Claude vs Gemini: I opened every source they cited
The more quotable the number, the sooner I open its source.
I gave Claude and Gemini the same research question, one I needed answered for real that how do AI answer engines decide which websites to quote?
Then I opened every source they cited. All 97 of them.
Claude came back with 50. 37 checked out completely, 13 had small slips like a wrong date, and none were made up.
Gemini came back with 47. 15 checked out completely. 12 didn't exist at all.
Here's the part I didn't expect. The fakes weren't filler. They carried the report's biggest numbers.
"4.8x more pipeline", credited to a study that doesn't exist. "28% fewer hallucinations", from research nobody can find. A "2.4x" pinned on a paper that turned out to be about classifying buildings.
Those were the exact numbers its recommendations rested on.
A third tool, ChatGPT, never got going. It sent back a template for an "unspecified topic", so it gets a rerun.
My habit now is simple. The more quotable the number, the sooner I open its source.
