Updated on 7 October 2026, by João Leite, Founder of murmur.

AI answers with what it learned in training and, when the question calls for current information, it searches the web at that moment. Training happens before launch: the model learns from a mass of texts gathered up to a certain day, the cutoff date. In search, the AI looks for pages at the moment of the question, writes with what it found and can show the sources. ChatGPT and Claude decide on their own when to search, and the person asking can also request the search. Google AIO, an AI-written summary at the top of Google search, builds its text on the results of Google search itself.

  • AI uses both, training and web search. Without searching, it answers with what it read up to the cutoff date; when it searches, it writes from the pages it found at that moment.
  • Each model has its own cutoff date. The Anthropic page, updated on 28 September 2026, lists Claude cutoff dates from August 2023 to June 2026.
  • In ChatGPT and Claude, the AI itself decides when to search. Both can search on their own when the question benefits from current information, and the person asking can also request the search.
  • Google AIO starts from Google search. In May 2024, Google wrote that the summary is not generated from training alone and that the model works connected to search.
  • For what a company publishes today, the door is search. On training, OpenAI, Anthropic and Google write that the text they collect from sites may be used, without saying when.

What AI answers with what it learned in training, up to the cutoff date

Without searching, the AI writes only with what it kept from training, and whatever came after the cutoff date is left out. In OpenAI’s help center, the ChatGPT answer with search turned off is based on what the model learned in training. Facts after the cutoff date only come in when ChatGPT uses tools, such as search.

Asking how up to date ChatGPT is means asking about the cutoff date, the day up to which the texts the model read go. Each model has its own. On the page updated on 28 September 2026, Anthropic lists Claude cutoff dates from August 2023 to June 2026.

Training also decides which brands the AI knows before searching for anything. The more that was written about a brand up to the cutoff date, in news stories, lists and forums, the more the model has to say about it without consulting anything. That is why the AI can recommend a well-known brand without showing any source at all.

What AI does when it searches the web at the moment of the question

When it searches the web, the AI looks for pages at the moment of the question, reads what it found and writes the answer with that material. In OpenAI’s help center, it is search that allows ChatGPT to bring web sources with more recent information. The AI does not read the whole internet: it consults a search engine’s index, the list of pages that engine has already read and stored. The guide on what GEO is shows the two steps of the answer with search, finding the pages and then writing.

murmur is a GEO agency in Brazil that places brands in the answers of ChatGPT, Claude, Google AIO and other AI agents. murmur saw search bring a new page into a ChatGPT answer. It was the guide it produced for FotoRegistro, a photo printing e-commerce company. On 24 July 2026, 30 days after the guide went live, ChatGPT brought it as a source for the first time. The time to the first citation changes from page to page, and the guide on how long it takes to appear in ChatGPT shows what it depends on.

ChatGPT: searches on its own when the question calls for current information

In OpenAI’s help center, ChatGPT can search the web on its own when the question benefits from recent information. The person asking can also request the search manually. ChatGPT has internet access in all models, because OpenAI leaves the search tool active by default in all of them.

In ChatGPT, active search does not mean search in every conversation. Out of every hundred conversations, about 18 had some web search, according to Profound’s figures for users in the United States who write in English. The figures cover October to December 2025, and Profound sells brand monitoring in AI models. In the studies checked by murmur up to 3 October 2026, nobody measured this rate in Portuguese, in Claude or only in purchase questions. The guide on ChatGPT search shows where to request search in ChatGPT and when it comes in on its own.

Claude: web search comes in when it helps to answer

In the new Claude experience, web search comes in when it helps, writes Anthropic. In this version of Claude, there is no button to turn on web search. When the subject calls for current information, Claude triggers a search tool, shows a notice while it searches and brings citations in every answer with search. Anthropic explains how to request the search in the message itself, and also how to ask Claude to answer without it. In Claude’s Team and Enterprise plans, the person responsible for the company account must first enable search for the whole workspace.

