AI Answer Repeatability Report

$249

one-time

Checkout for this service opens by emailed invoice link. Tell us where to send it and we reply the same working day.

Free specimen and paid-delivery boundary

We ran this method on ourselves and published the result, including the part that is bad for us. That run is the sample: the same panel shape, the same outcome buckets, and the same refusal to score an unreadable answer as a miss.

Paid delivery contract

You send:

• One client domain and exactly five prompts
• Three fresh-session manual runs per prompt across one to three fixed assistants
• Your written authorization to record the transcripts

You get back:

• PDF and CSV separating citation, domain-text mention, brand-text mention, disagreement, and unreadable
• Planned, delivered, excluded, unreadable, and usable record counts, reconciled
• The response hash behind every scored record

Limits: This is a private internal-QA artifact over a frozen panel of manually collected transcripts. It measures what a set of assistants answered on the dates we ran them; it does not predict future answers, and it is not a ranking service. Our own run above scored 0 answer-URL mentions across 6 prompts — a real result, not a demo tuned to look good.

Each prompt is run in a fresh session, the readable answer text is hashed and stored, and the outcome is decided by a fixed rule against the target URL and the exact asset name. An answer we could not read is 'unreadable', never a miss, because those are different facts. This is bounded evidence of the dated method stated here; it does not establish affiliation, endorsement, authorization, or any current outcome beyond that scope.