Field notes on AI accuracy.
What the models get right, where they disagree, and how the answer keeps changing.
Why three models matter — and which three we picked
Consensus is a function of model diversity. We tested eleven LLMs against the same sixty brand questions. Three providers captured 94% of the disagreement signal. Here's the math, the methodology, and the one surprising omission.
The llms.txt file is doing more than you think
A 4,000-domain audit reveals that brands with a well-formed llms.txt score 28 points higher on average — even when their JSON-LD is identical.
The AI Overview takeover, charted
Google AI Overviews now appear on 47% of commercial queries. We pulled 12,000 SERPs to see which industries got hit hardest.
How a holiday-tech startup rebuilt its AI footprint
Christmas or Not went from "ChatGPT doesn't know who you are" to top consensus in under a month. Here's the playbook.
Measuring agreement: the κ statistic for LLMs
Cohen's kappa was built for human raters. We adapted it for three stochastic models. The math is uglier than you'd hope.
Monitor Pro: weekly scans, drift alerts, badge API
Brands move slowly. Models move fast. Monitor Pro closes the gap with weekly automated re-scans and a CDN-cached AI Readiness Badge.