Web search and citations: reading past the summary

Web search fixes the most obvious limitation: the model can look things up instead of answering from training data of unknown vintage. Answers come back with citations and source links.

Which introduces a subtler problem in place of the obvious one. A citation is evidence that a page was retrieved and read. It is not evidence that the page supports the specific sentence it has been attached to.

What the link actually certifies

The model searches, reads results, and writes an answer with links back to what it drew on. The summary is a synthesis, and synthesis is where a claim can drift from its source: two pages get merged into one sentence, a qualification gets dropped, a figure from one context is applied to another. The link still points at a real page that really was read.

Anthropic's own guidance says as much without hedging: "While Claude strives for accuracy: Cross-reference cited sources to understand the full picture. Use authoritative sources for critical decisions." The vendor is telling you to check. It is worth taking them at their word.

The same article also notes that "Occasionally, website links may not function" — so a broken link is a known behaviour rather than a sign that something was fabricated.

Where drift concentrates

Not evenly. Some shapes of claim go wrong far more often:

• Numbers with units or dates attached. A figure that was quarterly becomes annual, or a 2024 value gets reported as current.
• Attributed positions. "X's position is Y" is the construction most likely to compress a nuanced document into a slogan.
• Anything aggregated across sources. A sentence citing three links is three chances for one of them not to say it.
• Superlatives and firsts. "The largest", "the first" rarely survive contact with the cited page unqualified.

A single-source factual claim with no arithmetic in it is comparatively safe. Spend your checking where the shape is risky.

Turning it on

Web search is available across current models. On Team and Enterprise accounts, an Owner or Primary Owner enables it in Organization settings before members can use it, so if it appears missing for you and present for a colleague, that is the likely reason.

Try this

Take the last answer you accepted that carried citations. Pick the single claim you would be most embarrassed to repeat wrongly. Open its source and find the sentence that supports it. If you cannot find one within a minute, you have a candidate for the failure this page is about.

What goes wrong

Reading the citation count as a confidence score. More links means more retrieval, and it carries no information about whether any of them support the sentence. A heavily cited paragraph can be the least reliable one.

Checking that the link resolves. A page loading proves the URL is valid. The question is whether the content backs the claim, and that needs the page read, not merely opened.

Treating a search result as a primary source. Search surfaces what ranks, which mixes primary documents with commentary about them. For anything that matters, follow the chain to the document itself — this site's own source list exists for that reason.

Assuming recency. Web search means the model can retrieve current pages. It does not mean the page it retrieved is current, and an undated blog post ranks as happily as a regulation.

How to check it worked

Pick one claim and verify it at the source before acting on the answer. If it holds, you have calibrated how much this particular kind of question can be trusted. If it does not, you have found the failure mode described here in your own work, which is considerably more persuasive than reading about it. Where you cannot verify a claim at all, treat it the way you would treat anything the model invented.