1\section{Systematic search strategy and study selection}2\label{sec:search}34The search, screening and selection follow the PRISMA 2020 principles \citep{page2021prisma} adapted to economics research, with a5protocol written before any coefficient was extracted (Appendix~\ref{app:protocol}). Figure~\ref{fig:prisma} gives the flow of6records generated from the actual screening files and Table~\ref{tab:coverage} the counts at every stage.78\subsection{Sources and queries}910Records were identified on 2026-10-20 from four machine-readable sources: the Semantic Scholar bulk search API, with Boolean11queries combining the hedonic vocabulary with residential-price terms and with the twelve multilingual phrasings listed in the12protocol; the Crossref bibliographic search, capped at 4{,}000 relevance-ranked records per query; the arXiv API for the13machine-learning, vision and text literature; and OpenAlex, which could be used only for verification because its unauthenticated14access is limited to a shared daily budget. Forward citation snowballing was applied to the two Sirmans papers through the15Semantic Scholar citation graph. In total \nQueries{} queries were executed and \nRecordsIdentifiedRaw{} records retrieved;16deduplication on DOI and on normalised title plus year left \nRecordsAfterDedup{} unique records, \nRecordsWithAbstract{} of17which carry an abstract. SSRN, Google Scholar, IDEAS/EconPapers and institutional repositories have no compliant bulk interface18and were not harvested directly; the grey literature is covered only insofar as Semantic Scholar and arXiv index it. The audit19trail (queries, dates, hit counts, retrieval failures) is in the replication material.2021\subsection{Screening}2223Stage-one screening proceeded in two steps. A rule-based scorer applied the keyword groups of the protocol (hedonic vocabulary in24twelve languages, residential-price outcomes, regression and characteristic signals, hard exclusions for non-housing hedonic25applications) to titles and abstracts and retained \nStageOneCandidates{} candidates. Each candidate was then screened on title26and abstract by an independent agent working from the written eligibility criteria, with a coded reason for every decision;27\nLlmExcluded{} records were excluded, \nLlmUnclear{} were flagged as unclear for human review, \nIncludedHedonicStageOne{} were28included as hedonic coefficient studies and \nIncludedMlPredictiveStageOne{} as machine-learning or automated-valuation studies29that feed the synthesis of Section~\ref{sec:ai}. The most frequent exclusion reasons were non-housing applications of the hedonic30method (sensory science, art, wine, labour markets), aggregate-level price regressions without dwelling characteristics, and31reviews. The coding conventions that required a decision across screeners, such as the treatment of hedonic price-index32papers, of community-level Chinese studies and of short-term-rental pricing, are recorded in the methodological-decisions log.3334\subsection{Full-text retrieval and eligibility}3536Full texts were sought for every included hedonic study and every unclear record (\nFulltextSoughtHedonicOrUnclear{} records)37from legal open-access locations only: Unpaywall locations, Semantic Scholar open-access links and arXiv. \nFulltextRetrieved{}38PDFs were obtained (\retrievalRate{} percent); the remainder are paywalled or have no open-access copy and are counted as not39retrieved. This is the binding constraint of the study, and Section~\ref{sec:selection} compares the retrieved and not-retrieved40sets on observable metadata. Of the retrieved full texts, \nFulltextAssessedExtracted{} were assessed and extracted by the41pipeline described in Section~\ref{sec:extraction}; \nFulltextEligible{} were eligible for the coefficient synthesis and42\nFulltextIneligible{} were excluded at full text, most often because the unit of analysis was an aggregate (city, tract, housing43project) or because characteristic coefficients were controlled for but not reported.4445\begin{figure}[H]46\centering47\includegraphics[width=0.95\textwidth]{fig01_prisma_flow.pdf}48\caption{PRISMA flow of records, generated from the screening and retrieval logs of the pipeline.}49\label{fig:prisma}50\end{figure}5152\begin{table}[H]53\centering54\caption{Literature coverage and study selection}55\label{tab:coverage}56\small57\input{tables/tab01_coverage}58\end{table}59