Cosmetic Plastic Surgery AI Search Visibility Across Four Platforms

Fix the cosmetic-surgery panel before collecting answers

The protocol and 18 exact query cells were fixed before paid collection. Three convenience-selected US markets were crossed with rhinoplasty, facelift and breast augmentation, using ordinary-selection and surgeon/facility/evidence wordings. No practice brand was seeded in a query.

Collect the four surfaces separately

DataForSEO LLM Scraper collected ChatGPT and Gemini with Live delivery. ChatGPT used force_web_search=true. Google AI Overviews and AI Mode used Standard queue delivery at normal priority 1 from preflight onward. AI Overviews used depth 10 and asynchronous loading; AI Mode used its separate Google surface. US-English country-level location 2840 was used; Google requests used desktop and Windows settings, and city names were explicit in each query.

Returned model labels — ChatGPT: gpt-6; Gemini: 3.5 Flash-Lite. Google AI Overviews and Google AI Mode did not return model-version labels. This collection does not establish equivalence to every signed-in consumer session.

Keep repeats and collection states visible

The core contains 216 scheduled observation slots: three sequential same-day waves, 54 per platform. 0 original temporary failures received one prospectively admitted exact-input recovery after every original first-pass slot was terminal. A successful answer or no-answer capture was never replaced. Original and recovery delivery modes, statuses and timestamps remain in the observation metrics. There were 216 core requests in total. 4 technical collection attempts are excluded: 4 preflight attempts across 4 fixed technical cells and 0 separately admitted diagnostic probes. 526 read-only queue GET calls retrieved already-posted tasks; they are not additional observation attempts.

210 final core slots contained an AI answer; 6 checks returned no AI answer; 0 retained an unresolved collection error. Original failure states and both timestamps remain in the public observation metrics. Scheduled repeat labels are not a claim that delayed recoveries occurred in their original chronological wave. Provider-side caching was not controllable.

Count final-answer citations and source-set similarity

ChatGPT/Gemini use the provider’s declared final-answer sources, not unused search results. Google uses references inside top-level AI answer items, not ordinary organic listings, local packs or People Also Ask. URLs are deduplicated within an answer after hostname/www, tracking, fragment and trailing-slash normalization. Hostname means the normalized host, not corporate ownership.

Jaccard is intersection divided by union. Empty/empty citation sets are undefined, one-empty sets are zero, and both-nonempty summaries are reported separately. Within-query comparisons match exact query and platform; cross-platform comparisons match exact query and scheduled repeat label. Delayed recoveries are not represented as chronologically simultaneous captures.

Inspect a fixed first-reference source sample

The qualitative sample uses the first plan-ordered selection wording, scheduled wave 1, then the first declared normalized citation. A same-input recovery of a temporary failure may fill that original slot later; the source ledger flags recovery and actual timestamps. No-answer, no-source and unresolved-error slots remain explicit. Dated source-page evidence supports only what was actually visible on that page.

The published codebook classifies URL navigation patterns conservatively. Treatment paths take precedence over generic location or practitioner prefixes. A path label is not a review of page quality or proof that a particular block caused a recommendation.

The public data omit full AI answers, API identifiers, credentials and private ClinicTruth project information. They retain prompt cells, observation metrics, citations, pair inputs and review evidence needed to inspect the analysis.