An AI market report is only as trustworthy as the provenance of the numbers inside it, and the whole problem is that AI writes every number in the same confident voice. A rent comp from a recorded transaction, a vacancy figure estimated from a survey, and a rent-growth projection a model invented all arrive in the same clean paragraph, with no visible seam between the fact and the guess. For a principal at a small commercial real estate firm with no data analyst to vet the output, the skill that matters in 2026 is not generating these reports — any associate can do that in a minute — but reading one and knowing which sentences you can take to a lender and which are decoration. This is how to sort the signal from the filler.
What an AI market report is actually made of
Every number in a commercial real estate market report comes from one of three places, and knowing which one you are reading is the entire skill. The three provenances are a recorded transaction, a survey estimate, and a model’s projection — and they carry wildly different weight.
A transaction-derived number is the strongest: an actual sale, a signed lease, a recorded price. This is what data providers like CoStar, Crexi, CoreLogic, and Altus Group are built to collect and verify, and it is the closest a market report gets to fact. A survey estimate is softer — a vacancy or asking-rent figure assembled from what owners and brokers reported, useful for direction but subject to who answered and when. A projection is the weakest: a rent-growth or cap-rate forecast a model produced, which is a hypothesis wearing the costume of a data point.
The trouble with an AI-generated report is that it dissolves these three into one unbroken narrative. A general assistant summarizing the open web will write “office vacancy is 14.2% and rents are projected to grow 3% next year” without telling you the first figure is a survey estimate of unknown vintage and the second is a guess. The report reads as if every sentence is equally solid, and your job as the reader is to put the seams back in.
This is not an argument against AI reports but for reading them like a broker’s pro forma — assuming optimism until you find the source. The same discipline that separates broker projections from trailing actuals in an underwriting model applies here, and we walk through that reconciliation habit in the anatomy of an AI-assisted underwriting workflow.
The signal test: two questions per claim
Signal is any claim you can trace and check; filler is any claim you cannot. Two questions sort almost every sentence in a market report into one bucket or the other.
Question one: is it sourced to something I can open? A signal-grade claim names its origin — “per CoStar Q2 sale comps,” “based on the property’s T-12,” “from the county recorder” — specifically enough that you could pull the same record. A filler-grade claim floats free: “market rents are trending upward,” “the submarket is seeing strong demand,” “cap rates have compressed.” If you cannot follow the sentence back to a document, treat it as atmosphere, not evidence.
Question two: is it dated? Commercial real estate moves faster than most reports admit. Federal Reserve commercial-property price data has tracked value corrections exceeding 20% in the hardest-hit asset classes inside 24-month windows, which means a “current” vacancy or cap-rate figure with no vintage attached might be describing a market that no longer exists. A dated claim — “as of June 2026” — can be judged; an undated one is a coin flip.
A claim that passes both questions is signal: sourced and dated, you can act on it or verify it in one step. A claim that fails both is filler — the report’s connective tissue, not its content. Most AI market reports are a thin layer of the first wrapped in a thick layer of the second, and the confident prose hides the ratio until you apply the test deliberately.
The five filler tells
Once you know to look, filler announces itself. Five patterns recur in nearly every AI-generated market report.
- Round-number confidence. Real transaction data is lumpy and specific — a 4.9% cap rate, $28.50 per square foot. Filler rounds: “roughly 5%,” “around $30.” Suspiciously clean numbers often mean the model is estimating, not reporting.
- Unsourced projections. Any forward-looking claim — rent growth, absorption, cap-rate movement — is a model’s opinion unless it cites a specific forecaster and methodology. Treat every undated projection as the report’s least reliable sentence.
- Generic macro boilerplate. Paragraphs about interest rates, “economic uncertainty,” or “shifting demand drivers” that would read identically for any submarket in the country are pure filler — the model had little local data and padded the gap.
- Uncheckable citations. A report that references “CBRE’s latest market report” or “recent JLL data” without a title, date, or link is citing a vibe, not a source.
