChatGPT](https://help.openai.com/en/articles/11128490-shopping-with-chatgpt-search) and are not ads, nor influenced by any OpenAI partnerships.” The choice rests on the **structured metadata** supplied by sellers and intermediaries, such as price and description, on what the model already knows before searching, and on OpenAI’s safety rules. Stores are ranked by availability, price, quality, and whether they are the maker or the primary seller of the item. Labels such as “Budget-friendly” or “Most popular” and the review summaries are written by the model, and OpenAI warns that they “*are not guarantees or verified statements*.” For stores on Shopify the data arrives automatically from the catalog; others can send a product feed. This is why [paid ads inside ChatGPT](https://www.thinkai.it/en/blog/from-geo-to-gem-ads-inside-ai-answers/) stay separate from the answers and the product cards: buying a slot is one thing, being the source the model picks is another. ## What changes for a brand Three things can be done, one has to be understood. First: **check robots.txt**. In 2023 many sites blocked OpenAI’s bots wholesale, when GPTBot was the only one. Today that block removes the site from search answers, which is the channel that brings visits. The two settings are distinct and can be managed separately. Second: **write pages that answer a precise search**. ChatGPT turns the question into targeted queries and picks the pages that satisfy them. A page that answers a single question, with data, dates, and a title that says what it is about, gets found; a generic page does not. This is [the principle of GEO](https://www.thinkai.it/en/blog/what-is-geo-generative-engine-optimization/), generative engine optimization, applied to the letter. Third, for anyone who sells: **structured product data**, with price, availability, and description, is the only thing OpenAI says it considers for the product cards. Without it, the product exists only if someone else talks about it. The thing to understand: **nobody can promise a citation**. OpenAI writes “placement is not guaranteed,” rewrites the question differently for each user, and changes sources from one session to the next. Whoever guarantees first place on ChatGPT is selling something they do not control. What can be measured, and improved, is how often you appear in the answers to a set of questions repeated over time. ## Frequently asked questions **Does ChatGPT Search use Bing?** OpenAI says the search “partners with other search providers” and, in its official help page, links to the privacy policies of Microsoft and Shopify. Added to these are the content of its publishing partners and OpenAI’s own index. The name Bing does not appear, but Microsoft is the only web search engine named. **How do I get my site to appear in ChatGPT Search?** The technical condition is to allow the OAI-SearchBot crawler, in robots.txt and from the IP addresses OpenAI publishes: a site that blocks it stays out of the answers. Then the pages matter: ChatGPT rewrites the question into targeted queries and picks the pages that answer them precisely. OpenAI guarantees placement to nobody. **How many sources does ChatGPT cite in an answer?** Across 450 questions about cars, motorcycles, and motorsport asked in Italian between September and October 2026, ChatGPT listed an average of 25 sources per answer, against 11.6 for Perplexity and 5.2 for Google’s AI Overviews. At the top: Sky, gov.it government sites, Gazzetta, ACI, and manufacturers’ sites; YouTube and Facebook appear in less than 4% of answers.