The Search Inside ChatGPT: How GPT-6 Changes When and Where ChatGPT Looks Things Up
For marketing leaders who want to understand how AI search decides what it reads before it recommends anyone.
Picture a marketing leader who types a simple question into ChatGPT: who is the best provider in my category for a company our size? The answer names three companies, explains why each one fits, and cites a handful of pages. Her company is not among them. She checks Google, and her company sits at the top of the first page for the same question. Both results are real, and both were produced by something we call search, but they were not the same search.
That gap is no longer a corner case. In an August 2025 survey of nearly 2,000 U.S. consumers, McKinsey found that half of consumers already use AI-powered search. The same research found that a brand's own websites make up only 5 to 10 percent of the sources AI search references when it answers questions about that brand.
With GPT-6, which is rolling out this week, OpenAI has changed two things about how ChatGPT finds information: when it decides to search, and where it searches. Together they change what it means to be found.
From answering from memory to looking things up
Every AI answer starts with a choice about whether to search at all. An AI model like the one behind ChatGPT can answer a question in two ways. The first is from memory. During training, the model reads a very large amount of text, and it can answer many questions from what it learned then. The problem is that this memory has an end date, called a training cutoff, and anything that happened after that date is missing. The second way is to look things up first. The model runs a search, reads the pages it finds, and writes its answer from those pages. The technical name for this is retrieval-augmented generation, or RAG. In plain terms, RAG means search first, then answer.
For most of ChatGPT's history, the choice between these two methods was left largely to the model's judgment, and the model often chose to answer from memory. That is why older answers sometimes sounded confident and out of date at the same time. For a brand, it also meant that what the model "knew" about you was mostly fixed at training time. If your product launched after the cutoff, or your pricing changed, or a competitor overtook you, the model might never notice.
Each new model has moved further toward searching first. GPT-6 is the clearest step yet.
What GPT-6 changes about when ChatGPT searches
The new instructions give the model clearer rules for when to search, and the threshold is lower. After comparing GPT-6's search instructions with those of the models before it, the most important change is that the model now gets much more specific guidance on when it should trigger a search. [RYAN: confirm this description matches your comparison.] The practical effect is that ChatGPT searches far more often than it did.
The key rule concerns time. If a prompt is judged to involve information that may be time-sensitive, GPT-6 will always run a search before it answers. That sounds narrow until you look at what counts as time-sensitive. It ranges from sports, where scores and standings change every day, to how-to questions, where the steps change whenever software, products, or rules change, to "who is the best?" questions, where the answer depends on the current market.
That third group deserves attention, because it is where buying decisions happen. A question such as "what is the best payroll software for a 50-person company" or "who is the best estate attorney in Pasadena" is, by GPT-6's standard, a question whose answer can change over time. So the model will most likely look it up instead of relying on what it remembers. The recommendation the buyer reads is assembled from what the model finds on that day.
The implication is simple to state and harder to act on. For the questions that matter most to a business, ChatGPT's memory of your brand now matters less, and what it can find about you right now matters more.
Search no longer just means Google and Bing
With each model, OpenAI relies more on its own index of the web.
When people hear that ChatGPT "searched," most assume it ran the question through Google or Bing and read the top results. For a long time that was close enough to true, and it led to a reasonable shortcut: if you were visible in traditional search, you were probably visible in ChatGPT.
That shortcut is getting weaker. With every model, OpenAI is relying more and more on its own internal index. [RYAN: confirm this is from your model comparison.] An index, in this sense, is the stored collection of web pages that a search system has visited, read, and organized so it can find them quickly later. Google has one. Bing has one. OpenAI is building its own, and it leans on that index more with each release.
OpenAI's crawler, the program that visits websites to build this index, is called OAI-SearchBot. OpenAI's documentation says the crawler "is used to surface websites in search results in ChatGPT's search features," and it states plainly that "sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers."
That sentence is worth reading twice. A website can block OpenAI's crawler without anyone on the marketing team knowing. It can happen through an old line in the site's robots.txt file, which tells crawlers what they may visit, or through a security or bot-blocking setting at the hosting level. When that happens, the site can still rank well in Google while being left out of ChatGPT's answers entirely.
So when we say a model searched, we now have to ask which search. It may be Google. It may be Bing. Increasingly, it is OpenAI's own index, and each one has its own rules for what gets in, how often it is refreshed, and what it trusts.
What the model reads once it searches
Being in the index is only the first step; the model weighs what others say about you. Getting into the index makes a brand eligible. It does not make the brand chosen. Once ChatGPT runs a search, it reads several sources and writes one answer from them, and the McKinsey finding above suggests where most of that reading happens. If a brand's own sites are only 5 to 10 percent of the sources AI search references, the other 90 percent or more comes from somewhere else: reviews, comparison articles, forums, news coverage, industry directories, and the sites of partners and customers.
This is where the shift in search behavior meets the way models build trust. A model reading ten pages about a category will most likely give more weight to a claim that appears in several independent places than to a claim that appears only on the company's own website. In practice, the model is checking whether the outside world agrees with what you say about yourself. A brand that describes itself one way while reviewers, directories, and partners describe it another way gives the model a reason to hesitate.
