\section{Implications for real estate economics and appraisal}
\label{sec:implications}

For appraisal and mass valuation the central result is the pair of numbers attached to each characteristic: a precisely
estimated global mean and a wide prediction interval. The mean is a defensible prior for a market without local evidence, and the
prediction interval is the honest statement of how much local evidence can be expected to move it. For the structural core the
interval is narrow enough to be useful: a living-area elasticity between 0.5 and 0.8, an age effect of a fraction of a percent
per year, a bathroom premium of a few percent, a distance gradient of a few percent per kilometre. For amenities and
disamenities the interval is wide, and transferring a premium from one market to another without local calibration is not
supported by the evidence. For research practice, three implications follow from Sections~\ref{sec:methodological}
and~\ref{sec:selection}. Reporting standards matter more than estimator choice: a large share of the extracted coefficients could
not enter the pooled analysis because standard errors were replaced by stars, units were not stated, construction years were
used in place of age, or the dependent variable's transformation was not documented, and the meta-regressions show that the
estimator itself explains little. Second, the field would benefit from reporting implicit prices in a common unit (percent per
square metre, per kilometre, per year) alongside the raw coefficient, which would make the harmonisation of Section~\ref{sec:extraction}
unnecessary. Third, the AI literature should report hedonic benchmarks on the same test set, out-of-time validation, and the
incremental contribution of AI-derived features, which only a minority of studies currently do; the evidence that such features
add interpretable characteristics (condition, view, layout) is more promising than the evidence that they add accuracy.