Key Takeaways
What the parcel record actually looks like, what the gap costs, and where AI stops helping.
- The National Academies found there is no single national program for parcel data in the United States, and set out a vision for one in 2007 that still does not exist.
- Parcel records are held county by county across 3,144 counties and county-equivalents, each with its own format and its own level of digitisation.
- The cost of that gap is mispricing, not slow sales. NAR’s land survey finds most land sells in under 60 days, roughly in line with housing.
- AI closes the aggregation gap by assembling scattered public records into one view of a parcel in seconds.
- It cannot close the coverage gap, and a model trained on a thin, historically mispriced market will repeat that market’s errors with confidence.
Ask a simple question about a vacant parcel, what is it worth, and the answer is unexpectedly hard to get. No clean comparables, a county assessment that was never built to price a sale, and a handful of listing sites that each show a slice of the market.
That is not bad luck. It is a structural gap in how land data is collected in the United States, and it is the reason land trades at the wrong price more often than any other major asset class. This sets out what the record actually looks like, what the gap costs, what AI has genuinely closed, and what it has not.
Call it the land data gap. For the methods that work around it today, the pillar guide treats comparable sales as the single most reliable way to reach a number.
Quick verdict: the fragmentation is real and documented, but the usual conclusion drawn from it is not. Land is not measurably slower to sell than housing. It is measurably harder to price, and those are different problems with different fixes. AI has closed most of the aggregation gap and none of the coverage gap, so an automated estimate is worth exactly as much as the records behind that particular county.
What Does the Parcel Record Actually Look Like?
Fragmented by design, and documented as such by the National Academies nearly twenty years ago.
Housing has an MLS, national portals and decades of clean transaction data. Land has none of that. Ownership and sale records are held at county level, and USDA’s county typology documentation counts 3,144 counties and county-equivalent geographies in the United States. Each maintains its own records, in its own format, to its own standard.
That is not an impression. In National Land Parcel Data: A Vision for the Future, a National Research Council committee examined the question directly and found that “there is not a single dedicated program for development of nationwide parcel data within the United States.” The same report identified “a major digital divide in terms of parcel data” across the country, and recorded 2,925 counties acting as the primary responsible entity for collecting and managing it.
That report is from 2007, and its specific digitisation percentages are far too old to quote as current. The structural finding is what matters, along with what happened next: a national body laid out a vision for national parcel data, and nearly two decades later no national parcel database exists.
The consequence is the one economists describe as information asymmetry. When one side of a transaction can see the record and the other cannot, the informed side does not merely win the negotiation. It sets the price.
Does the Data Gap Actually Make Land Slow to Sell?
No, and this is where most versions of this argument, including earlier versions of this page, get it wrong.
The Federal Reserve’s FRED series for median days on market, sourced from Realtor.com listing data, reads 60 days as of August 2026. That is the housing benchmark.
Now the land side. NAR’s REALTORS Land Market Survey is the only national land transaction dataset that exists, and it reports that “Most land is purchased in under 60 days,” with “a notable 25% of these transactions are wrapped up in less than 30 days.” In six of its regions, sales close within 45 days.
Sixty days against under sixty days. The dramatic gap that gets asserted in commentary about land is not in the data.
Two honest caveats. The NAR figures come from REALTOR-represented transactions, which skew toward brokered, marketable parcels, and the survey is not annual in the way housing statistics are. So it undercounts exactly the unbrokered, remote, hard-to-market parcel that takes eighteen months to move.
But notice what that caveat is. It is a data coverage problem, not a speed problem. The parcels that genuinely sit are the ones no dataset sees, which is the thesis restated rather than refuted. Where a parcel does have a narrow buyer pool, that is a demand fact about that parcel, not a national property of the asset class.
So What Does the Data Gap Actually Cost?
Mispricing. A market can clear quickly and clear at the wrong number, and that is what thin comparable data produces.
The mechanism is straightforward. A house sits among dozens of near-identical neighbours that sold recently, so a valuation has a dense evidence base. A five-acre parcel may have no genuine comparable within miles and no recorded sale for a decade. Everything that follows from that, the wide bid-ask spread, the seller anchoring to a tax figure, the buyer with better records naming the number, traces to the same absence.
The tax figure is worth singling out because it is the most common substitute for real evidence. An assessed value is the product of an assessment ratio applied by a county for taxation, on a reassessment cycle, under standards that formally tolerate more dispersion on vacant land than on housing. It answers a different question, and a seller who reads it as a price signal is not making a small error.
The scale is not trivial. NAR’s survey puts the value of land held by households at $18.6 trillion, roughly 41 percent of total household real estate assets, as of the third quarter of 2023. That is a large fraction of American wealth priced with the worst data of any major asset class.
What Has AI Actually Closed?
The aggregation problem, which was always the real bottleneck.
The raw material existed all along, sitting in county and federal databases. What was missing was any practical way to collect, clean and connect it at parcel level. That is a data-engineering problem before it is an intelligence problem, and it is the part that has genuinely been solved.
A modern system pulls parcel boundaries and ownership from county records, overlays spatial layers for flood zone, terrain and zoning, and reaches federal datasets that were previously effectively invisible to a private buyer. USDA’s Web Soil Survey is the clearest example: soil class and agricultural capability for any parcel in the country, free, and consulted by almost nobody buying land by hand.
