The commoditization thesis is right about parsing and wrong about coverage. The moat survives, and it is already occupied.
The question
"Does the 'public data access is now commoditized' thesis kill the county-permits data-marketplace entry point, or does a curation/freshness moat survive for hyperlocal, poorly-structured public records?"
Two vault docs assert things that cannot both be fully true, and neither references the other. This is a reconciliation, not a market survey. Filed from the 2026-08-11 curiosity sweep.
What we already know (from the vault)
- The commoditization claim, verbatim. [[2026-02-16-vertical-software-selloff]] ranks "Public Data Access" third in its destroyed-moats list: "FactSet's and LexisNexis's core value was making technically-public-but-hard-to-use data searchable. Building that required armies of parsers, NLP pipelines, domain-specific NER models. Frontier models already know how to parse a 10-K... The parsing layer is now a commodity baked into the model." It adds that MCP is "eliminating the 'making it accessible' premium." Basis: Bustamante's own operating experience at Doctrine and Fintool, published into the ~$1T vertical-software selloff. No permit-specific or municipal-data evidence is offered.
- The entry point, verbatim. [[01-projects/data-marketplace/index]] lists four candidate entry points; the second is "County permits: public data that's valuable when curated and made searchable." That is close to a word-for-word restatement of the thing the selloff doc declares commoditized. The project has been
status: exploratorysince March 2026. - The selloff doc already resolved this, but only half-correctly. Its own Actionable section claims the data marketplace is "exactly the bet that this analysis validates," reasoning from moat #6 (Private and Proprietary Data, stronger). That holds for exclusive or hard-to-assemble datasets. It does not hold for an entry point whose stated value proposition is making public data searchable. The doc applied moat #6 to the project and never applied moat #3 to this specific entry point. That omission is the tension.
- We have first-hand field evidence, and it points at semantics, not parsing. The Tampa Accela pulls for the home build ([[04-finance/house-fund-tracker]], [[2026-10-17-build-cost-by-delivery-path]]) produced per-permit records including contractor, license number, declared job value and square footage, and surfaced the trap: "declared permit job values are routinely understated, so they are NOT build cost." A model that reads that permit perfectly still yields a wrong dataset.
- Freshness is already known to be the weakest thing to sell. [[2026-05-26-x402-mpp-monetizable-endpoints-ranked]] found live data feeds converting at 0-0.8% against skills at 33% in a 368-probe field study, and adding a
data_freshness_secondsfield actively lowered conversion. Whatever freshness is worth, our own best evidence says it is not worth much in the agent channel.
What the web says
- The structural number that breaks the analogy. Shovels' own documentation: "there are over 20,000 permit jurisdictions in the US, and no one has them all" (Shovels docs). A 10-K corpus is one regulator, one taxonomy, one free bulk filing system. Permits are ~20,000 schemas, no regulator, no central filing system, no bulk API.
- Access is gated by permission, not by parsing. Accela's API has "no self-serve path to data: each agency's administrator must approve and install an app before it can transact," with "no webhooks; extraction is record-by-record," "no bulk export," undocumented rate limits that "surface only as 429s at runtime," and the approval step explicitly "blocking horizontal multi-tenant integrators" (Supergood API report card). Bias flag: Supergood sells a competing API-wrapper product and its business is strongest where official access is poor; the page is also internally inconsistent, grading Accela's docs GOOD while assigning an overall F. Treat the specific mechanics as plausible and the grade as advocacy. Comparable vendor-published claims say OpenGov has "no public, broadly documented developer portal," with access contract-dependent; Tyler's open-data product is Socrata. Nothing surfaced on Clariti. None of this was confirmed against the vendors' primary docs.
- Open portals carry real record-level permit data, but as agency exports. San Francisco's Planning Department dataset on data.gov is described as exported from the Planning Department's Accela system and updated nightly (data.gov catalog). So the publication cadence is a property of the agency's export job, not of the consumer's polling.
- Shovels collects it itself, and says what hurts. It scrapes rather than licensing from aggregators, combines "jurisdictions that publish nothing online, which our team activates directly," and files FOIA requests it "tries to avoid because the process is manual and doesn't scale." Its stated limits: "The biggest hangup in this is the variance in required fields across jurisdictions"; digitization gaps "in the rural counties and unincorporated areas";
job_valueis "an extremely nuanced data point to understand, due to how frequently it's under-reported"; and scraping carries "linearly-scaling maintenance costs" (Shovels docs). Claimed coverage is 2,820+ jurisdictions, all 50 states, ~85% of US population, with its own quarterly index conceding it "strives for 100% permit coverage, but isn't there yet" (Shovels API, Shovels Q3 index). The jurisdiction count is inconsistent across its own and third-party pages (2,820+ / 2,750+ / 1,800+); not reconciled. - The free federal source is aggregate and cannot substitute. The Census Building Permits Survey publishes "national, state, CBSA, county and place" statistics of "housing units authorized by building permits" - counts, not records. No addresses, no contractors, no per-permit job values. Revised data land on the 17th workday of the month and final 2025 annual estimates were released 2026-05-14, and responses are voluntary (Census BPS). It answers "how many units in this county," never "who pulled a permit on this street last week."
