# Study 9 strengthening (ChatGPT local recommendations vs Google Maps). Frozen before the second wave was collected
# (2026-09-27). Observational only: no content interventions exist in this study.
seed: 20260927
bootstrap_reps: 2000

wave1:
  date: "2026-09-26"
  runs: [1]
  wording: original            # "Who is the best {service} in {city}?"

wave2:
  date: "2026-09-27"
  instrument: "DataForSEO LLM Scraper, ChatGPT consumer app, location = country (same as wave 1)"
  maps: "DataForSEO Google Maps SERP, depth 20, keyword 'best {service} in {city}', location = country (same as wave 1)"
  original_runs: [2, 3, 4]     # same wording as wave 1, three runs minutes apart
  variants:                    # one run each; the type says what the variant changes
    p1: {type: semantic_paraphrase, template: "Which {service} in {city} would you recommend?"}
    p2: {type: semantic_paraphrase, template: "I need a good {service} in {city}. Who should I go to?"}
    p3: {type: list_format, template: "List the top {services} in {city}."}
    p4: {type: source_required, template: "Who is the best {service} in {city}? Please cite your sources."}

# Services as asked (US uses "physical therapist" for physiotherapist, as in wave 1)
plurals:
  dentist: dentists
  plumber: plumbers
  personal injury lawyer: personal injury lawyers
  accountant: accountants
  physiotherapist: physiotherapists
  physical therapist: physical therapists

# Not testable in this study (documented, not simulated): content interventions, placebo/length controls,
# fidelity-constrained interventions, retrieval and context inclusion (the consumer app exposes neither),
# cross-engine field comparison (Gemini's capture returns text only, no structured business listings).

# Added 2026-09-28, before the third collection. Purpose: separate run-to-run variation from day-level variation
# (not to claim that ChatGPT changed). Same instrument and settings as wave 2.
wave3:
  date: "2026-09-28"
  start_utc: "2026-09-28T16:15:00Z"  # 24 h after wave 2 and at the same time of day, so day and time of day are not confounded
  original_runs: [5, 6, 7]     # original wording only, three runs
  maps_pulls: [a, b]           # two Google Maps pulls the same day (a before, b after the ChatGPT runs): same-day Maps baseline
  raw: cite/data/raw/day3_localrec/{llm,maps_a,maps_b}

# Analysis revisions decided 2026-09-28 before the third collection was analysed:
#  - entity resolution is branch-aware: a shared registrable domain counts as the same business only when the street
#    key (number + first street word) agrees or one side has no address; otherwise "same company, other branch"
#  - wording-sensitivity outcomes: list shown, cites any source, business-set overlap with the original wording
#    (vs the original-vs-original baseline), identical-link share; the Maps-top-20 share is dropped (Maps query unchanged)
#  - coverage is reported in both directions with explicit denominators; citation traceability keeps an
#    "undetermined (page not fetched)" category in the denominator

# Deviation log (after the model coders coded the targeted sample, before any human labels):
#  - 2026-09-28: the postcode of a Maps entry is now read from its full address text first, because Google's
#    structured zip field is sometimes wrong (item H030: zip "1811" = street number). General fix, not item-specific;
#    it moved 8 of 12,271 listings from other_branch to confirmed. Found through a validation item, so disclosed.
#  - 2026-09-28: the third collection (wave3) was cancelled by the study owner before it ran; no day-3 data exists.
#    The study ends with two dates (26 and 27 September).