Work that required a title company, a surveyor and several days now assembles in seconds. That does not replace due diligence. It changes the starting position from a blank page to a real picture, which is a bigger shift than it sounds.
The valuation layer follows from the aggregation layer. With enough parcels observed, a model can learn how acreage, access, zoning and location relate to price across a region and apply that to a parcel with no direct comparable, which is precisely the case a traditional appraisal handles worst. A defensible range beats the shrug that used to be the honest answer.
What Can AI Still Not Fix?
Coverage, and its own inherited errors. Both limits are worth naming precisely, because vague ones get ignored.
Coverage. A model reaches only the records that exist in machine-readable form. The National Academies documented a digital divide in parcel data, and while digitisation has advanced substantially since 2007, it has not advanced evenly. Estimates are weakest in remote counties with few recorded sales, which is exactly where land is cheapest and where buyers most need help. As the AI valuation explainer puts it, data thins out precisely where land gets most interesting.
Inherited error. This is the limit almost nobody names. A model trained on comparable sales learns the pricing behaviour of that market, including its mistakes. If a county’s land sales were themselves mispriced by years of opacity, a system trained on them reproduces the error and reports it with confidence. Better data narrows this over time. In a thin market today, an automated estimate is a hypothesis to test, not a fact to rely on.
And the ground itself. No model walks the access road in March, checks whether the boundary is disputed, or runs a perc test. Those decide real value and none of them are in the record.
The correct use is as a first screen that tells you where to spend verification effort, and the correct output is a range with a confidence attached, not a number with a decimal point.
What Should a Buyer or Seller Do About It Now?
Use the information advantage, and know which parts of it are load-bearing.
For a seller: start from a data-driven estimate rather than the tax card, then document the things that decide value and that no dataset contains. Legal access, zoning, utility availability, survey status. Every one of those you answer in advance is a question a buyer would otherwise discount for.
For a buyer: use automated estimates to filter and to sanity-check an asking price, then verify physically. The estimate tells you which parcels deserve a site visit. It does not tell you which to buy.
For either side, the method that resolves a disagreement is the same one it has always been. Pull genuinely comparable sales, adjust honestly for access, zoning, topography and size, and reconcile to a final value you can defend with evidence. AI has made assembling that evidence fast. It has not changed what counts as evidence.
Where Does This Go From Here?
Toward better coverage, unevenly, and slowest in exactly the places that need it most.
Digitisation continues county by county. More recorded sales feed the models. Estimates get sharper where transaction density is highest, which is the suburban fringe and the growth corridors, and improve slowest in the remote rural counties that make up most of the country’s acreage.
The defensible prediction is narrow: parcel-level estimates will become a standard feature of land listings rather than a specialist tool, and the confidence attached to them will diverge sharply by county. The gap will not close evenly, and anyone promising otherwise is selling something.
What changes for individuals is already true. The information edge that belonged to people with county access and professional relationships is now broadly available, and the cost of not using it is paid in the price. You can see what comparable parcels are asking across vacant land for sale by acreage and location, which is the closest live read on demand available without pulling records yourself.
Once a number is built on that evidence rather than on the tax roll, you can put a parcel in front of buyers and let the market test it.
The land data gap was never a permanent feature of the market. It was an accident of administrative history, a market that grew up without the infrastructure housing was given. That accident is being corrected, slowly, and unevenly, and mostly by private aggregation rather than by the national program that was recommended in 2007.
Frequently Asked Questions
Why is vacant land so hard to price?
Because there is no national parcel data system. Records are held across 3,144 counties, each in its own format and at its own level of digitisation, and the National Academies found no single dedicated national program for parcel data. Comparable sales are thin because parcels are unique and trade rarely, and county assessments answer a tax question rather than a price question.
Does land actually take longer to sell than houses?
Not according to the only national land dataset. NAR’s REALTORS Land Market Survey reports most land is purchased in under 60 days, with 25 percent inside 30 days, against a FRED housing median of 60 days in August 2026. The survey covers REALTOR-represented sales, so it undercounts unbrokered remote parcels that genuinely sit.
Is AI land valuation accurate enough to trust?
As a first screen, yes. As a verdict, no. Where transaction density is high and records are digitised, estimates land in a useful range. In thin rural markets uncertainty rises sharply, and a model trained on historically mispriced sales will repeat those errors confidently. Treat any automated figure as a hypothesis to test against local comparables.
Does AI replace appraisers and due diligence for land?
No. AI aggregates records and produces fast estimates. It cannot inspect a parcel, verify legal access, or take professional responsibility for a value. Appraisals remain necessary for loans, disputes and high-value transactions, and physical due diligence remains necessary everywhere, because access, drainage and boundary problems are not in any dataset.
Resources and Further Reading
- National Land Parcel Data: A Vision for the Future is the National Academies study that found no single national program for parcel data in the United States.
- USDA ERS county typology documentation gives the count of counties and county-equivalent geographies.
- FRED, Housing Inventory: Median Days on Market tracks the housing benchmark used here, sourced from Realtor.com listing data.
- REALTORS Land Market Survey is the only national survey of land transaction timing and value.
- USDA Web Soil Survey is the public soil dataset that automated parcel profiling draws on.
