When it can search, and what that fixes

What grounding actually is

A model with a search tool does something specific: it issues a query, receives some results (usually snippets, sometimes full pages) puts them in the context, and answers from them. The answer is then produced from retrieved text rather than from training data alone.

This fixes several things. Recent events. Current prices and figures. The existence of a source. Anything that happened after the training cut-off.

It fixes them by turning a memory problem into a reading problem, which is the right trade. It also creates new failure modes, and those are less discussed.

What grounding does not fix

The page can be wrong. Retrieval finds text; it does not evaluate it. A confidently written blog post, an out-of-date government page, a competitor’s marketing claim, a forum answer from 2019: all are text, all can be retrieved, all can be reported with the same steady tone.

The snippet may not be the page. Search results often supply a fragment. A claim built on a snippet can miss the sentence beginning “until 2024, this was the case”.

The query may be wrong. The model chooses what to search for. A bad query returns plausible results for a slightly different question, and the answer will be about that question without saying so.

The web itself is increasingly machine-written. A generated article citing another generated article looks exactly like corroboration. Two sources agreeing is much weaker evidence than it used to be, and this is getting worse rather than better.

Recency confusion. Pages are undated, or dated by their last edit. “Current rate” from a page updated in 2021 is reported as current.

How to prompt a grounded search well

Say the date, and say you want current information. “Today is 6 September 2026. I need the position as of today, not as of your training data. Say the date of every source.”

Ask for the sources with the answer. Not at the end as a list: beside each claim.

Name the kind of source you will accept. “Only official government or regulator pages for the rules; news outlets only for what happened.” This does more than any instruction about accuracy.

Ask for the query it used. One line, and it exposes the commonest silent failure: it searched for something adjacent to your question.

Ask what it could not find. “List anything I asked about that you could not verify from a source.” That converts a smoothly complete answer into an answer with visible gaps, which is more useful and more honest.

Reading a grounded answer

Three habits, each a few seconds:

1. Open one link. Not all of them. One, chosen as the one carrying the claim that matters most.
2. Check the date on the page, not in the answer.
3. Ask whether the page is a source or a repeat. A news article quoting a regulator is not the regulator. For anything consequential, get to the primary page: the regulator, the company’s own announcement, the statute, the published paper.

Deep research modes

Several products now offer a mode that searches many sources over several minutes and returns a long report with citations. These are useful for orientation: the shape of a field, who the players are, what the main positions are.

They do not remove the checking. A twenty-page report with sixty citations is sixty citations nobody has opened, and the length makes checking feel less necessary at exactly the moment it has become more work. The honest use is to treat the report as a map and the citations as places to go, then read the four sources that carry your actual decision.

The one-line summary

Search converts “what does the model remember” into “what did it find and read”. That is a large improvement and it moves your job rather than removing it: from checking whether a source exists to checking whether the source says it and whether the source is any good.

Knowing whether it searched at all

A practical detail that catches people. Products differ in when search fires: some search on every message, some only when the question looks time-sensitive, some when you toggle it, and some route to a model without the tool at all.

So an answer about a current figure may be a retrieved fact or a remembered one, and both arrive in the same voice. If there are no links, assume it did not search. If it matters, say so directly (“search the web for this and cite what you find”) rather than trusting that a question about current prices triggered a lookup.