Research and search tools

Three products that look similar and are not

Search-grounded assistants. Perplexity and the search modes inside the major assistants. They run a search, read a few results, and answer with citations. Fast, good for current information, bounded by what the search returned.

Deep research modes. The assistant runs many searches over several minutes, reads dozens of pages, and produces a long structured report with sources. Genuinely useful for a survey of an unfamiliar area. Expensive in tokens and in waiting.

Grounded-on-your-own-documents tools. NotebookLM and similar. You supply the sources; the tool answers only from them, with citations back to the passage. This is a different and often better proposition, because the failure mode changes: instead of finding the wrong source, it tells you your sources do not say.

The failures, and how to catch them

The citation that does not support the claim. The most common failure and the easiest to miss. The link is real, the page exists, and it does not say what the answer says. Click two citations on anything you will rely on. Not all of them: two is enough to tell you whether this answer is the careful kind.

Search-engine junk as a source. The tool reads what the search engine ranked, and for commercial questions that is often marketing copy or an article written to rank rather than to inform. An answer built from three such pages is fluent and worthless.

Recency confusion. The model’s training gives it an out-of-date belief; the search gives it a current page; the answer sometimes blends them. Watch for answers that mix a current fact with an obsolete one, particularly on prices, versions and job titles.

The one-source answer. Some answers rest entirely on a single page dressed up with several citations to the same domain. Check the source list for variety, not just for length.

Missing what is not on the web. Anything behind a paywall, in a book, in a database, or in a language the search did not query, is invisible. A survey of the literature built only from what is freely readable is a survey of what is freely readable.

What deep research is for

It is at its best when you know nothing about an area and need a map: who the main players are, what the standard positions are, what vocabulary to use, what to read next. As an orientation exercise it can save a day.

It is at its worst as a final answer on anything contested, because the length and structure create an impression of thoroughness that the underlying source selection may not deserve. Read the source list before the report. If the sources are weak, the report is weak however well it reads.

The grounded-on-your-own-documents option

For anyone with a body of material (a syllabus, a regulation, a set of contracts, a year of meeting notes) this is the most reliable configuration available to a non-programmer. The answers come from documents you chose, the citations point to a passage you can check in one click, and questions outside the material get “the sources do not cover this” rather than an invention.

It is not perfect. Retrieval can miss a passage that is relevant but worded differently, and an answer that says the sources are silent is sometimes wrong about that. But the direction of the failure is far safer than the alternative.

Choosing between them for a given question

The rule is short. If the answer must be current (a price, a rule, a fixture, who holds an office) use a search-grounded tool and click the citation. If the answer must come from a specific body of material you hold, use the grounded-on-your-own-documents route. If you need to understand an unfamiliar field, use deep research and treat its output as a reading list.

And if the answer is stable general knowledge, a plain assistant with no search is often better, because searching introduces the possibility of a bad page where the model’s own knowledge was adequate. More retrieval is not uniformly safer.

The habit

For anything you will act on: read the sources, not the summary. The summary is a convenience for finding which sources to read, and treating it as the finished product is where people get caught. A researcher who uses these tools well spends the saved time on reading the three papers that matter rather than on the forty they no longer had to skim.