Reading time: 14 minutes
ChatGPT works by predicting the next word in a text, using patterns it learned in training. When a question needs current information, it also searches the web and cites sources. In addition, memory, plan and location change the answer. That is why this guide explains each piece and what it changes for brands.
In short
- ChatGPT works word by word: it builds each answer from patterns learned in training. Since October 2026, it has used GPT-6 Sol and GPT-6 Luna. In the API, these models have knowledge cutoffs in April and May 2026.
- When it needs current data, it searches the web. According to Semrush, ChatGPT turned on search for 34.5% of queries in the U.S. in February 2026.
- Word sampling, plan, memory and location change the answer. So, in a SparkToro study, the chance of ChatGPT repeating the same list of brands was below 1 in 100.
- According to OpenAI, AI makes things up because training and evaluations reward guessing. In a test by Hashtag Treinamentos, for example, ChatGPT got the names of the company’s partners wrong.
How does ChatGPT work inside? What is an LLM?
Inside, ChatGPT works with an LLM, or large language model. An LLM learns from huge amounts of text how to predict the next word. The app is the interface, and GPT models are the LLMs behind it. So each sentence of the answer comes out word by word.
According to OpenAI, the model breaks text into tokens, which can be whole words, parts of words or punctuation. During ChatGPT training, the model gets incomplete sentences such as “Instead of turning left, she turned ___.” Early on, its answers are mostly random. But after millions of sentences, it learns to predict the most likely next word.
Training leaves behind numbers known as weights or parameters. OpenAI says the model does not store copies of what it read. ChatGPT also does not “copy and paste” from its training data. Instead, when it answers, it uses the weights to create new text.
GPT stands for Generative Pre-trained Transformer: a pre-trained model that generates text with a neural network design called the transformer. For a brand, this has a direct effect. The model stores patterns, such as which companies appear with each topic. It does not keep an updated file on each one.
How does ChatGPT learn?
ChatGPT training happens before it ever talks to you. First, in pre-training, the model reads a huge amount of data and learns to predict words. Then post-training adjusts the model, and evaluation continues after launch. According to OpenAI, data used in these stages helps improve performance, reliability and safety.
For ChatGPT training, OpenAI names three sources of data. They are public internet content, data from partners, and information from users, human trainers and researchers. For public content, the company says it uses only what is freely accessible. It also says it does not collect pages behind paywalls or from the dark web on purpose. In addition, filters remove hate speech, adult content, spam and sites that gather personal data. The company also uses more and more synthetic data made with its models.
Site owners can block GPTBot in robots.txt to signal that their content should stay out of training. However, this training block does not remove the site from ChatGPT search, which uses another crawler, OAI-SearchBot. In personal accounts, OpenAI may also use chats to improve models when Improve the model for everyone is on.
What is the ChatGPT knowledge cutoff?
It depends on the model in your plan. According to the release notes, Plus, Pro, Business and Enterprise have used GPT-6 Sol since October 7, 2026. Free and Go, on the other hand, use GPT-6 Luna. In the API documentation, Sol has an April 20, 2026 knowledge cutoff, and Luna’s is May 18. All training has a cutoff date, and the ChatGPT model did not learn what came after it. So a new price or brand that came after the cutoff is not in what the model learned.
Does ChatGPT answer from training or search the internet?
Both. Without search, ChatGPT works from what it learned in training, up to the cutoff date. But when a question needs current information, it can search the web on its own. Users can also ask for a search. So an answer with search can include citations with links to sources.
How does ChatGPT search work?
According to OpenAI’s Help Center, when a question “would benefit from current information,” ChatGPT may search the web automatically. Sometimes it rewrites the question into one or more queries and sends them to partner search providers. Then it reviews the results and writes the answer. In this mode, ChatGPT works like a reader that summarizes what it found.
But search does not always happen. Semrush, which sells SEO tools, analyzed browsing data from U.S. users. In February 2026, ChatGPT turned on search for 34.5% of queries, down from 46% in late 2024. The rate also swung from 15% to 66.3% during the period. In other words, for Semrush, most ChatGPT answers still come from training alone.
