Reading time: 13 minutes
ChatGPT does search the internet when a question needs current or specific facts, like news, stock quotes or nearby businesses. In other cases, however, it uses what it learned in training. So even though ChatGPT has internet access on every plan, it does not always search the web. For brands, this decides whether a new page gets into the answer.
In short
- According to OpenAI, ChatGPT may search the internet on its own when current information would improve the answer. You can also turn on search in the tools menu.
- In an AI+Automation test of 391 questions, search showed up in 73% of discovery questions but in only 10% of informational ones.
- A ChatGPT answer that used web search may include citations and the Sources button. On the other hand, with no web source at all, the answer tends to come from training.
- ChatGPT has internet access on every plan, even without logging in. However, only questions that trigger a search open the door to a brand’s new pages.
Does ChatGPT search the internet on its own or only when I ask?
Both. ChatGPT may search the internet on its own when current information would improve the answer. You can also ask it to search. So the decision changes from question to question. That is why many answers come from training alone, without checking a single page.
According to OpenAI’s Help Center, web search works on chatgpt.com and in the ChatGPT mobile and desktop apps. But search is a tool the model can use, not a step in every answer. On the page Does ChatGPT tell the truth?, the company sums up the rule like this:
“While ChatGPT does not search the web unless specific tools are enabled, these tools are enabled for all models by default.”
OpenAI, Help Center (2026)
In other words, ChatGPT has internet access by default. Still, ChatGPT only runs a web search when it judges that search will help. Semrush, which sells SEO tools, says most answers still come from training alone. In the U.S., it saw search on 34.5% of queries in February 2026. The full path, from training to memory, is in the guide on how ChatGPT works.
How to make ChatGPT search the internet
If you need the answer to use the web, turn on search before you send the question. According to OpenAI, there are three ways:
- Tools menu: select View all tools, then Search.
- Shortcut: type / in the message box and choose Search.
- Existing answer: in the refresh control, choose Search the web, when the option appears.
In a voice conversation, just ask ChatGPT to search. In every case, however, searches count toward your plan’s usage limits.
What kind of question makes ChatGPT search the internet?
ChatGPT will mostly search the internet when the answer depends on new or specific facts. Examples include news, sports scores, stock quotes, weather and nearby places. For definitions or general knowledge, however, it tends to answer from what it learned in training. In other words, having internet access does not make ChatGPT search on every question.
The official cases come from OpenAI itself. At the launch of search in October 2024, the company cited sports scores, news and stock quotes. Today, the Help Center recommends search for current events, unfamiliar topics and source-backed answers. In addition, ChatGPT can use your approximate location to show local results, such as restaurants and weather.
OpenAI has also tuned search over time. According to the release notes, ChatGPT began running multiple searches automatically for complex questions in June 2025. Then, in September 2025, the company said search had gotten better at detecting shopping intent.
What tests show about the questions that trigger search
AI+Automation, a consultancy that sells AI SEO services, tested when ChatGPT will search the internet. It used 391 questions about brands and products, taken from Google’s “People Also Ask” and autocomplete. Each one ran once, in the ChatGPT interface, on paid accounts. Overall, ChatGPT web search showed up in 42% of them. However, the type of question mattered more than the topic:
| Question type | Example from the test | With search |
|---|---|---|
| Discovery | “best CRM for small business” | 73% |
| Review-seeking | “best service desk software reddit” | 58% |
| Validation | “Is Mailchimp good for beginners?” | 44% |
| Comparison | “Salesforce vs HubSpot” | 29% |
| Informational | “What is a CRM?” | 10% |
According to the consultancy, unknown brands led to a search, while ChatGPT usually answered questions about famous brands from training. For example, the comparisons that did search included at least one lesser-known brand.
Semrush, looking at U.S. browsing data, also found that queries with search were shorter. In early 2026, they averaged 8.7 words, compared with 13.5 for queries without search. In other words, search shows up more on short prompts, although that gap has narrowed since early 2025.
How can you tell if a ChatGPT answer came from the web or from training?
Look for sources. When ChatGPT does search the web, the answer may show in-text citations and a Sources button. The button lists the links it used. But with no web source at all, the answer tends to come from training. Training only goes up to the model’s cutoff date.
According to OpenAI’s Help Center, three signals show whether ChatGPT did search the internet, and one test helps you compare:
- In-text citations: select a citation to open the source. On desktop, point to it to see a preview.
- Sources button: when it appears, it shows the cited sources and other relevant links.
- Search icon: check whether search or deep research is turned on in the message box.
- Search test: in the refresh control, choose Search the web and compare the two answers.
A link in the answer, however, does not always prove there was a search. In the AI+Automation test, for example, six answers had links that came from training, with no web lookup at all. So check the Sources button and the search icon too.
A cited source does not guarantee the facts either. OpenAI warns that search results and citations can be incomplete, outdated or incorrect. In other words, search improves accuracy but does not end made-up facts, the topic of the guide on AI hallucination. When an answer gets a company wrong, the source of the error changes the fix. The guide on why ChatGPT makes mistakes explains each case.
What is RAG, and how does it decide which pages make it into the answer?
RAG is the technique of retrieving content and adding it to the model’s prompt before it generates the answer. The acronym stands for retrieval-augmented generation. In ChatGPT web search, that retrieved content comes from pages on the internet.
OpenAI compares RAG to an open-book exam. Without it, the model answers from what it memorized in class, that is, in training. With it, on the other hand, the model looks up the material on the spot. The same guide warns, however, that RAG only works with the right material. After all, too much irrelevant context buries the useful information and leads to hallucinations.
