% ============================================================================
\section{Related Literature}
\label{sec:literature}
% ============================================================================

This paper connects four strands of research.

The first is the classical \textbf{hedonic pricing} literature. Building on
\citet{lancaster1966new}, \citet{rosen1974hedonic} formalized the market for
differentiated goods as a market for their underlying characteristics, providing the
theoretical foundation for interpreting regression coefficients as implicit prices.
The empirical housing literature that followed is vast; \citet{sirmans2005composition}
survey the characteristics most commonly included in hedonic house-price models and
document substantial heterogeneity in specification, \citet{malpezzi2003hedonic}
provides a critical review of functional-form and identification choices, and
\citet{hill2013hedonic} surveys the use of hedonic methods for house-price
measurement. A recurring theme is omitted-variable bias from unmeasured quality:
the state of repair, the standard of finishes, and the visual quality of a dwelling
are priced by the market but rarely recorded in the structured fields available to
the econometrician. \citet{nowak2020quality} show that this omission is not
innocuous even for aggregate statistics, since quality-adjusted repeat-sales indexes
built with information from listing remarks differ materially from conventional
ones. Our contribution speaks directly to this measurement problem. Our log-linear
specification, with logged continuous attributes and a dependent variable in logs,
follows the mainstream of this literature, and we read indicator coefficients with
the \citet{halvorsen1980interpretation} transformation; the small-sample refinement
of \citet{kennedy1981estimation} is immaterial at our precision.

The second strand concerns \textbf{spatial structure, inference, and prediction}.
\citet{can1992spatial} and \citet{anselin1988spatial} emphasize that unobserved
locational amenities induce spatial dependence that, if ignored, biases coefficient
estimates and understates uncertainty; we absorb the first-order component of this
problem through municipality fixed effects, re-estimate with finer forward
sortation area (RTA) effects, and cluster inference at the level of the fixed
effects \citep{cameron2015cluster,abadie2023when}, complementing the analytical
standard errors with a cluster bootstrap because our clusters are of very unequal
size \citep{mackinnon2017wild}. The same dependence bears on validation.
\citet{bourassa2010predicting} show that spatial structure that is not exploited by
the model surfaces in prediction errors, \citet{pace2020examining} find that tree
ensembles recover information left in the residuals of hedonic and spatial models,
and \citet{roberts2017cross} document that random cross-validation is optimistic
whenever observations are spatially clustered, recommending blocked (spatially
grouped) folds instead. We report both. On the prediction side, \citet{bin2004prediction}
compares parametric and semi-parametric hedonic predictors, \citet{mullainathan2017machine}
frame machine-learning methods as estimators of the conditional expectation
$\mathbb{E}[P\mid\mathbf{z}]$ suited to prediction tasks, and \citet{kok2017bigdata}
document the rise of automated valuation models (AVMs) in industry. This literature
generally finds that flexible learners beat linear hedonics on accuracy at some cost
in interpretability; our estimator ladder (Section~\ref{sec:benchmark}) reproduces
that pattern on a common feature set, and our focus is on how much of the gap is
closed not by a better \emph{estimator} but by better \emph{measurement}.

The third strand is the use of \textbf{unstructured listing content}---text and
images---in real-estate valuation. \citet{gentzkow2019text} survey text-as-data
methods in economics, and \citet{dell2025deep} and \citet{korinek2023generative}
survey, respectively, deep learning and generative models as research tools for
economists. In real estate specifically, \citet{nowak2017textual} show that words in
MLS remarks predict transaction prices and reduce hedonic pricing errors;
\citet{nowak2021dictionaries} show that the price-relevant vocabulary is not stable
across markets or over time, a warning against transferring word lists; and
\citet{shen2021information} extract quality signals from property descriptions with
supervised machine learning, finding that the text carries information about
unobserved quality that is capitalized into prices. \citet{zhang2024describe}
benchmark text encoders for price prediction and find that textual and
non-textual features are complementary, and a companion working paper on Quebec
listings \citep{boucher2026decoding} documents that semantic embeddings of
descriptions carry out-of-sample price information and that the language of the
description is a first-order confounder in a bilingual market---a finding we
build on here. On the image side, \citet{ahmed2016house} combine visual and
textual features for price estimation; \citet{bessinger2016quantifying} and
\citet{poursaeed2018vision} show that learned measures of exterior and interior
quality carry independent predictive power; \citet{naik2017computer} and
\citet{glaeser2018vision} use street-level imagery to measure the physical
appearance of neighbourhoods and houses and study how appearance relates to
incentives and change; \citet{law2019take} and \citet{kang2021understanding}
combine street-view and satellite imagery with structured attributes to predict
prices and appreciation; \citet{zhang2018greenery} price street-visible greenery
from street-view images; \citet{johnson2020curb} construct a curb-appeal index from
Google Street View images and estimate its capitalization into sale prices; and
\citet{lindenthal2021architecture} classify architectural styles from photographs
and document style premia. Relative to this work, our approach differs in the
\emph{technology of measurement}: rather than training bespoke supervised models on
labelled data---which requires labels and task-specific engineering, and yields
opaque scores---we use general-purpose large language and vision--language models
with schema-constrained decoding to produce a fixed set of interpretable,
human-readable attributes (``renovated kitchen,'' ``brick façade,'' a 1--10
curb-appeal score) at a marginal cost of a fraction of a cent per listing, and we
embed those attributes in a transparent hedonic regression whose coefficients can be
read as implicit prices. The reliability of such zero-shot annotation is itself an
active research question; \citet{gilardi2023chatgpt} find that general-purpose LLMs
match or exceed crowd workers on standard text-annotation tasks, but the extraction
error of any particular field must be assessed, which we do by confronting the
machine-generated attributes with independent sources
(Section~\ref{subsec:validation}).

The fourth strand is the \textbf{microstructure of listing markets}, which frames two
institutional features of our data. Our sample pools a brokered MLS platform
(Realtor.ca) with a for-sale-by-owner platform (DuProprio); \citet{hendel2009mls}
compare exactly these two market forms and document differences in seller
composition, marketing, and outcomes, motivating our platform indicator and
platform-specific robustness checks. And because we observe asking rather than
transaction prices, the interpretation of our dependent variable rests on the
list-price literature: the asking price is a strategic choice that anchors search and
bargaining \citep{horowitz1992list,knight2002listing,han2015microstructure}, and
\citet{han2016asking} show that it acts simultaneously as a commitment device and
as a signal in a directed-search environment. We return to this limitation in
Section~\ref{subsec:limits}.