For a brand, there are two doors in. In an answer without search, what counts is what the web said about the brand up to the cutoff. In an answer with search, on the other hand, what counts is the pages the search engine finds today. So a brand launched after the cutoff depends on the second door. How to tell which door each answer came from is in the guide on when ChatGPT searches the internet.
How does ChatGPT decide which sites and brands to cite?
It depends on where the answer comes from. With search, ChatGPT chooses among the pages it found. According to OpenAI, it ranks results by several relevance and reliability factors, with no guaranteed placement. Without search, on the other hand, ChatGPT tends to name the brands that training linked most strongly to the topic.
On the search route, the site must allow OAI-SearchBot. It also needs to be among the pages the search engine finds. The steps, including crawlers, Bing and timing, are in the guide on how to show up in ChatGPT.
On both routes, what the web says about the brand counts. Ahrefs analyzed 75,000 brands in ChatGPT, AI Mode and AI Overviews. Branded web mentions had a 0.66 to 0.71 correlation with presence in these AIs. In ChatGPT, however, branded search volume had a weaker correlation, at 0.352. But Ahrefs sells SEO tools and notes that correlation is not causation.
That is why what the press, comparison sites, reviews and forums say about the brand counts, not just its site. See which sources AI cites most and how to earn a place in them. According to OpenAI, for shopping questions, ChatGPT considers product metadata, such as price and description, and third-party content.
Why does the answer change every time and from person to person?
Because the model picks each word at random from likely options, and because ChatGPT works a little differently for each person. The plan sets the model. In addition, memory, custom instructions and location adjust the answer. Finally, the choice to search or not changes the sources. So two identical questions can bring different brands.
OpenAI explains that there are many possible ways to complete a sentence. That is why, according to the company, the same question can produce different answers. In the API, this sampling has a control called temperature. Higher values make the output more random, while lower values make it more focused.
How does ChatGPT personalization work?
The plan also matters, because Free and Plus use different models. With memory on, according to OpenAI, ChatGPT can use past chats, saved memories and custom instructions. Memory can also change the query it sends to the web. Meanwhile, rough location, based on the IP address, changes local results.
The effect shows up in brand lists. SparkToro ran a study with Gumshoe.ai, which sells AI tracking. In total, 600 volunteers made 2,961 runs of 12 prompts in ChatGPT, Claude and Google’s AI. As a result, the chance of ChatGPT repeating the same list of brands was below 1 in 100.
In other words, a single check does not show whether the brand appears. So test in a temporary chat, which by default does not use memory or custom instructions, and repeat the questions. The method is in the guide on how to measure brand mentions and citations. The role of temperature and the other causes is in the guide on why the same question gets different answers.
Why does AI make up information about companies?
Because ChatGPT works on probability: in training, it learned to complete text, not to check facts. According to OpenAI, training and evaluations reward guessing over saying “I don’t know.” Also, rare facts in the data, such as a birthday, follow no pattern. For example, company details like the names of business partners and prices can be facts like that.
OpenAI explained the problem in Why language models hallucinate, from September 2025. In that text, the company compares the model to a student taking a multiple-choice test. Guessing gives some chance of being right, while leaving it blank guarantees zero. Spelling and parentheses follow patterns, so those errors fade as models scale. On the other hand, patterns alone cannot predict a pet’s birthday. The types of hallucination and the rates of the main models are in the guide to AI hallucination.
“If the main scoreboards keep rewarding lucky guesses, models will keep learning to guess.”
OpenAI, Why language models hallucinate (2025)
Hashtag Treinamentos, a Brazilian school of tech courses, showed the effect in a test published in February 2025. The team asked ChatGPT who the company’s partners were. The answer changed one partner’s name, made up another’s surname and invented a third who does not exist. But then the team asked for the company’s CNPJ, its Brazilian tax ID. After that, ChatGPT searched the web and got the list right.
How can a brand reduce ChatGPT’s errors about it?
The defense is basic data in plain text, kept the same on the site, profiles and directories. After all, search helps but does not fix everything. OpenAI warns that search results and citations can be incomplete, outdated or incorrect. The types of errors and what to do about them are in the guide on why ChatGPT gets companies wrong.
What changes when an AI agent researches and buys for the user?
