DailyTech

Research

Measurements, not estimates

The desk publishes original measurement work when the answer to a question cannot be found in existing public material. Each release names its research question, the exact prompts or inputs used, the collection date, the sample size, the scoring formula, what was excluded and what cannot be concluded. The underlying data is downloadable, so the numbers can be checked rather than taken on trust. We never publish a score without showing how it was calculated.

AI Recommendation Stability Index

How stable are AI product recommendations? 160 repeated prompts across two models

We asked two production language models the same eight buying questions ten times each and recorded every name they returned. Stability varies more by category than by model: infrastructure answers barely move, while service categories such as SEO agencies and cyber insurance change from run to run.

Release 2026.09-v1 · Collected 18 September 2026 · 9 min read · Data: JSON and CSV

Standing rules for this section

  • A number is published only with the method that produced it.
  • Recurring measurements keep their prompts and scoring fixed between releases.
  • Raw data ships with every release, under a stated usage licence and citation.
  • Corrections are published against the release version rather than edited in silently. Send them to corrections@dailytech.wiki.