- Disagreement between two runs. Generate the same report twice. Where the versions agree on a specific number, there is probably a real source; where they diverge — one says 12% vacancy, the other 16% — the model is guessing, and neither number is signal.
That last tell is the most useful and least used: running a report twice costs a minute and exposes which figures the model actually knows versus which it fabricates fresh each time.
The fabricated-citation trap
The single most dangerous filler pattern is the citation you cannot find, because it looks exactly like signal. This is not hypothetical. In June 2026, a widely covered incident saw a major consulting firm publish an AI-assisted report in which, by one audit, only five of forty-five citations correctly pointed to their stated source — the rest were mangled, vague, or fabricated. Observers named the pattern “vibe citing”: the model stitches together fragments of real sources and invents references that look convincing until someone clicks them.
A CRE market report has the identical failure mode. An AI summary asserting “office absorption rose per Cushman & Wakefield’s Q2 report” is worthless if that report says no such thing, does not exist, or covers a different metro. Named sources are signal only when the source is real and says what the report claims.
The defense is a spot check: pick the one or two numbers you would actually stake a decision on and trace each to its cited source. If the source opens and confirms the figure, the claim is signal. If it cannot be found, does not contain the number, or turns out to be a homepage rather than the specific dataset, discard the claim and grow suspicious of the rest. Blindly trusting fluent, well-cited-looking output is exactly how unverified AI work produces expensive mistakes, a pattern we examine in why most underwriting automations fail.
Reading a report in ten minutes
You do not need a data team to grade an AI market report — you need a repeatable pass that takes about ten minutes.
First, separate the numbers from the narrative. Ignore the prose on the first read and pull out every specific figure — vacancy, rents, cap rates, absorption, comps. The narrative is where filler hides; the numbers are what you can test.
Second, run the two-question signal test on each figure. Sourced and dated stays; unsourced or undated goes into a “verify or discard” pile. Most reports lose half their numbers here, which is fine — you keep the half that means something.
Third, spot-check the two figures that matter most. For the numbers a lender, investor, or committee memo would rely on, open the cited source and confirm. If a figure has no citable source, verify it against a transaction-data provider or leave it out of anything you send onward.
Fourth, note what is missing. AI reports are strong on easily summarized metrics and weak on the local color that decides deals — a competing project two blocks away, a tenant that just gave notice, a zoning change in committee. If the report reads clean but tells you nothing a resident of the submarket would know, it is a starting point, not an analysis. Where this screening feeds a repeatable deal process, the deal-screening framework for a lean team shows how to make it routine rather than ad hoc.
What a small firm should and should not use these for
AI market reports are genuinely useful, but only for the jobs their reliability supports — matching the tool to the task is what keeps a small firm out of trouble.
Use them for speed on low-stakes screening. When a broker blast lands and you need to know in two minutes whether a submarket is worth a real look, an AI report is a fine first filter — you are deciding whether to spend more time, not committing capital. Use them to draft the boilerplate sections of your own market write-ups, which you then verify and localize, and use them to surface questions: a figure that surprises you is worth chasing to its source whether or not the report got it right.
Do not use them, unverified, in anything a counterparty relies on. A number that flows into a lender package, an investor update, an offering memorandum, or an investment-committee memo has to be traced to a real source first — the report is a draft of that number, not the number itself. And do not treat an AI summary of the open web as equivalent to a transaction-data subscription: when the decision is large enough, licensed comps from a provider like CoStar or Crexi are worth their cost precisely because their provenance is known, a trade-off we cover in the broader CRE deal-analysis playbook.
The firms that get the most out of these tools are not the ones that trust them most, but the ones fluent enough to read a report critically in minutes and route each claim to the right level of scrutiny — the kind of practical AI fluency that lets a lean shop out-operate a larger competitor, the argument at the heart of the small CRE firm AI manifesto.
FAQ
Are AI-generated CRE market reports accurate?