- The commercial lane is occupied and visibly bifurcated. Enterprise tier is sales-gated: Shovels, ATTOM and Construction Monitor all quote rather than publish, with Construction Monitor estimated at $200-800/mo by territory and volume and Shovels' entry plan publicly reported around $599/mo; BuildZoom enterprise data is custom-quoted. Beneath that sits a cheap tail: PermitStack at $29-79/mo with a free tier, Permit Ledger at $39/mo flat, and off-the-shelf Apify scrapers covering 34 cities (PermitStack pricing comparison, Permit Ledger comparison, ColdIQ on Shovels, Apify). Bias flag: every one of these price sources is a vendor comparing itself to rivals. No figure here was read off the named vendor's own pricing page. Dodge, Cotality (CoreLogic) and BuildCentral pricing and coverage: not found.
Convergences and contradictions
- The contradiction is narrower than it looks, and it resolves by layer. Bustamante's moat #3 is a claim about document parsing over a standardized corpus. The county-permits entry point's difficulty is coverage across 20,000 non-conforming publishers. Both sides are correct about different layers: the parsing layer genuinely collapsed, and the thing that made permit data hard was never primarily parsing. The vault's own framing is what is wrong - "curated and made searchable" describes the commoditized layer, not the surviving one.
- Convergence on where the hard part lives. Shovels (the operator), Accela's integration surface (the gatekeeper) and Census BPS (the free alternative) independently point at the same place: field variance across jurisdictions, per-agency permission, and non-digitization. None of the three points at parsing.
- Convergence that freshness is the weakest leg. The web shows cadence is set upstream by the agency's export job; the vault shows agents will not pay for freshness and penalise it when advertised. Freshness is a real product attribute and a poor moat.
- The cheap tail is evidence for both theses at once. A $29/mo, 34-city permit API existing at all is moat #3 playing out exactly as predicted on the easy jurisdictions. No cheap vendor claiming the long tail is moat #3 failing on the hard ones. The market has already priced the layer boundary.
Synthesis for RDCO
Verdict: the thesis does not kill the entry point. It kills the entry point's description. Decomposing the difficulty four ways and marking what agent tooling actually cheapened: (a) ACCESS - split. Reading a published portal page is now near-free, which is real deflation. But an Accela tenant requires each agency's administrator to approve an app, offers no bulk export and no webhooks, and a meaningful share of jurisdictions publish nothing online or keep paper. An LLM does not shorten an approval queue, file a FOIA for free, or digitize a rural county. Shovels' own "linearly-scaling maintenance costs" is the whole economics: agents lowered the per-jurisdiction unit cost and did nothing to the exponent. Twenty thousand jurisdictions times a smaller linear cost is still a payroll. (b) NORMALIZATION - largely not cheapened, and this is where the moat concentrates. Schema reconciliation is a solvable mapping problem, but "variance in required fields" is not a mapping problem: a field the jurisdiction never collected cannot be recovered by a better reader. Missing-at-source is information-theoretically irreducible, and the residual work is per-jurisdiction semantic calibration - knowing that Tampa's declared job value understates true cost, which we learned by hand, not by model. (c) FRESHNESS - technically cheapened, bounded upstream, and not monetizable. Polling is free; the agency's nightly export is the ceiling; our own conversion evidence says buyers discount it. (d) ENTITY RESOLUTION - partially cheapened, still the hardest. Embedding and LLM fuzzy-matching genuinely help link permits to parcels, owners and contractors, but the ground-truth keys (parcel numbers, state license registries) are themselves per-jurisdiction assets you have to go get. A secondary source attributes Shovels an ~88% contractor match rate; unverified, and if true the uncovered 12% is the tell that nobody has published how to close it.
So a moat survives, but it should be renamed. It is a coverage-and-semantics moat, not a curation/freshness moat, and its defining property is that its cost is linear in jurisdiction count with no shortcut. That is the worst possible moat shape for a solo operator and a fine one for a funded team, which is precisely who holds it. Shovels claims 2,820+ jurisdictions and ~85% of population with a five-year collection head start; ATTOM, Construction Monitor, BuildZoom, Dodge and Cotality sit in the same lane; the easy slice is already a $29/mo commodity. The honest verdict is not "the thesis killed it" and not "the moat is open." It is "the moat survives, it is already occupied, and it is made of a cost RDCO cannot pay." Retire county permits as a business entry point. This costs the data-marketplace project nothing it had: [[2026-04-23-project-next-moves-audit]] already recommended defaulting to the clinical-research use case over county permits in April, and that recommendation has sat unexecuted for six months. Given RDCO is dormant (founder 2026-09-23), the right action is to strike the county-permits bullet and record why, so the next sweep does not re-raise it.
