VERIFYSearch & retrieval
I asked an AI 52,350 questions.
Asif@akbuilds_
0:25TakeYou give it a page and a list of questions, and it sends back a yes, a no or a score for each one, with a probability. 25 judgments about a page take about a third of a second.
I asked an AI 52,350 questions. It cost 41 cents.
The model is Jev, from TypeSafe. It doesn't write text. It judges. You give it a page and a list of questions, and it sends back a yes, a no or a score for each one, with a probability. 25 judgments about a page take about a third of a second.
At that price you stop sampling and judge everything.
So I judged every page ChatGPT, Claude, Gemini and Perplexity cited for 574 real buyer questions. That's 1,920 pages, 25 questions each, in 4 minutes.
It turns out every AI has a type.
ChatGPT goes to the source. 91% of what it cites is a page about one product.
Claude wants a straight answer.
Gemini loves this year's ranked list. 54% of its citations, against 15% for the others.
Perplexity reads the comparisons.
I wouldn't have run this on a normal LLM. It was too slow and too expensive to even try.
Then I put Jev on a live page. Type your site and watch it judge every page against the questions your buyers ask AI. 2,000 judgments in 6 seconds, my cost is for about a cent, and nothing on the screen is simulated.