\section{Research questions and contributions}
\label{sec:questions}

The paper addresses eight questions. (RQ1) Which housing characteristics consistently affect transaction prices across countries
and periods? (RQ2) How large are their marginal contributions and how uncertain are they? (RQ3) How do implicit prices vary with
geography, market conditions, property types and socioeconomic characteristics? (RQ4) Have valuations changed over time,
particularly around financial crises, monetary-policy regimes and the COVID-19 pandemic? (RQ5) Do traditional hedonic, spatial
and machine-learning approaches produce systematically different estimates? (RQ6) What additional price information is captured
by textual, visual and other AI-derived characteristics? (RQ7) How much of the observed variation reflects genuine economic
heterogeneity rather than specification, measurement and publication selection? (RQ8) Which characteristics and markets remain
understudied?

The contribution is defined by what the Sirmans programme could not do with the tools and the literature of 2005. First, its
coverage was the United States, single-family houses and nine structural characteristics; the present synthesis is global,
covers all residential property types and the full taxonomy of structural, interior, exterior, locational, environmental,
neighbourhood, market and AI-derived characteristics, and treats rents and listing prices as separate outcomes. Second, the
2005 paper counted signs and the 2006 paper regressed raw coefficients across functional forms; the present synthesis harmonises
coefficients into effect-size families with explicit formulas, weights them by precision and reports prediction intervals that
describe the dispersion of implicit prices rather than only the precision of their mean. Third, the Sirmans meta-regressions did
not model the dependence of multiple estimates from the same study or test for publication selection; the present synthesis uses
three-level models with robust variance estimation, funnel-based and selection-model diagnostics, and Bayesian and
machine-learning cross-checks. Fourth, it adds the methodological and AI dimensions that did not exist in 2005. The claims of
this paper are bounded by its own search and retrieval process, documented in Section~\ref{sec:search}: it synthesises the
open-access literature that could be retrieved and extracted within a reproducible pipeline, and it compares its coverage with
every competing review it could verify (Table~\ref{tab:competing}) rather than asserting that it is the largest.