This is the research layer of blimpp’s AI search work: original, dated studies built on first-party measurement rather than commentary on other people’s data. This page sets out the programme before its studies publish, so the questions, methods and standards are on record in advance. Studies appear here as they complete; none is announced with a date it has not earned.
The planned studies
Citation volatility. How often does the same AI search produce a different source set, and how does stability vary by engine and question type? This study establishes citation volatility as a measured baseline and demonstrates why single-response screenshots are unreliable evidence of anything.
The Reddit influence index. When and where do Reddit threads shape commercial AI recommendations? Measured through the citation rate and carry rate defined under Reddit influence, across prompt classes and engines.
What AI engines cite for buying questions. Which domains and source types feed AI answers when buyers ask for recommendations in a category, starting with beauty and DTC. Produces source-type share and publisher findings per the source-type classification.
Cross-engine source survival. Which sources are cited by multiple engines answering the same prompt set, and which live on one engine only? Builds on cross-engine citation overlap.
The standard every study must meet
The research question is stated before the finding. The prompt universe and sampling method are defined per prompt sampling. Probabilistic outcomes use repeated runs per the repeated-run protocol; a single answer is never treated as a stable result. Engine, date and locale are recorded. Citations, mentions and recommendations are reported as separate outcomes. Every percentage carries its sample size, exclusions and failures are disclosed, and every study includes a limitations section with enough methodological detail for another practitioner to approximate the analysis.
Using this research
Findings published here may be cited with attribution to the study title, blimpp as publisher, the year, and the study’s canonical URL. Where a dataset can be shared safely, studies include an aggregate table for reuse.
Editor: Harpal Singh · Last reviewed: 7 August 2026