Paper B: Field Baseline Results. Does presentation format affect blind entity retrieval and brand attribution in production generative engines?

Abstract

Results for the pre-registered field baseline deposited as version 1.0 of this record (DOI 10.5281/zenodo.22798718, 16 September 2026). This version reports the registered dependent variables, completes Section 9 of the pre-registration from the session logs, and deposits the evidence package. Across 40 confirmatory sessions run from the UK on 1 October 2026, neither engine retrieved any of the four target pages. The design covered four entities across ChatGPT and Gemini, with five fresh sessions for every engine-entity combination, using the registered URL-free prompt through the desktop web interfaces. retrieved was 0/5 in all eight cells, and retrieval_shown was 0/40. Brand attribution could not be estimated because no session retrieved an entity. A supplementary cross-day run on 2 October also returned 0/8. The combined result was 0/48, with a Wilson 95% upper bound of 7.4%. All four pages were confirmed as indexed on Google from the UK on 1 October using the registered site: check. During this test window, that indexation did not lead to blind retrieval on either consumer surface. With retrieval at zero throughout, the registered primary comparison (presentation format on attribution) cannot be estimated. The result supports a narrower statement: during this window, placing a new entity on a single indexed source did not lead ChatGPT or Gemini to retrieve it from a blind prompt. Engine behaviour differed in texture and is reported as observation, not mechanism, in line with Section 12 of the pre-registration. Blind second view: 8 of 40 sessions (20%) coded independently by the co-author, raw agreement 100%. Cohen's kappa is undefined because both coders produced zero variance; the registered halt condition (kappa below 0.60) was not triggered. Deviation log D1 to D7 is included in the results document. Nothing was reconstructed after the fact; where a field was not captured, the record says so. Files: paperB-results-v1.0.pdf and .md (results document); PaperB_field_scoring_v1.0.xlsx (completed scoring workbook, both coders); paperB_raw_confirmatory_2026-10-01_UK.txt (40 confirmatory and 8 void sessions, full response text); paperB_raw_supplementary_2026-10-02_UK.txt (8 sessions); paperB_raw_G2_gate_2026-09-24_UK.txt (24 gate sessions); paperB-prereg-v1.0.pdf and .md (the pre-registration as deposited in version 1.0); MANIFEST.md5. Competing interests: Julio Arévalo Piedra founded and operates KusiGEO, a commercial GEO-audit platform. This study did not test KusiGEO or any other commercial product, and no KusiGEO tooling was used in its design, deployment or measurement. Julio did not run any confirmatory sessions; his measurement role was limited to the blind second-view coding. Artur Ferreira operates The GEO Lab, an independent GEO research site, and declares no competing interest in the outcome. Drafting disclosure: the confirmatory sessions, gate checks, coding, verification and data handling were carried out by the authors. The text of the results document was drafted with AI assistance using Anthropic Claude, working from the session logs and the deposited pre-registration. The authors reviewed, corrected and finalised the document and take full responsibility for its contents.