Who visits the site changes. Instead of just citing pages, the agent opens sites, compares options, fills in forms and can prepare the purchase. In ChatGPT, that role belongs to ChatGPT Work, launched on July 9, 2026, and not available on Free and Go. However, payments and bookings require the user’s confirmation.
ChatGPT’s old agent mode is no longer available. OpenAI also retired the Atlas browser, with a shutdown date of August 9, 2026. According to the release notes, agentic browsing moved into ChatGPT and Codex.
What does ChatGPT Work do?
With the cloud browser in Work, ChatGPT reads pages, clicks buttons and fills in forms. OpenAI’s examples include requesting quotes from businesses and comparing product availability and local stock. However, the company warns that some sites block automated agents, so the task may not move forward.
For a brand, ChatGPT works here like a visitor that decides based on what it can read. That is why the firewall and CDN should not block OpenAI’s browser, which signs requests to prove their origin. In fact, OpenAI publishes instructions to allow it in Cloudflare, Akamai and other services. Also, price, stock, delivery time and terms need to appear as text on the page.
For Brazil, OpenAI has not yet documented buying inside the chat, as the guide on GEO for ecommerce shows. What changes for stores and brands is in the guide on agentic commerce.
What does the way ChatGPT works change for brands?
A brand does not control the model, but it influences what the model finds. After all, since ChatGPT works through four routes, each one calls for a different action. This work has a name, GEO, which is the topic of the guide to what GEO is.
| Route | Where the brand comes from | What the brand can do |
|---|---|---|
| Training | Patterns learned up to the cutoff | Earn mentions in third-party sources |
| Web search | Pages found in real time | Allow the crawler and be in the index |
| Products | Data from merchants and third parties | Keep price and description accurate |
| Agent | Sites it opens and reads | Keep the site open to the agent |
- Allow OpenAI’s search crawler and agent: OAI-SearchBot in robots.txt and the Work browser in your CDN and firewall.
- Write basic data as text, such as partners, address, prices and service area, the same on the site and profiles.
- Earn mentions in third-party sources, such as the press, comparison sites and review sites.
- Measure with fixed questions, repeated in a temporary chat, and compare month to month.
With many brands or questions, however, data collection needs automation. What you can automate, and what still needs people, is in the guide on AI for SEO and GEO. If the brand still doesn’t show up, see why AI doesn’t cite your brand and how to test each cause.
What we found analyzing 8 Portuguese-language pages on how ChatGPT works
We analyzed the 8 pages in Portuguese that showed up most across 7 searches. The searches included “como o chatgpt funciona” (“how ChatGPT works”). The most recent page is from August 2025. Only 2 explain next-word prediction, and only 2 say that ChatGPT searches the web. In addition, three describe it as a database consulted at answer time.
None of them covers memory, agents or how ChatGPT chooses the sources and brands it cites. Likewise, none explains why AI makes up information, and only one explains why answers vary. Finally, none links to OpenAI’s documentation.
Frequently asked questions
Not directly. With memory on, the correction may apply to your future chats, but only in your account. In personal accounts, OpenAI may use chats to improve models when Improve the model for everyone is on. So, to change what other people see, fix the information on your site and in third-party sources.
In ChatGPT training, OpenAI picks the sources: public internet content, data from partners and information from users, trainers and researchers. Filters remove spam and adult content. In search, on the other hand, ChatGPT ranks results by several factors, and OpenAI says there is no guaranteed placement.
No. According to OpenAI, ChatGPT ads appear below the answer, marked as sponsored. In addition, advertisers cannot influence, rank or change the answers. Organic presence depends on content, mentions and crawler access, as the guide on how to show up in ChatGPT explains.
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, How ChatGPT and our foundation models are developed (2026).
- OpenAI, Searching the web with ChatGPT (2026).
- OpenAI, Memory in ChatGPT (2026).
- OpenAI, Why language models hallucinate (2025).
- OpenAI, ChatGPT release notes (2026).
- OpenAI, Using cloud browser in ChatGPT (2026).
- Semrush, analysis of 17 months of ChatGPT clickstream data (2026).
- Ahrefs, brand visibility factors in ChatGPT, AI Mode and AI Overviews (2025).