One correction propagates back. The selloff doc's claim that the data-marketplace bet is "exactly the bet that this analysis validates" should be scoped: validated for exclusive or hard-to-assemble datasets, contradicted for any entry point whose pitch is making public data searchable. That line in [[2026-02-16-vertical-software-selloff]] is the one that let this tension sit unexamined for eight months.
What is worth keeping is the demonstration, not the dataset. The four-layer decomposition above is a genuinely useful public artifact for the founder's stated near-term goal of credibility in agents-in-production for phData sales. The reason is that hyperlocal public records are the cleanest available counterexample to the lazy version of the commoditization story: model capability is not the binding constraint, harness design against thousands of non-conforming sources with missing-at-source fields and no ground truth is. We hold a non-hypothetical rep - the Tampa Accela pulls - and a concrete trap a naive agent pipeline gets wrong (declared job value read as build cost). "An agent that reads the permit perfectly and still produces a wrong dataset" is an original re-frame rather than a restatement of Bustamante, which is the bar for a Sanity Check piece. That is a writing candidate, not a venture, and it should be queued as one only if the founder wants a piece in that lane. No build, no data acquisition, no venture.
Why this is in the vault
It resolves the specific unsynthesized contradiction between moat #3 of [[2026-02-16-vertical-software-selloff]] and the county-permits bullet in [[01-projects/data-marketplace/index]], and supplies the strike-and-why for that bullet plus a scoping correction to the selloff doc's "validates the data marketplace" line. It also converts a dead entry point into a named credibility-artifact candidate for the agents-in-production positioning.
Open follow-ups
- The one number nobody publishes: the actual filing-to-availability lag distribution across jurisdictions. Shovels' documentation states no latency figure, and no vendor does. This is measurable from public portals without credentials, and it is the empirical crux of whether freshness is a differentiator at all or just marketing.
- Does the missing-at-source field-variance finding generalize to other hyperlocal record classes (code enforcement, business licenses, ABC/liquor licenses, septic and well permits)? If it does, the public artifact gets a general claim rather than a permit-specific one.
- Unverified in this brief: Shovels' true jurisdiction count (2,820+ / 2,750+ / 1,800+ across sources), the ~88% contractor match rate (secondary source only), all pricing above (vendor-comparison pages only, none read off the vendor's own pricing page), Dodge/Cotality/BuildCentral coverage and pricing at all, Accela/Tyler/OpenGov/Clariti API mechanics against primary vendor docs, and whether the 2015 Accela/Zillow/BuildFax open permit-data standard was ever released or adopted. No paywalls were hit.
Related
- [[2026-02-16-vertical-software-selloff]]
- [[01-projects/data-marketplace/index]]
- [[2026-05-26-x402-mpp-monetizable-endpoints-ranked]]
- [[2026-04-23-project-next-moves-audit]]
- [[06-reference/concepts/products-for-agents]]
- [[04-finance/house-fund-tracker]]
- [[2026-10-17-build-cost-by-delivery-path]]
Sources
Vault
06-reference/2026-02-16-vertical-software-selloff.md01-projects/data-marketplace/index.md06-reference/research/2026-05-26-x402-mpp-monetizable-endpoints-ranked.md01-projects/positioning/2026-04-23-project-next-moves-audit.md06-reference/concepts/products-for-agents.md04-finance/house-fund-tracker.md01-projects/life/home-build/2026-10-17-build-cost-by-delivery-path.md
Web
- https://docs.shovels.ai/docs/foundations-understanding-permits (primary: operator's own stated limitations)
- https://www.census.gov/construction/bps/ (primary: Census Building Permits Survey scope and cadence)
- https://supergood.ai/api-report-card/accela (secondary, competitor-published; bias-flagged)
- https://www.shovels.ai/api
- https://www.shovels.ai/blog/shovels-quarterly-permit-index-q3-2025/
- https://catalog-old.data.gov/dataset/planning-department-records-projects
- https://permit-stack.com/blog/building-permit-data-api-pricing-compared.html (vendor-published; bias-flagged)
- https://permitledger.com/compare/construction-monitor (vendor-published; bias-flagged)
- https://coldiq.com/tools/shovels (secondary)
- https://apify.com/spriggs-data/us-building-permits