% ============================================================================
\section{Conclusion}
\label{sec:conclusion}
% ============================================================================

We have estimated a hedonic price model for Quebec houses that augments the standard
set of structural, locational, and amenity attributes with two machine-generated
feature layers: structured specifications extracted from free-text descriptions by a
large language model, and visual attributes extracted from façade photographs by a
vision--language model. Schema-constrained decoding renders these outputs as
ordinary, auditable econometric variables, obtained at a marginal cost of a fraction
of a cent per listing. The central message is methodological: general-purpose AI
models, disciplined by strict output schemas, function as a low-cost
feature-engineering technology that measurably expands the characteristics set
available to hedonic analysis---and the discipline must extend to the input, since
the text handed to the model may contain the very price one is trying to explain.

The gains are economically meaningful but unequal across the two layers. Together
they add 3.8 points of in-sample $R^2$ to an already rich log-linear specification
(0.717 to 0.755 with living area observed), of which 3.0 come from the façade
photograph and 0.8 from the text; they raise the out-of-sample accuracy of every
estimator in our ladder, from OLS to gradient boosting and an embedding-based
residual neural network; and the best blend attains an out-of-sample $R^2$ of 0.779
with a median absolute percentage error of 14.0\% under random folds, 0.734 for
gradient boosting when entire postal areas are held out. The vision layer adds more
to the linear model than switching from OLS to a random forest does. The extracted
attributes carry interpretable implicit prices---curb appeal, renovation level,
luxury cues, and exterior materials all price sensibly---that survive
municipality-clustered inference, a cluster bootstrap, finer postal-code fixed
effects, and a control for the language of the description. The text layer's
contribution, by contrast, is modest once keyword amenities are already in the model,
and part of its measured gain on the for-sale-by-owner platform reflects the asking
price embedded in those descriptions rather than information about the house.

Three extensions are natural. First, the same schema-constrained approach scales
from a single façade image to a small panel of interior photographs (kitchen,
bathroom, living space), which our design already accommodates; a repeat extraction
of a subsample at a fixed temperature, and a human-labelled validation set, would
also quantify the extraction error that we could only measure indirectly. Second,
the DuProprio coordinates and the postal codes of all listings permit continuous
spatial features---distance to the urban core, to water, or to amenities---that would
refine the geography beyond discrete area effects. Third, and most important for
external validity, linking listings to registry transaction prices would establish
how much of the AI-measured quality premium survives bargaining. For appraisal,
automated valuation, and property-tax assessment---applications that consume hedonic
models at scale---the practical implication is that the unstructured majority of a
listing's information content is no longer out of reach, provided that it is read
with the same care as any other regressor.