{
 "question": "How do AI assistants construct and communicate a brand\u2019s reputation from external evidence when asked whether the brand is legitimate, and how does that construction differ by engine?",
 "population": "79 brands (40 named most widely and 39 named by one assistant in study 7, eight industries) x 4 engines (ChatGPT and Gemini consumer apps, Google AI Mode, Perplexity sonar API), one run each, US, 26 September 2026: 316 answers. Version 1.1 re-analyzes the same answers; no new engine queries.",
 "inclusion": "Every answer, including incomplete answers and answers with no citation (kept as outcomes).",
 "exclusion": "",
 "calculations": [
  "Items: each answer with inline citations replaced by source tags; engine name hidden from coders.",
  "Primary coding (Claude Opus): answer status, first-sentence frame, legitimacy position, overall structure (six categories), uncertainty, and every atomic reputation claim with polarity, function, attached source tags, paragraph and prevalence wording.",
  "Second coding (Claude Sonnet), blind to engine and to the primary answer-level codes: frame, legitimacy, structure, uncertainty and negative-first-paragraph for all answers; claim polarity and function for the claims of 80 randomly drawn answers. Agreement as percentage and Cohen\u2019s kappa.",
  "Source taxonomy: each cited registrable domain assigned one of eight classes by a model coder (Claude Opus); domains owned by the brand or its parent, identified per brand by the same coder, are reclassed as brand-owned.",
  "Concentration: domains counted once per answer; Herfindahl-Hirschman index over domain shares, effective number of sources = 1/HHI, top-1/3/5 shares.",
  "Evidence diversity: distinct source classes per answer against distinct domains per answer (Spearman).",
  "Claim support: 60 cited claims drawn at random per engine (seed 20260928); evidence = the passage Gemini\u2019s link quotes (text fragment) or the cited page fetched over plain HTTP on 2026-09-28; judged by Claude Opus and independently by Claude Sonnet; claims with no readable source stay in the denominator and bounds are reported.",
  "Engine differences: GEE logistic regression clustered by brand (exchangeable), fixed effects engine, brand group and industry; brand random-intercept logistic model as a check; engine x group Wald test.",
  "Intervals: 95% bootstrap resampling brands (2,000 resamples, seed 20260926)."
 ],
 "limitations": [
  "One run per engine and brand on one date: the stability of a reputation across runs, wordings and dates is not measured.",
  "Citation is not influence: the data show which sources the engines cite, not that those sources caused the answer.",
  "All coding and support judgments are model-coded (two Claude models); no person coded the sample.",
  "Review platforms (Trustpilot, BBB, ConsumerAffairs) block plain fetching, so support could be judged mainly for other sources and for Gemini\u2019s quoted passages; pages were fetched two days after the answers.",
  "Source independence (whether several cited sites repeat the same underlying reviews) is not measured; source classes are a proxy for diversity.",
  "Perplexity was tested through its API; the brand sample leans toward well-known US companies."
 ],
 "update": "quarterly",
 "@context": "https://schema.org",
 "@type": "Dataset",
 "name": "\u201cIs this brand legit?\u201d How AI assistants build a reputation",
 "version": "1.1",
 "dateCreated": "2026-09-26",
 "creator": {
  "@type": "Organization",
  "name": "Underneath",
  "url": "https://underneath.agency"
 },
 "url": "https://underneath.agency/research/is-it-legit-ai-reputation-study",
 "temporalCoverage": "2026-09-26",
 "isAccessibleForFree": true,
 "license": "https://creativecommons.org/licenses/by/4.0/",
 "code": "cite/pipeline/ (collection, validation and analysis scripts)",
 "variableMeasured": [
  "collected",
  "reanalyzed",
  "answers",
  "brands",
  "engines",
  "answer_outcomes",
  "source_classes_share_of_answers",
  "source_classes_share_of_citations",
  "concentration",
  "top_domains_share_of_answers",
  "domain_class_counts",
  "evidence_diversity",
  "complete_answers",
  "structure",
  "legitimacy",
  "first_sentence_frame",
  "first_sentence_affirms_any",
  "legit_but",
  "uncertainty",
  "claims",
  "answers_with_function",
  "polarity_by_source_class",
  "engine_profile",
  "models",
  "by_group",
  "cross_engine",
  "validation",
  "support",
  "page_fetch"
 ],
 "dateModified": "2026-09-28",
 "research_questions": {
  "RQ1": "Which classes of source do the engines cite when judging a brand\u2019s legitimacy and reputation, and how concentrated are they?",
  "RQ2": "Do the engines draw on different evidence ecosystems, net of brand, brand group and industry?",
  "RQ3": "Which positive and negative claims do the answers make, and how prominent are the negative ones?",
  "RQ4": "How often do the cited sources support the claims attached to them?",
  "RQ5": "How stable is the reputation an engine gives a brand across runs, wordings and dates? Not answered: one run per engine and brand.",
  "RQ6": "Are source classes associated with the polarity of the claims attached to them? Answered descriptively only."
 },
 "framework": "Reputation representation R = (L, S, C, E, U): legitimacy framing, overall reputation structure, complaint profile, evidence structure, uncertainty; each coded per answer."
}