None of the major AI companies publishes exactly how its model ranks or weighs sources, so this is a description of likely behavior rather than known rules. But the pattern is consistent with what OpenAI has said about its crawler and with what McKinsey measured about source mix.
The Choice Economy, applied to search
When the model searches more often, more of the buyer's decision happens inside the answer.
We use the term Choice Economy to describe a market where the customer journey no longer begins on a company's website, and increasingly not on a list of search results either. It begins with a question to an AI system, and the system's answer narrows the field before the buyer visits anyone.
GPT-6's search behavior makes that narrowing more frequent and more current. Every time a buyer asks a "best" or "how do I" question, the model now goes out, reads, compares, and returns with a short list. The scale is large. OpenAI said in February 2026 that ChatGPT had reached 900 million weekly active users. Even if only a small share of those users ask buying questions in a given week, that is a very large number of decisions being shaped by what the model happened to find.
For a marketing leader, this changes the unit of competition. The old unit was the ranking: where your page sat in a list. The new unit is the answer: whether your brand is named, how it is described, and which sources the model cites to support that description. You can rank first and still be absent from the answer. You can also be named in the answer without ranking first anywhere, if the sources the model reads describe you clearly and consistently.
What to expect over the next 6 to 18 months
The direction of each release so far points to a few likely developments.
The gap between visibility in Google and visibility in ChatGPT will most likely grow. As OpenAI relies more on its own index, the overlap between what Google shows and what ChatGPT reads will shrink, and teams will need to measure both instead of assuming one stands in for the other.
More questions will most likely be treated as time-sensitive. Almost any question about which product, service, or provider to choose can change as new options, prices, and reviews appear, so it is reasonable to expect the share of buying questions that trigger a search to keep rising.
Freshness will matter more than it has in years. When the model reads your pages at the moment of the question, a page with last year's pricing or an outdated product name is not a small problem. It may be the page the model reads.
Measurement will get harder before it gets easier. An answer built from your page does not always produce a click, so website analytics will most likely undercount how often ChatGPT uses your content. Teams will need to watch ChatGPT referral traffic, branded search volume, and direct checks of AI answers together to see the full picture.
Where to begin
The first steps are about access, freshness, and what others say about you.
Start with access, because nothing else matters if the crawler cannot reach you. Ask your web or IT team to confirm that OAI-SearchBot is allowed in your robots.txt file and is not blocked by any firewall or bot-protection service. OpenAI notes that changes to robots.txt can take about 24 hours to take effect, so this is worth doing before any other work.
Next, list the questions your buyers actually ask, especially the "who is the best" and "how do I" questions, and run them in ChatGPT. Record which brands are named, how they are described, and which sources are cited. That record is your baseline, and it will show you which outside sources the model trusts in your category.
Then update the pages that answer those questions so that prices, product names, and dates are current, and make your how-to content specific enough that a model reading it today would find it accurate. Finally, look at what those outside sources say about you, and correct the gaps between how you describe yourself and how the rest of the web describes you, because that agreement is most likely what the model is checking.
Search used to be a place your buyers went. In GPT-6, it is something the model does on their behalf, more often than before, and increasingly in an index that OpenAI controls. You no longer need to rank. You need to be the thing the machine quotes.
Appendix: Reference tables
Table AHow GPT‑6 handles different kinds of prompts
| Prompt type | Example | What it does | Why |
|---|---|---|---|
| Sports | Who won last night’s game? | Searches Every time | Scores and standings change daily. |
| How-to | How do I set up two-factor login on my bank app? | Searches Every time | Steps change when products or rules change. |
| Who is the best? | What is the best payroll software for a 50-person company? | Searches Every time | The answer depends on the current market. |
| Stable fact | What is the boiling point of water? | Inferred, not tested Most likely answers from memory | The answer does not change over time. |
Table BStatistics cited in this article
| Statistic | Source | Date | What it means |
|---|---|---|---|
| Half of consumers use AI-powered search | McKinsey AI Discovery Survey, 1,927 U.S. consumers. | Fielded Aug 2025 Published Oct 2025 |
AI search is already a mainstream way people find information. |
| A brand’s own sites make up 5 to 10% of the sources AI search references | McKinsey | Oct 2025 | Most of what a model reads about a brand comes from other sites. Our own corpus puts brand-owned content at 29% for B2B SaaS, well above this general-web figure. |
| ChatGPT reached 900 million weekly active users | OpenAI Company figure, reported by TechCrunch. | Feb 2026 | A self-reported number, and a measure of how many people may see an AI-written answer each week. |
Sources
McKinsey, "New front door to the internet: Winning in the age of AI search," October 16, 2025: https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/new-front-door-to-the-internet-winning-in-the-age-of-ai-search
OpenAI, "Overview of OpenAI Crawlers": https://developers.openai.com/api/docs/bots
TechCrunch, "ChatGPT reaches 900M weekly active users," February 27, 2026: https://techcrunch.com/2026/02/27/chatgpt-reaches-900m-weekly-active-users