Noise Floor Measurement for AI Citation Experiments: Platform Variance, Recording Artifacts, and a Four-Test Diagnostic Protocol

Abstract

This paper reports a five-day noise floor measurement for AI citation rate experiments on thegeolab.net, conducted 13–17 April 2026 with zero content changes. Using a 10-query universal set across three platforms (Perplexity sonar-pro, ChatGPT via chatgpt.com scraper, Google AI Overviews), we recorded 150 individual citation checks and established a combined citation rate range of 13.3%–20.0%. Perplexity showed zero inter-day variance (4/10 cited on all five days, identical pages, identical annotation order across eight total observations including three intraday reruns). ChatGPT ranged 0–2 citations per day, with one outlier day (Day 2: 2/10) attributable to web-search non-determinism. Google AIO triggered on most SERPs but cited thegeolab.net zero times across all five days. A key finding was that an apparent Day 1→2 URL "flip" was a recording artifact (single-URL capture of a multi-URL annotations array), not platform variance — confirmed via intraday replication. Zero mention (M) results were recorded across all 150 checks: citation is binary. The paper presents a four-test diagnostic protocol (intraday repeat, content comparison, GSC impressions, crawler log analysis) for attributing daily citation variance before assuming platform noise. The established noise floor (max 20.0%) means future experiments on this domain must exceed a 22.0 percentage-point lift threshold to be considered interpretable signal. Version — 1.0

Ähnliche Arbeiten