When an AI answers by citing recent sources, many people imagine it "browses the internet" the way you might consult a library. That's not what happens. It formulates queries to a search engine, receives a few page excerpts, and writes based on those excerpts. Every link in this chain can fail, and understanding which ones explains most of the errors you'll encounter.
This capability has become central, and how it works remains largely opaque to users. Here's the chain, step by step.
Step 1: deciding to search
The model first has to determine whether your question requires a search. A question about a definition doesn't; a question about a recent event does.
This is a decision, and it can go wrong in both directions. A model that doesn't search when it should will answer from outdated knowledge, with the usual confidence. A model that searches systematically is slower and costs more.
Step 2: formulating the query
The model translates your question into search queries. This translation isn't neutral: the terms chosen determine what gets found.
If your question touches on a topic where vocabulary matters — a technical term, an ambiguous proper noun, a contested phrasing — a poorly calibrated query will bring back off-topic results, and the model will diligently write from those.
Step 3: receiving excerpts
Here's the most misunderstood link. The model generally doesn't receive full pages: it receives excerpts, a few sentences per result.
These excerpts are selected by the search engine according to its own criteria. The model therefore inherits everything that structures that engine: its ranking, its optimisation, its commercial biases.
Important consequence: on topics heavily optimised for SEO — product recommendations, health, personal finance — the quality of available excerpts is often poor, and the answer suffers accordingly.
The model then produces an answer from these excerpts, attaching links to it. The trap is here: a displayed link doesn't guarantee the sentence comes from that page. The model read several excerpts, synthesised, then attributed. This attribution is itself a generation, and therefore fallible. A correct claim can end up attached to the wrong source, or a nuance present in the excerpt can disappear in the synthesis.
What this means in practice
Open the links that matter. Not all of them, but those carrying a figure, a date, a quote, or a claim you're going to rely on. That's the most cost-effective verification, and it takes thirty seconds.
Be wary of over-optimised topics. If the question is about a product to buy, a treatment, or an investment, the available raw material is of uneven quality.
Rephrase if sources seem weak. You can explicitly ask to search a specific source, or to favour primary documents. This genuinely changes the result.
The date matters. A search returns what's well-ranked, not what's most recent. On a fast-moving topic, specify the time period.
Why it's still a considerable improvement
Despite these limits, this capability fixes the most annoying flaw in these systems: the knowledge cutoff date. A model that can search is no longer frozen at the moment of its training.
And it makes answers verifiable, which is a major improvement over a claim pulled from nowhere. A link you can open is infinitely better than an assurance without a source, even if the link is sometimes misattributed.
What to remember
An AI that searches the web isn't an AI that knows. It's an AI that quickly reads a few excerpts selected by a search engine and gives you a synthesis of them.
That's extremely useful, and it doesn't replace reading a source when the stakes are real. The right way to use it is to treat it as a fast document assistant: it gets you to the right pages much faster than you'd have found them, and it's up to you to read them.