E101: Per-Render Citation Variance in Google AI Overviews — Frozen First-Party Dataset (5 August 2026)

This dataset is a frozen, first-party observational record (GEO Lab experiment E101) of per-render citation variance in Google AI Overviews. On 5 August 2026 one query ("web designer for tradespeople UK") was run ten times in a single afternoon against one page the author owns (the AJ Web Design trades page), on Google AI Overviews, with two DataForSEO API pulls alongside browser captures. Nine of ten realizations resolved; across those nine the page was cited in six and absent in three (a rate near 60% for this one page-query pair on this one day), with every browser observed more than once producing both outcomes. Each resolved render was a materially different document, with 9 to 19 sources. Finding: under the supplied access configurations, the same nominal page and query produced non-mergeable citation outcomes across renders, so citation presence is a per-render draw, not a stable state. Four candidate predictors (an API-versus-browser blind spot, browser family, template framing, and a dated drop) were generated and rejected on the same day's data. Methodological payload: a single observation of an AI Overview citation is not evidence of presence or absence for an intermediate-rate page; E101 establishes a minimum-N rule of no citation presence or absence claim from fewer than five renders across at least two declared access configurations, reporting the rate and the sample size. Bounds, stated deliberately: n equals one page, one query, one day, so the ~60% figure is an existence proof of per-render variance at intermediate rates, not a prevalence estimate; API captures are wire-level and hash-verified, whereas browser captures are declared-profile screenshots with incomplete sentinels, so the record evidences observed variance motivating replication, not a demonstrated within-declared-profile split. Integrity: the canonical evidence packet is receipt-zero-v2-corrected.zip; its manifest file receipt-zero_PACKET.json has SHA256 4584652f643127d612894fa77385397b1c16461e026e1f305aec413d477fa1b3, independently byte-verified by a second party (match, 16 August 2026), and it lists per-file hashes for all raw API JSON captures and all eight browser screenshots. Attribution: this is the author's independent first-party finding; the continuity-receipt and drift-control methodology draws on the PHI-OMEGA Runtime framework by Massimiliano Brighindi, credited as method, and a preregistered replication round on rate stability and mechanism is a separate, jointly-governed study in preparation, linked rather than merged.