# Methodology

## Fixed design

Three convenience-selected US cities × three cosmetic-surgery procedures × two wordings = 18 exact cells. Three sequential same-day waves across ChatGPT, Gemini, Google AI Overviews and Google AI Mode produce 216 scheduled slots. The plan and protocol were fixed before paid collection. No clinic or practitioner brand was seeded. The evidence wording intentionally requests more information and is not an isolated causal prompt experiment.

Fieldwork: 2026-10-30T14:08:10.698Z–2026-10-30T15:08:51.169Z. Excluded technical collection attempts: 4 (4 preflight attempts across 4 fixed technical cells; 0 diagnostic probes). Read-only queue GET operations: 526; they are retrievals, not additional collection attempts.

- ChatGPT: 54 answer-bearing; 0 successful no-answer; 0 collection-error states from 54 planned slots.
- Gemini: 54 answer-bearing; 0 successful no-answer; 0 collection-error states from 54 planned slots.
- Google AI Overviews: 48 answer-bearing; 6 successful no-answer; 0 collection-error states from 54 planned slots.
- Google AI Mode: 54 answer-bearing; 0 successful no-answer; 0 collection-error states from 54 planned slots.

## Actual requests and recovery

First-pass core requests: 216; original technical failures: 0; approved missing-only recovery attempts: 0; recovered captures: 0; total core requests: 216. Observations are not request attempts. No-answer/no-source outcomes are not replaced. Original status and both timestamps remain in public metrics. Any delayed recovery retains its scheduled label and actual completion time.

## Observable surfaces

DataForSEO LLM Scraper live/advanced for ChatGPT and Gemini; ChatGPT force_web_search=true. Google AI Overviews and AI Mode used Standard queue task_post and task_get/advanced at normal priority 1 from preflight onward. AI Overviews retained depth 10 and load_async_ai_overview=true; AI Mode retained its distinct Google surface. US-English location 2840; Google desktop/Windows; city named in the query. Provider-returned model labels are retained where supplied. Physical city IP, signed-in personalization and provider-side caching were not controlled. These are provider-collected surfaces, not ClinicTruth owner reports or consumer-UI replicas.

## Delivery counts

- chatgpt: 54 retained Live captures; 0 Standard queue captures.
- gemini: 54 retained Live captures; 0 Standard queue captures.
- ai_overviews: 0 retained Live captures; 54 Standard queue captures.
- ai_mode: 0 retained Live captures; 54 Standard queue captures.

Scheduled wave labels and actual collection timestamps remain separate. Standard Google queue delivery was prospectively fixed, not a post hoc platform substitution.

## Citation inputs and normalization

ChatGPT/Gemini use declared answer sources, not unused search results, image/ad/navigation URLs. Google uses references inside top-level AI answer items, not ordinary organic/local/PAA results. HTTP(S) references are deduplicated per answer after hostname/www, fragment, trailing slash and known tracking normalization. Nontracking query parameters and HTTP versus HTTPS remain distinct. Hostname is not corporate ownership.

## Source-set comparisons

Jaccard = intersection / union. Both answers must exist. Empty/empty source sets are undefined, one-empty pairs are zero, and a separate both-nonempty sensitivity view is provided. Means weight defined pairs equally. Three within-cell pairs × 18 cells × 4 surfaces = 216 planned within pairs. Six platform combinations × 54 exact-query/scheduled-wave matches = 324 cross-platform pairs. Actual capture timestamps remain available. No probability confidence interval or patient-demand weighting is implied.

## URL-codebook

Known directory/social hosts are coded first, then homepage, editorial, treatment/services, location/contact, explicit practitioner/team, and unclassified. Specific procedure paths take precedence over generic provider or location prefixes. These are navigation proxies, not content-quality or causal-source labels. The complete rules are in url-codebook.mjs and every recorded type is independently recomputed.

## Predetermined source follow-up

36 slots: city × procedure × surface, ordinary-selection wording, wave 1, first declared normalized reference. The exact originally selected URL remains; no convenient replacement is made. Repeated selected URLs retain every slot but each unique page is inspected once. Saved source bytes retain SHA256 and retrieval time privately. Positive procedure/surgeon/location/contact/fee-context evidence is recorded only for the inspected scope. Unknown or inaccessible facts remain unknown. Static text alone cannot establish whole-site absence or missing image/dynamic content. Consultation-location and operating-facility evidence are separate; an address alone does not establish where surgery takes place. Named practitioner information is not license or clinical-quality validation. No automated clinic-accuracy rate is inferred.

## Interpretation and provenance

This convenience panel is not all US clinics or searches. Same-day repeat source variation does not establish long-term stability. No website intervention was applied; practical recommendations are Rotgar interpretations and testable hypotheses, not measured causal ranking effects. Historical Plastic Surgery source-overlap data retain their original population and surface definitions; nonsurgical medspa, Dental, IVF and other datasets are not relabeled as cosmetic-surgical findings. No human methodology/clinical review is claimed.

Protocol SHA-256: 86a567b0de7a7cc232238285d37874ba5309d55a0b1b35e2f2b6aca695bee5b4
Analytical JSON SHA-256: 348475805609f4089002e5203a57e7359fc0d50e2cafbe95ba1d5a5fd4cbbb97