They are as accurate as the data underneath them, which varies enormously. A tool sitting on verified transaction data can be reliable for the metrics that data covers; a general assistant summarizing the open web produces a mix of real figures, survey estimates, and outright projections in the same confident tone. Accuracy is not a property of “AI” but of provenance, and the reader has to establish it because the report will not.
How can I tell if a market report’s numbers are real or estimated?
Ask two questions of each figure: is it sourced to something you can open, and is it dated? Transaction-derived numbers tend to be specific and citable; estimates tend to be round and vague. A precise, dated, sourced figure (“4.9% cap rate per Q2 sale comps”) is likely real; a clean, undated, unsourced one (“cap rates around 5%”) is likely estimated. When in doubt, generate the report twice — figures that change between runs are being guessed.
What does “signal versus filler” mean in an AI market report?
Signal is any claim you can trace and verify — sourced, dated, and specific. Filler is the connective narrative you cannot check: unsourced trends, undated projections, and generic macro commentary that would read the same for any market. Most AI reports are a thin layer of signal wrapped in a thick layer of filler, and the confident prose makes the ratio hard to see until you test each claim deliberately.
Can AI make up statistics in a market report?
Yes, and it does so fluently. Language models generate text that sounds right based on patterns, so they can produce specific-looking figures and citations with no basis — a 2026 consulting-report case where most citations were fabricated earned the label “vibe citing.” A CRE report can invent an absorption figure or attribute a number to a brokerage report that says nothing of the kind. Any statistic you would act on needs to be traced to its source first.
Should I trust an AI market report for an investment decision?
Not on its own. Use it as a fast first screen to decide whether a deal deserves real work, but trace every number that would enter an underwriting model, a lender package, or an investment-committee memo back to a verifiable source first. The report is a draft of your analysis, not the analysis. The judgment — and the accountability for a bad number — stays with a person.
What’s the difference between survey data and transaction-derived data in CRE?
Transaction-derived data comes from recorded events: actual sales, signed leases, filed prices. Survey data comes from what market participants reported when asked, which is useful for direction but subject to who responded and when. A recorded sale comp is stronger evidence than a surveyed asking-rent average, and an AI report will often blend the two without distinction. Knowing which you are reading tells you how much weight the number can bear.
Why do two AI reports on the same submarket sometimes disagree?
Because where the model lacks underlying data, it generates a plausible number fresh each time, and two runs land on different guesses. Agreement between runs suggests a real source underneath; divergence suggests fabrication. Running the same report twice is a cheap, effective way to see which figures the tool actually knows versus which it invents to fill space.
What should a small firm use AI market reports for, and not for?
Use them for fast, low-stakes screening, for drafting boilerplate you then verify, and for surfacing questions worth chasing. Do not use them, unverified, in anything a lender, investor, or committee relies on, and do not treat a free assistant’s summary of the open web as a substitute for a licensed transaction-data subscription when the decision is large. Match the tool’s known reliability to the stakes of the task.
Key takeaways
- Every number comes from a transaction, a survey, or a model, and an AI report blurs all three into one confident voice — putting the seams back in is the reader’s job.
- Two questions sort signal from filler: is the claim sourced to something you can open, and is it dated? Neither means filler.
- Five filler tells recur — round-number confidence, unsourced projections, generic macro boilerplate, uncheckable citations, and disagreement between two runs of the same report.
- Fabricated citations are the most dangerous filler because they mimic signal; spot-check the sources behind any figure you would stake a decision on.
- Use AI market reports for fast screening and drafting, never unverified in anything a counterparty relies on, and do not mistake a summary of the open web for licensed transaction data.
Want to know where AI market reports fit in your firm’s deal process — and where a verified data source or a custom screen would serve you better? A short conversation about your deal flow, your current tools, and the decisions you make from market data will map it faster than any checklist. Book your free AI-readiness assessment → and we will show you where these reports earn their place and where they should never be trusted alone.
Dirk Jan van Veen, PhD
