Ask ChatGPT which payment gateway to use and you get a confident, well reasoned answer. It sounds like the model knows the market. It doesn't, at least not from training. The answer comes from a pipeline: the model searches the web, picks top results, reads passages, reasons over them, then writes. The language model is one stage in that chain, and not the one that decides whether your brand is in the running.
The Short Version The language model is a frozen snapshot. Its knowledge stops months before you ask. The apps hand the model tools. For product questions it makes at least two tool calls: a web search, then a fetch of pages worth reading. Your prompt gets rewritten into several search queries. Ordinary search ranking decides which pages reach the model at all. The model reads passages, not whole pages, then weighs them against what it already believed. Every step is a filter. To come out the other end, a brand has to be indexed, fetchable, quotable and backed up by other people.
A large language model is trained in two phases. Pretraining feeds it a huge slice of the public web so it learns patterns. Fine-tuning then teaches it to be helpful and follow instructions. Both phases finish on a fixed date, and after that date the weights do not change. Nothing new gets in.
That date is the knowledge cutoff, and it usually sits further back than people assume. GPT-5 shipped in August 2025 with a September 2024 cutoff. Gemini 2.5 Pro has a January 2025 cutoff. Newer models narrow the gap, but there is always a gap. A model released this month still knows nothing about the pricing page you shipped last week.
There is a second, quieter problem. Inside the cutoff, what the model holds about your brand is a statistical impression rather than a record. It has seen your name near some words more often than near others. It cannot reliably tell your 2024 feature set from your 2025 one. Ask a model with no tools for the best CRM for a 10 person agency, and you get a confident answer built from patterns that are a year old. It will not flag that for you.
The model's memory is out of date. GPT-5 was 11 months behind on launch day and 2 years behind today.
So the product teams did the obvious thing. They stopped asking the model to know things and started asking it to look things up.
The mechanism is called tool calling, and it is simpler than the name suggests. The model cannot open a web page. All it can do is produce text. So the app gives it a short menu of actions written in plain language: search the web, open this URL, find this phrase on the page. When the model wants one of those, it writes a small structured request instead of an answer. The app catches that request, runs the real search or fetch, and pastes the result back into the conversation. Think of it as a colleague handing the model a printout.
The model reads the printout, thinks about what it says, and writes the answer. If a page looks worth opening, it asks for that too before writing.
This is the whole idea. A model that cannot see the internet becomes a model that sends errands and reads what comes back. OpenAI, Google, Perplexity and Anthropic all run some version of this. Google's developer docs for Gemini lay it out in five stages: the model looks at the prompt, decides whether a search would help, writes one or more queries, runs them, then works the results into a response. ChatGPT's browsing tool does the same job with different plumbing.
The rest of this post goes through those steps one at a time, because each step throws pages away. The pages that survive are the ones that get your brand named.
Not every prompt triggers a search. Explain compound interest usually does not. Best project management tool for a remote team of 12 almost always does, because the model has learned that product questions go stale. In ChatGPT a small decision layer makes this call before any search runs. In Gemini the model makes it as part of its reasoning.
For anyone trying to get a brand named, the practical point is short. Anything that names a category, a price, a comparison or the current year goes to the web. Your brand's fate on that prompt depends on what the web says today, not on what the model absorbed in training. Which is a relief, because you can change the web.
The model does not paste your prompt into a search box. It rewrites it. One question becomes a set of narrower searches, each aimed at one facet of what you asked. Google calls this query fan-out and describes it as issuing multiple related searches concurrently across subtopics and multiple data sources. ChatGPT does the same thing, it just has not named it.
Take which payment gateway should a UK subscription business use. A plausible fan-out looks like this: payment gateways for subscription billing UK Stripe vs GoCardless vs Adyen recurring payments payment gateway fees comparison 2026 SCA and 3D Secure support subscription payments best payment gateway reviews small business UK
Notice that your page does not need to rank for the original question. It needs to rank for at least one of the sub-queries, and the narrow ones are the easier target because fewer strong pages compete for them.
ChatGPT's shopping answers go one step further. Profound looked at about 200,000 shopping prompts and found a second fan-out layer that turns the prompt into product-specific sub-queries before any candidate is picked. Same pattern: many narrow searches rather than one broad one.
Each of those queries runs against a real search index. For ChatGPT that is a mix of providers with Bing as the main one, plus whatever OpenAI's own crawler, OAI-SearchBot, has indexed. For Gemini, AI Overviews and AI Mode it is Google's index, its Knowledge Graph and, for products, its Shopping Graph. Perplexity runs its own index and ranker.