In practice, when ChatGPT does search the internet, picking the pages follows four steps:
- Decides whether to search: in the API, OpenAI describes search as a tool the model may or may not use, depending on the prompt.
- Rewrites the question: sometimes ChatGPT turns the prompt into one or more queries and sends them to partner search providers.
- Ranks the results: the company says it uses multiple relevance and reliability factors, with no guaranteed placement.
- Reads and cites: the model gets part of the content from the results and cites the most relevant references.
To compete here, a site must allow crawlers such as OAI-SearchBot. The steps are in the guide on how to show up in ChatGPT.
Why does ChatGPT read more pages than it cites?
Because citations show only the most relevant references. According to the API web search docs, the list of pages consulted is often longer than the list of citations. Likewise, in the app, the Sources button shows other relevant links besides the cited sources. So when ChatGPT does search the web, a page can help shape the answer without getting an in-text citation.
In reasoning models, in fact, the search goes further, according to the same docs. The model searches, opens pages, looks for passages inside them and decides whether to keep searching. Meanwhile, in the AI+Automation test, 92% of answers with search fired two queries. Those answers cited a median of six sources.
Does the free version of ChatGPT search the internet?
Yes. According to OpenAI, the free version of ChatGPT can search the internet. So can the Go, Plus, Pro, Business, Enterprise and Edu plans. Even people who use ChatGPT without logging in get internet access. Across plans, what changes is the usage limit and the model that decides when to search.
In 2023, ChatGPT could already browse the web, but only on paid plans, according to the release notes. ChatGPT web search as it exists today, however, launched on October 31, 2024, first for Plus and Team users. Later, on December 16, 2024, it reached all logged-in users. Finally, since February 5, 2025, according to OpenAI, it has worked even without an account. So pages that treat search as a paid feature are out of date.
What changes between plans and models
First, the limit. OpenAI says searches follow each plan’s usage limits, and Plus has higher limits than Free. Second, the model. According to the release notes, Free and Go use Luna, the lowest-cost model in its family. For harder questions, these plans have the Think button. On Plus and Pro, on the other hand, users choose how much ChatGPT thinks.
Since the model decides when to search, the same question can take different paths on each plan. So a test run on Plus may not show what Free users see. In addition, at companies, internet access in ChatGPT also depends on the admin. Admins can turn off web search for each role.
What changes for a brand when ChatGPT uses internet search?
It changes what the brand can influence right now. When ChatGPT uses internet search, the pages the search engine finds today come into play, including new ones. Without search, however, only what the model learned up to its cutoff date counts. That is why the diagnosis starts by separating questions with search from questions without it.
For example, picture a buyer asking ChatGPT for packaging suppliers in São Paulo. If the answer shows sources, the race is for the pages search found, from the brand’s site or third parties. If not, the list probably came from training, so a new site won’t appear in that answer.
The difference matters more for well-known brands. In the AI+Automation test, answers about those brands usually came from training, which may hold old data. Unknown brands, by contrast, led to a search, so the site and third-party pages shape what appears.
How to separate questions with and without search
- List the client’s questions in the language its customers use: discovery, comparison, reviews and brand.
- Run each one in a temporary chat, several times, and note whether ChatGPT ran a web search and which domains appeared.
- Sort the questions into three groups: always with search, sometimes with search and without search.
- For questions where ChatGPT does search the internet, check whether the site appears among the sources and which third-party pages hold the citations.
- For questions without search, compare the answer with the brand’s current data and work on mentions in third-party sources.
Run the test in the interface, not just through the API, because the two can decide differently. In a smaller AI+Automation test with 24 questions through the API, every comparison triggered a search. In the interface, with other questions, only 29% did. The full data collection method is in the guide on how to measure brand mentions and citations.
If the brand is missing from every group, see why AI doesn’t cite your brand and how to test each cause.
What we found analyzing 9 Portuguese-language pages on ChatGPT search
We analyzed the 9 pages in Brazilian Portuguese that showed up most across 6 searches. Two of the searches were “chatgpt pesquisa na internet” (“does ChatGPT search the internet”) and “chatgpt tem acesso a internet” (“does ChatGPT have internet access”). All of the pages came out between 2023 and 2025.
In addition, 6 present an access condition that has since changed as if it were current. Examples include search as a paid feature, required login or a browser extension. Only 4 say ChatGPT decides on its own when to search, and only 2 say search works without logging in. Finally, none links to OpenAI’s Help Center, and only 1 mentions sites that want to appear in the answers.
Frequently asked questions
Because the Search tool may be unavailable in that chat, in the app or on your plan. ChatGPT still has internet access by default. So OpenAI suggests starting a new chat and looking for Search in the tools menu. Next, update the app and check your account and usage limit. At companies, ask your admin whether web search is turned on for your role.
No. According to OpenAI, search quickly brings recent information from the web and returns a short summary with links. Deep research, on the other hand, takes longer, reads and analyzes many sources and delivers a detailed, cited report. In addition, deep research usage varies by plan.
No. OpenAI warns that technical issues, paywalls or robots.txt rules can keep the model from getting information from a site. To appear in answers with search, a site must allow the OAI-SearchBot crawler. Its host or CDN must also accept the crawler’s IP addresses.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- OpenAI, Searching the web with ChatGPT (2026).
- OpenAI, Does ChatGPT tell the truth? (2026).
- OpenAI, Introducing ChatGPT search, with the availability updates (2024).
- OpenAI, ChatGPT release notes (2026).
- OpenAI, API web search documentation (2026).
- AI+Automation, test of 391 queries on when ChatGPT searches the web (2026).
- Semrush, analysis of 17 months of ChatGPT clickstream data (2026).