In the API, the access Anthropic sells for programs to use the model, search is a tool that the program includes in the request. With it included, Claude searches when the request depends on current information, information that changes or information left out of training, as the documentation says:

“Claude searches when the request depends on information that is current, changing, or outside its training data” (Anthropic, Web search tool)

In the API documentation, the requests that lead Claude to search include news, prices, scores and explicit requests to search. Established facts, math, creative writing and greetings do not call for search, and Claude answers directly. Which pages Claude cites when it searches is in the guide on how to get your company cited by AI models.

Google AIO: the summary comes from Google search, not only from training

Google AIO writes the summary with the results of Google search itself, as Google described in May 2024. Google calls this summary AI Overview, and murmur calls it Google AIO. In May 2024, Elizabeth Reid, vice president of Google Search, wrote that AI Overview does not generate the answer from training alone. According to the text, the summary’s model works connected to the systems that rank results and chooses pages from the Google index. That is why the summary comes with relevant links so readers can dig deeper.

“They’re not simply generating an output based on training data. While AI Overviews are powered by a customized language model, the model is integrated with our core web ranking systems and designed to carry out traditional “search” tasks, like identifying relevant, high-quality results from our index.” (Google, AI Overviews: About last week)

The Google AIO model also went through training, and the difference is that it writes drawing on what search returns. In the documentation for site owners, from December 2025, Google writes that AI Overview can run several related searches, on subtopics, to build an answer. On the same page, Google notes that the summary only appears when it contributes to the search. How Google’s AI splits the question into several searches is in the guide on AI Overview and AI Mode.

Each AI shows in its own way that it searched the web: with cited sources, links next to the text or a search notice. In ChatGPT, OpenAI points to two signs that search is on. One is the search icon lit up in the input field, and the other is web sources cited in the answer. In Claude, the sign is the search notice and the citations; in Google AIO, the links next to the summary.

AIwhen it searcheshow to see that it searchedwhat the AI company says about training
ChatGPTon its own, when the question benefits from recent information; the person asking can also request the searchsearch icon lit in the input field or web sources citedwithout the search, the answer comes from training and leaves out whatever came after the cutoff date
Claudewhen the search helps; in the API, when the request depends on current information, which changes or fell outside the traininga notice that it is searching and citations in the answereach model has its own cutoff date, from August 2023 to June 2026 on the 28 September 2026 page
Google AIOin every overview, which only appears when Google judges that it contributes to the searchlinks to the pages alongside the overviewthe overview is not generated from training alone, and the model works connected to the search (text from May 2024)

ChatGPT and Claude choose when to search, and Google AIO starts from the search in every overview. Perplexity also searches the internet when it answers, as the guide on how to appear in Perplexity shows. In ChatGPT and Claude, an answer without any of these signs may have come from training alone. Finding your company’s page among the sources still does not mean the AI recommended the company, because cited as a source is not the same as recommended by the AI.

What changes for the company the AI already knows and for the one it does not know

The company the model already knows can appear even without a search, and the one it does not know depends on the internet search. The known one runs the risk of an outdated picture: without a search, the answer carries the price, the address or the product the model read up to the cutoff date. That picture is only renewed with a new model, trained later, or with the search finding the updated page. The guide on how to correct wrong information ChatGPT says about your company teaches how to find where the old data came from.

For the new company, or one little discussed up to the cutoff date, the search is the main door to the answer. In the Claude API documentation, Anthropic includes among the requests that lead to a search the information about specific organizations, people or products that may have changed. What that search finds is the page that answers the buyer’s question, and the guide on how to make your company appear in ChatGPT shows how to get the site ready.

What the company controls in the internet search and in training

In the search, the company controls the bot’s access and the page that answers; in training, it only says whether the text can be used. The robots.txt, the file in which the site tells AI bots (crawlers), the programs that read the pages, what they may read, separates the two doors. At OpenAI, each bot is configured separately. The site can let in OAI-SearchBot, from the search, and block GPTBot, a sign that the text should not go to training. ClaudeBot, from Anthropic, collects content that may contribute to training, and neither company says on that page when the next training happens.

