How Do AI Systems Find Information?

The short answer

An AI system draws on three reservoirs when it answers a question: knowledge absorbed during training, information it retrieves live from external sources at the moment you ask, and context you supply in the conversation itself — a document you upload, a page you paste, a database it is connected to. Which reservoir it uses depends on the question, the product, and the settings, and that choice determines what your answer is actually built from.

This matters more than it might seem. The same system can answer one question from memory and search for the next one, and the two answers rest on completely different foundations: one on patterns learned months ago, the other on pages fetched seconds ago. Understanding which reservoir is in play is the first step to understanding why an AI answer says what it says.

Reservoir one: knowledge inside the model

During training, a large language model processes enormous amounts of text and adjusts its internal settings — its weights — so that it gets better at predicting language. Facts, relationships, and patterns from that training data end up encoded in the weights. When you ask a question, the model can answer from this internalized knowledge without consulting anything outside itself.

Two properties of this reservoir are documented and important. First, it is frozen at a point in time: training ends, and the model’s built-in knowledge has a cutoff. A model answering purely from training cannot know about your business’s new location, last month’s price change, or yesterday’s news. Second, the knowledge is not stored as a lookup table of verified facts. It is a pattern of statistical associations, which is why models can state things fluently that are outdated, conflated, or simply wrong — the mechanism that produces an answer is prediction, not verification.

For stable, widely documented subjects — the capital of France, the plot of a famous novel, how photosynthesis works — model knowledge is usually adequate, because those facts appeared consistently across the training data. For specific, recent, or obscure subjects — including almost everything about a particular small business — it is the weakest of the three reservoirs.

Reservoir two: live retrieval at answer time

Many AI products do not rely only on the model. When a question seems to call for current or specific information, the system can search an external index — typically a web index — retrieve the most relevant pages or passages, and hand them to the model as source material for its answer. This is the mechanism behind answers that include citations and links.

Platforms document this behavior in their own materials, though each implements it differently. Google describes its AI features as grounded in its search index. OpenAI documents crawlers that feed ChatGPT’s search functionality and a search feature that retrieves pages when answering. Perplexity describes itself as an answer engine built around retrieved sources. Microsoft documents Copilot’s use of web search. Anthropic documents a web search tool for Claude. The shared pattern is real and documented; the details — how often search triggers, how sources are ranked, how passages are chosen — are not fully published by any of them.

What triggers retrieval? Providers do not publish complete rules, but the observable pattern is consistent: questions about current events, specific businesses, prices, schedules, and anything the model is likely to lack or get wrong are more likely to be answered with retrieval, while general-knowledge questions are often answered directly. Treat that as a description of observed behavior, not a documented specification — the same question can be handled differently on different days, in different products, or even in different runs.

This reservoir is where a business’s published information matters most. When the system retrieves, it retrieves what exists: the business’s own website, its listings, its reviews, the pages that describe it. If that material is consistent, readable, and complete, the answer built on it can be accurate. If it is stale, contradictory, or locked in formats machines cannot read, the answer inherits those flaws.

Reservoir three: context you provide

The third reservoir is whatever is placed in front of the model in the current session: a document you upload, text you paste, a file from a connected drive, records from a business system the tool is integrated with. Enterprise products document versions of this — an assistant connected to a company’s internal documents answers from those documents rather than from the open web.

Provided context is the most controllable reservoir and the narrowest. It can make an AI tool accurate about one company’s procedures or one project’s files, and it knows nothing outside what it is given. For business owners, the practical significance is mostly on the other side of the counter: when a business uses AI tools itself, what it feeds those tools governs what they produce. When customers ask public AI systems about that business, reservoirs one and two are doing the work.

Why the same system answers two questions differently

Put the three together and the inconsistency people notice — “it knew this but had to search for that” — stops being mysterious. The system routes each question, by design choices its maker does not fully disclose, to some blend of memory, retrieval, and provided context. A question about a famous historical event comes from training. A question about a restaurant’s hours today demands retrieval. A question about a document you just uploaded is answered from your context.

The blend matters because each reservoir fails differently. Memory fails by being stale or confabulated. Retrieval fails by finding the wrong sources, or the right claim in an outdated copy. Provided context fails by omission — the crucial page was never included. Diagnosing a wrong answer starts with asking which reservoir produced it, because the fix is different in each case: you cannot edit a model’s memory, but you can correct the sources retrieval will find.

A concrete example

Ask an AI assistant two questions about the same bakery. “What is sourdough starter?” is answered from model knowledge — no search needed, and the answer is stable. “Does Harbor Lane Bakery sell sourdough on Sundays, and when do they open?” cannot be answered reliably from training; a specific small business’s hours may barely exist in the training data at all. A product with search will typically retrieve: the bakery’s website, its map listing, perhaps a review mentioning Sunday hours. If the website says the bakery is closed Sundays, an old listing still says it opens at eight, and a review from two years ago praises the Sunday sourdough run, the system must build one answer out of conflicting retrieved material — and which version wins is not something the platform publishes.

The bakery cannot reach into the model’s training. It can fix the listing, correct the website, and make the current facts the easiest version for retrieval to find and agree on. That is the entire practical lever, and it operates on reservoir two.

Misconceptions

“ChatGPT Googles it.” Sometimes an AI product searches; sometimes it doesn’t, and when it searches it uses its own provider arrangements and indexes, not necessarily the search engine you use — and the same question may be answered from memory an hour later.

“If it’s on the internet, the AI knows it.” Presence is not retrieval, and retrieval is not training. A page can exist for years without being in a model’s training data, and being indexed does not guarantee a given system will select that page for a given question.

“The AI checked its sources.” When retrieval happens, the model is working from fetched passages; generation still predicts text around them. Citations show where material came from — they are not a verification step the system performed on your behalf.

“Every answer works the same way.” The routing between memory, search, and context varies by platform, product tier, question, and time. Any claim of the form “AI always…” should be read as marketing, not mechanism.

The bottom line

An AI system’s answer is only ever as good as the reservoir it was drawn from. For information about a business, the reservoir that can actually be influenced is the retrievable public record — the pages, listings, and sources a system can find at answer time. That record is the subject of the rest of this pillar.