Why it is not a search engine, and what changes when you bolt one on

Two entirely different machines

A search engine keeps an index: a vast list of documents, and for each, where its words appear. You send a query, it finds matching documents, it returns them. Whatever comes back existed before you asked. It can be wrong about relevance. It cannot invent a page.

A language model keeps weights. You send text, it generates text. Nothing it returns existed before you asked. It cannot look anything up, because there is nothing to look in.

That is the whole difference, and nearly every disappointment people have with chatbots traces back to expecting the first while using the second.

Why citations from memory are unsafe

A reference has a recognisable shape: authors, year, title, journal, volume, pages, a DOI beginning 10.. The model has read hundreds of thousands of them and has learned that shape thoroughly.

Generating a well-formed citation therefore requires no knowledge of any actual paper. Plausible authors for the topic, a plausible year, a plausible journal, plausible page numbers. The output is indistinguishable in form from a real one, and it may be entirely fictional.

This is not a rare edge case. It reached court: in 2023 two New York lawyers were sanctioned for filing a brief containing six fabricated case citations produced by a chatbot, complete with quotations from the non-existent rulings. Comparable incidents have followed in several countries since, including from people who knew about the first one.

The rule is unconditional. A citation from a model with no search is a hypothesis about a document, not a document. Check that it exists before repeating it.

What “with web access” actually does

When a product searches, something specific happens, and it is worth knowing the steps:

1. Your question is turned into one or more search queries.
2. A real search engine runs them and returns real pages.
3. Some of that page text is pasted into the context window.
4. The model answers from the text now sitting in front of it, and links to the sources.

This is retrieval-augmented generation, RAG, and it is the same recipe whether the corpus is the whole web, your company’s documents, or a folder of your own notes. The retrieval step is often the embedding search from an earlier lesson, sometimes ordinary keyword search, usually both.

What it fixes is real. The answer is now grounded in text that exists, the cut-off is bypassed, and you get links you can check.

What it does not fix

Three things, and each has caused public failures.

Retrieval can bring back the wrong thing. If the search returns an outdated page, a satirical post, a forum joke or a competitor’s marketing, the model answers faithfully from it. Grounding guarantees a source, not a good one. The widely reported episode in 2024 of an AI search feature recommending glue on pizza came from faithfully summarising a joke comment.

The model can still misread. It may overstate a hedged finding, merge two sources into a claim neither made, or attribute a sentence to the wrong document. A link next to a sentence means the system used that page. It does not mean the page says that.

The link may not support the claim. This is the one to check, and it takes ten seconds. Open the source, search it for the specific figure or claim. Studies of AI search summaries have repeatedly found meaningful rates of citations that do not support the surrounding sentence.

The practical test

When an answer matters, ask two questions.

Did it actually search? Look for links, or for a visible indication that a search ran. If neither is present, the answer came from weights and any citation in it is unverified. Some products search only sometimes, deciding per question, and do not always make that obvious.

Does the source say it? Open one link (the one carrying the claim you would act on) and find the sentence. If the claim is not there, nothing else in the answer has earned your trust either.

That second habit is worth more than any amount of prompt technique. It converts a fluent paragraph into either a checked fact or a discarded one, in the time it takes to make tea.