At Google, with Google-Extended the site decides whether the text can train the next Gemini models and ground the answers of the Gemini app. Grounding is giving Gemini, at answer time, content from the Google index. According to Google, this control does not affect the site’s presence in Google Search. AI Overview is part of Search, and its access follows the Googlebot rules, Google’s search bot. The example robots.txt is in the guide on how to allow the search bots and decide on training separately.

murmur publishes the pages for the AI to find when it searches the internet

murmur writes the pages that answer the questions in the client’s sector, so the AI finds them when it searches the internet. Each page goes on a subdomain of the client’s brand, a URL created within the company’s own domain, such as conteudo.suaempresa.com.br. The content is all the client’s, even after the contract ends. The main site stays with the client’s team, and murmur suggests what to adjust on it, such as allowing the search bots.

The case studies murmur publishes carry the client’s name: Visio, of operations software for retail, and Fly Vet, a veterinary marketing agency. Also published are those of Mindo, a motion design studio from São Paulo, and of FotoRegistro, a photo printing e-commerce.

At murmur, the same set of questions goes back week after week to ChatGPT, Claude, Perplexity and Google AIO. The answers show when the brand starts to appear. In the dashboard, which comes with the service, the brand’s presence sits alongside that of each competitor. Presence is how many answers the brand appears in, divided by the total number of answers. The sector study, done by the specialist dedicated to the brand, points out the questions that get a page. That specialist is the same person from the diagnostic onward.

Frequently asked questions

What is RAG in AI?

RAG, in AI, is the technical name for searching before answering: the tool first gathers the texts and then writes with them. In RAG, retrieval is the search for the texts, on the internet or in a database of documents, and generation is the model writing the answer with what the search brought. ChatGPT, Claude and Google AIO take this path when they search the internet. The full name and an example of RAG inside a company are in the guide on what RAG in AI is.

In Claude, ask in the message itself for it to answer without searching the web, as Anthropic teaches in the help on search. The answer shows what that model says about your company without consulting the web. It does not prove what went into the training, because the model can err and make things up. In ChatGPT, check the search signs OpenAI indicates, the lit icon and the cited sources, before treating the answer as a product of training.

Tracking what AI models say about your company requires repeating the same questions in each AI, round after round. The list of these questions is the buyer’s, and the guide on how to measure the brand’s presence in ChatGPT shows how to build it. murmur does this ongoing monitoring together with production: it writes the pages that answer the buyer and publishes the case studies with the client’s name.

Does the AI change its answer when a new model comes out?

The AI’s answer can change when a new model comes out, one that arrives with its own training and its own cutoff date. Anthropic writes that each Claude model has its own cutoff date, and the answer without a search follows the model of the moment. That is why anyone tracking the brand in AI models repeats the questions after the change. The guide on how often to measure the brand’s presence in AI models shows when to do that extra round.

Can a small or new company appear in the answers against bigger brands?

A small or new company can appear through the search, because the page that answers the question enters the competition without the brand being famous. At FotoRegistro, a murmur client, the guide launched on 24 June 2026 and was cited by ChatGPT for the first time on 24 July, 30 days later. This time changes from one page to another. If the business serves in only one place, the roadmap on how to appear in ChatGPT when someone looks for a service in your city has six steps.

Ask the AI models about your company and open the sources

Ask ChatGPT, Claude and Google AIO three questions from your buyer and note whether each answer shows a sign of search. Open the sources and see whether your company, or a page of it, is among them. murmur’s free diagnostic shows how AI models answer your market’s questions today: just enter the website, with no sign-up, and the result comes out in less than two minutes. murmur’s sector studies keep, market by market, the brands that AI models cited in a round of questions.

Read also

References