Real estate generates enormous amounts of information that valuation models have never been able to harness. Listing descriptions, property photographs, inspection reports, lease documents, floor plans, street imagery, and market commentary all describe assets in detail, and almost none of it has historically made it into a valuation model. What made it in were the fields. Square footage, floorplate, lot size, year built, transaction history, these can all be put in a column and compared across properties.

An appraiser reading a listing extracts information about condition, layout quality, and neighborhood character that affects price. A model reading the same listing extracted nothing, because there was no reliable way to turn prose into a variable. The text got discarded and the model worked with what was left.

That is beginning to change, and the early results suggest the discarded material was worth considerably more than the industry assumed. A study published in the Journal of Real Estate Research in December 2025 used GPT models to extract quantitative features from residential property advertisements across 9,842 transactions in Oslo, then fed those extracted features into an XGBoost model alongside the structured data from the same listings. The combination improved root mean squared error by 24.31% and mean absolute percentage error by 15.25% over structured data alone.

The magnitude matters given how mature AVM methodology has become. Gradient boosting approaches have been refined for years against the same categories of input, and incremental gains have gotten progressively harder to find. An error reduction of that size coming from a new data source rather than a better algorithm indicates how much predictive signal has been sitting unused in documents nobody could read.

Earlier attempts had established the premise without capturing nearly as much value. A 2022 study using 30,218 rental offers from Berlin and 32,610 house offers from Los Angeles applied BERT embeddings to property descriptions and found error reductions of 17.09% on rentals and 5.66% on house prices. The text clearly carried information. The question was how to get at it.

The newer approach differs in method, and the difference explains the gap in results. Embedding converts text into numerical representations a model can process without ever identifying what the text actually says. Using a language model to extract specific features produces something else entirely, a set of discrete variables about renovation status, view, noise, layout, or condition. Those variables behave like structured data because they are structured data, just derived from a source that used to be illegible.

That distinction also addresses one of the persistent objections to machine learning in valuation. A model that improves because of embedded text vectors cannot explain what changed. One that improves because it now knows a property has a renovated kitchen and southern exposure can show its work. Where valuations get challenged by lenders, appraisers, and owners, being able to identify which features drove a number is not a minor consideration.

Text is the most tractable of the unstructured sources, which suggests the current results are an early version of what becomes possible. Images are the obvious next input, and some of that research is already underway. A study published in PLOS One in 2025 used Hong Kong as a test case for fusing exterior housing photos, street view imagery, and remote sensing data into a machine learning valuation framework. Street view captures things no listing mentions and no field records, including streetscape quality, the condition of neighboring buildings, tree cover, and the general character of a block.

Floor plans are another candidate. Two properties with identical square footage can have very different layouts, and layout affects usability in ways that price reflects but models cannot currently see. A model capable of reading a floor plan could distinguish between a well-organized 1,200 square feet and an awkward one.

Commercial valuation depends on unstructured context to an even greater degree, which is what makes the application there potentially more valuable and considerably harder. Lease documents, tenant credit information, property condition assessments, environmental reports, and market commentary all contain material information stored as prose. The same extraction approach that works on a residential listing could work on a lease abstract or an engineering report, and the resulting features would feed the same kind of model.

The limits are worth stating plainly. The Oslo study covers residential transactions in a single market, where listing conventions and disclosure practices may not resemble other markets. Property advertisements are also written to sell, which means they emphasize favorable attributes and omit unfavorable ones systematically. A model trained on promotional text inherits that bias, and whether extraction can correct for it is unresolved.

Access is the other constraint, and it cuts differently across property types. Residential listing text is abundant and largely public. Commercial lease documents and condition reports are neither, which means the organizations best positioned to apply these methods to commercial assets are the ones already sitting on large proprietary document archives.

What the research establishes is that the boundary of usable valuation data has moved, and the mechanism that moved it is not specific to listing text. Models built on structured fields were never limited by what practitioners understood to be relevant. They were limited by what could be read. As that constraint loosens, the competitive question shifts from what a model can process toward what an organization can actually get access to.

The post LLMs Are Pulling Value Out of Property Descriptions appeared first on Propmodo.

​