mirror-eval · perplexity · 29 days
13 of 63, then 63 of 63. A cluster of changes.
The note
A calendar strip of 29 days. A band lights the repo's dates for my changes, 7 to 20 August, six change cards drop in below it, and the counter under the strip runs from 13 to 63.
That counter is Perplexity: 13 of 63 probes about my work whose answer cited my site in the first wave of my AI search audit, 63 of 63 in the second.
In between I changed the surfaces the engines read. The repo counts six changes, five of them listed on the case-study page, and notes the cluster landed between 7 and 20 August. The second wave ran on 4 September.
The other engines did not move with it. ChatGPT went from 62 to 54 on the site count and 62 to 63 when repos are included. Gemini went 61 to 63. Claude stayed at 0.
I cannot tell you which change did it. They went out as a cluster, I have one subject, and the engines' versions were not verifiably frozen across the month. The repo has a noise floor, but it covers the composite quality score, not citation counts, so there is no significance test behind 13 to 63. There is a count, and a gap of fifty probes that one engine closed and another did not touch.
Next time I would stagger the changes by a week each and run a wave between.
Would you keep a fifty-probe win you cannot attribute, or unwind it one change at a time to find the lever?
Where the numbers live
The case study, with the five listed changes and both waves
One subject. Changes went out as a cluster. No test exists for the count. Versions not verifiably frozen.