Tested, not guessedTest plan

Research with sources: accurate answers, real citations?

Task: Answer research questions correctly and back every claim with a source that exists and says what the tool claims.

AI research tools cite sources, but links can be dead, irrelevant or not say what's claimed. We check answers against a key and open every link.

This is a test plan. Nothing has been run yet, so there are no scores or verdicts below.

How we tested

A set of research questions with answers a person has verified from authoritative sources.
Every tool gets the same questions, word for word, with web access on where the tool has it.
Answer accuracy is checked automatically against the answer key.
Citation validity: every link is opened by our runner. A citation counts only if it loads and, where we set a list, comes from an authoritative domain.
A person checks whether each source actually supports the claim it's attached to.
Cost and time are recorded.

Answer accuracy · objective · weight 3: Share of answer-key facts present, minus 25% per known wrong claim.
Citations that load · objective · weight 2: Share of cited links that open (and come from an allowed domain, where set).
Sources support the claims · human-judged · weight 2: Does each source actually say what the tool claims?

3–5 research questions real people ask (for example health, money or consumer questions) whose answers can be verified.
An answer key per question, and the list of authoritative domains for each topic.
Which tools and plans to test.