\section{Introduction}
\label{sec:introduction}

Two decades ago, \citet{sirmans2005composition} and \citet{sirmans2006value} asked a question that every appraiser, assessor
and housing economist faces: across the hundreds of hedonic regressions of house prices on characteristics, what is the value
of a bathroom, a garage, a swimming pool, a year of age or a square foot of living area, and does that value depend on where and
when it is measured? Their answer, built from a tabulation of signs and a meta-regression of raw coefficients from mostly
American single-family studies, was that the signs are consistent, that magnitudes vary by location and sometimes by data type,
and that there was no clear change over time. The literature has since grown by an order of magnitude, has moved to apartments,
rents and listing prices in Asia and Europe, has adopted spatial econometrics, quasi-experimental designs and machine learning,
and has begun to extract characteristics from text and images. The tools of research synthesis have changed as well: effect-size
harmonisation, multilevel models for dependent estimates, cluster-robust inference, publication-selection diagnostics and
Bayesian hierarchical models are now standard in economics meta-analysis \citep{stanley2012meta,nelson2009use}. No synthesis has
applied them to the hedonic housing literature as a whole: the meta-analyses that exist are attribute-specific
(transit, noise, green space, water, flood risk, contamination, energy efficiency, school quality), and the general reviews
are descriptive. This paper fills that gap to the extent that the accessible literature allows.

The paper answers eight questions: which characteristics consistently affect prices across countries and periods; how large
and how uncertain their marginal contributions are; how they vary with geography, market conditions, property types and
socioeconomic context; whether they have changed over time; whether conventional, spatial and machine-learning approaches
differ; what text, vision and other AI-derived characteristics add; how much of the observed variation is specification,
measurement and publication selection rather than economics; and what remains understudied. The approach is a systematic review
with a pre-registered protocol, a global variable taxonomy, a documented harmonisation of coefficients into effect-size
families, and a meta-analytic framework that respects the dependence of estimates within studies. Every step is reproducible
from the replication repository and every number in the text is generated from the result files.

Three features of the design deserve emphasis because they bound what the paper can claim. First, retrieval was restricted to
legal open-access full texts; the \retrievalRate{} percent retrieval rate is documented in the PRISMA flow and the extracted
sample is therefore a sample of the accessible literature, with the biases that this implies. Second, screening and extraction
were performed by language-model agents working from written protocols, with an independent double extraction of a random sample
to measure their reliability; the agreement on matched coefficients is complete but the recall of auxiliary variables is not, and
the paper reports both. Third, the synthesis distinguishes throughout between a descriptive synthesis (which characteristics are
studied, with what sign) and a statistically valid meta-analysis (which implicit prices can be pooled within a defensible
effect-size family), and it counts independent studies, not coefficients, when it describes its evidence base.

The results can be summarised briefly. The structural core of the hedonic model is global: living area, bathrooms, bedrooms
conditional on area, age and distance to the centre have precisely estimated mean implicit prices that are similar across
continents and decades, and the signature amenities carry premiums of the order documented in attribute-specific studies.
Every implicit price is heterogeneous beyond sampling error, with prediction intervals that span zero for most characteristics,
and the observable features of studies, including spatial estimators, explain little of that heterogeneity; funnel asymmetry is
widespread but cannot be read as selection under this degree of heterogeneity. Machine-learning models predict prices better than
hedonic benchmarks in most of the comparisons that could be retrieved, and AI-derived characteristics add modest accuracy and
some interpretable variables. The gaps are as informative as the estimates: interior amenities, market characteristics, several
environmental risks, AI-derived characteristics and three continents remain thinly covered.

The remainder of the paper proceeds as follows. Section~\ref{sec:theory} recalls the theoretical foundations and
Section~\ref{sec:history} the development of the literature; Section~\ref{sec:questions} states the research questions and the
contribution relative to the Sirmans programme. Sections~\ref{sec:search} and~\ref{sec:extraction} describe the search, selection,
extraction, validation and harmonisation, and Section~\ref{sec:framework} the econometric framework. Sections~\ref{sec:evidence}
to~\ref{sec:ai} present the evidence by characteristic, geography, period, method and AI, Section~\ref{sec:selection} the
publication-selection diagnostics and robustness, Section~\ref{sec:implications} the implications, Section~\ref{sec:limitations}
the limitations and Section~\ref{sec:conclusion} concludes. The appendices contain the protocol, additional tables and the
reproducibility statement.