Did AgriTerra really publish citations that were wrong?

Yes, and we corrected them. A self-audit in 2026 found several source attributions in the renewable-energy and cost sections that did not hold up, including references to lease-payment tables that do not exist in the reports they named. The underlying numbers were mostly in a defensible range, but the footnotes would not have survived scrutiny. We re-sourced every one of them from named primary documents and built checks so it cannot recur.

How is that different from what other land tools do?

Most land tools never show their sources at all, so a bad citation is invisible. Our whole promise is that every figure has a documented origin, which is exactly why we hold ourselves to opening the document and logging it. The differentiator is not that we never made a mistake. It is that we can tell you precisely which numbers we corrected and why.

What stops the AI writing step from making up a number?

The numbers are computed from the datasets before any sentence is written. The writing step can only reuse those computed numbers, every figure it produces is checked back against the computed value, and anything it cannot match renders as data unavailable rather than a guess. A section that cannot be produced on spec falls back to a fixed human-written template and the report is marked partial.

Where can I see the actual sources behind my report?

The full dataset-by-dataset account, with who publishes each source and how fresh it is, is on the methodology page at how we get our numbers. Every figure in a report also prints its own source line, for example "Source: USDA SSURGO, accessed 2026-06-03".

How we work

How we check our numbers, including the ones we got wrong

August 19, 2026 · AgriTerra

Our whole promise is a plain one: every figure in an AgriTerra report traces to a named public source and the date we read it. The method behind that, dataset by dataset, is written out on how we get our numbers. This page is the other half, the part most companies never put in writing. It is the story of how we keep the sourcing honest, and it starts with the time we caught ourselves getting it wrong.

We audited our own numbers, and did not like what we found

In the summer of 2026 we ran a hard audit of the data layer behind our reports: not a spot check, but a line-by-line pass through every figure and every source attached to it. Most of it held. The soil, flood, wetland, water, and cash-rent figures read straight from federal datasets and checked out. But when we opened the documents behind the renewable-energy and cost sections, several of the citations did not exist.

The specifics, so you can see exactly what we mean:

Our solar lease bands cited a lease-payment table in a national solar market report. That report has no such table and no landowner lease content at all.
Our wind lease bands cited two tables in a national wind market report. That report contains three tables, and neither of the two we named was among them.
Our recreational-lease rates cited an annual hunting-lease guide from a well-known outdoor brand. No such annual guide is published.
A crop-cost figure was attributed to a USDA source as a state average. That agency publishes cost of production by multi-state farm resource region and has never published a state-by-state series, so the attribution could not be right.

Here is the uncomfortable part. The numbers themselves were mostly in a defensible range. What failed was the paperwork under them. These attributions had the signature of confident text that was never checked against the actual documents, which is the exact failure mode a language model produces when it is allowed to write a citation instead of read one. For a company whose entire pitch is that every number has a documented origin, that is the most dangerous thing that can be in the product. The numbers could be fine and the footnotes could still not survive ten minutes with a land-grant economist, a journalist, or a lender's credit committee.

We re-sourced them. We did not re-invent them.

The wrong move would have been to quietly swap in a better-looking citation for the same number. We did the slower thing. We went back to primary documents, re-derived each band from sources we could name and open, and let the number move if the honest source said it should.

The solar and wind lease bands are now AgriTerra estimates anchored to named surveys and to per-megawatt payment math, not mirrored from any one rate card. As an example of where that lands, an Ohio mid-bucket parcel now carries a modeled solar lease band of about $800 to $1,225 per acre per year, midpoint $1,000, anchored to the Purdue Ag Economy Barometer solar-offer surveys and Ohio State University Extension's solar-leasing guidance, as of 2026-08-01. The recreational-lease figures are now labeled as market indications from university extension and brokerage sources. The crop-cost figure is re-anchored to university crop budgets and relabeled as a non-land production cost, which is what it actually measures. Every one of these is now the kind of citation a reader can follow to a real page.

We are telling you this on purpose. After the fix, "here is exactly where every number comes from, including the ones we corrected" is a stronger position than pretending the first version was perfect, and it is one almost no competitor can honestly take.

The machinery that makes it not happen again

A promise to be careful is worth very little. So the checks below are not intentions, they are steps that run in the build. If any of them fails, the release does not go out.

A person opens the document, and it is logged

For every dataset edition, someone opens the actual cited source and records that they did, with the edition and the date, in an append-only log. A citation becomes a thing that happened, not a string that got typed. That single habit is what the fabricated footnotes skipped.

The build refuses to ship an unverifiable citation

A check named npm run check:citations runs over every citation in the data layer and fails the build if a cited source cannot be reached or has no verification record. It also carries a tombstone list: every citation we removed as unverifiable is permanently blocked from coming back, so a fabrication cannot be reintroduced by accident, by a person or by a model.

Every dataset carries a machine-readable read date

Each source records the date it was last read, in a form the build can check. If a dataset goes stale relative to the code that depends on it, a freshness check catches it before release. That is the same date printed on the figure in your report and in the refresh column of the methodology page.

The writing step cannot invent or move a number

The numbers are computed from the datasets first. The step that writes the sentences can only reuse those computed values, and every figure it emits is checked back against the computed number before the report ships. Where a value is missing, the report says data unavailable instead of guessing, and a section that cannot be produced on spec falls back to a fixed, human-written template while the report is marked partial. The prose can describe the numbers. It cannot change them.

Two of those deserve a plain-language note. The tombstone list in check two is the direct answer to what we found: the fabricated citations are not just deleted, they are recorded as forbidden, so no future edit can bring them back. And check four is why the AI in the product cannot do the thing that caused the problem in the first place. It writes the sentences around numbers that were already computed and checked; it never gets to originate a figure or a source. Every dataset edition must have a verification record confirming a human opened the cited document; the tombstone gate blocks any purged fabrication from returning.

The same discipline, pointed at what a lease can actually earn

Checking a citation is one kind of honesty. Refusing to show income that cannot be collected is another, and it runs on the same principle. Since the 2025 federal energy law, a good solar or wind resource on paper is not enough on its own, so every energy scenario runs through a policy gate: county siting restrictions, the state posture, and the federal tax-credit vintage. Where a developer would not actually offer the income today, we demote it rather than print a number you cannot bank. As of 2026-08-02 that gate covers Ohio, Indiana, Iowa. The long version is its own piece: why we will tell you a lease won't happen on your land, alongside what the 2025 federal energy law means for your farmland.

What we deliberately do not do

Being able to check a number is only half the job. The other half is not claiming more than the data supports. So, on purpose:

No land valuations. We assemble and source the data. We do not tell you what your land is worth, and a Decision Report is a sourced parcel read, not an appraisal or an opinion of value.
No guaranteed-income claims. A band is what the market has paid, not a promise of what you will be offered. Estimates are labeled as estimates.
No treating queue presence as a promise. A project sitting in an interconnection queue is a signal, not a guarantee it gets built, and we say so.
No single-number precision theater. We show a low, a midpoint, and a high, because a lone number implies a certainty the data does not have.

Six federal datasets, each with a name over the door

Under all of it sit 6 federal datasets, and we name every one with its agency parentage rather than waving at "government data": USDA NASS, USDA NRCS SSURGO, NREL, FEMA, USGS, USFWS. Around them sit the lease and market bands we model from named surveys. That is the whole point of the exercise. A source with an agency and a read date is one you can open, argue with, and hold us to. A vague one is not, and the summer of 2026 is why we treat the difference as seriously as we do.

Go deeper: the full methodology, dataset by dataset, how the policy gate works, the landowner guides on solar leasing and wind lease rates, and how AgriTerra compares to AcreValue and AcreTrader.

Challenge any number in your report

Pin your parcel for a free instant read, each figure with its source and the date we read it. Point to a better source and we would rather hear it from you than from the person across the table.

Launch Scout