Every number FreeCodeChecks shows you traces back to the same underlying process: an ordinance document goes in, and structured, categorized rules come out the other end. We think that process is worth explaining in plain terms — not the marketing version, the actual mechanics, including the parts that are genuinely hard and the parts where the data is messier than a clean headline number suggests.
The number, measured today
We re-ran a live count against production on 2026-09-07. Our sign-rule table holds 660,173 individually extracted rules. Those rules carry 492 distinct category labels, and they are attached to 11,156 individual local governments — the subset, out of the 33,535 governments we track nationwide, that currently has at least one rule extracted from a retrieved ordinance.
That last distinction matters and is easy to blur. “Governments tracked” and “governments with extracted rules” are two different numbers, and the gap between them is our remaining work, not a rounding error. A jurisdiction we track but have not yet extracted is a jurisdiction where we can tell you who holds authority but not yet what their code says.
Sixteen categories carry 92% of the corpus
The 492 category labels are not evenly loaded. The sixteen largest account for 610,006 rules — 92.4% of everything we have extracted:
| Category | Rules | Share of corpus |
|---|---|---|
| Ground signs | 124,102 | 18.80% |
| Temporary signs | 114,993 | 17.42% |
| Building signs | 109,094 | 16.53% |
| Permit process | 66,285 | 10.04% |
| General standards | 27,910 | 4.23% |
| Prohibited signs | 26,196 | 3.97% |
| Directional signs | 24,310 | 3.68% |
| Definitions | 20,835 | 3.16% |
| Awning signs | 17,266 | 2.62% |
| Electronic signs | 14,371 | 2.18% |
| Off-premise signs | 13,226 | 2.00% |
| Window signs | 12,749 | 1.93% |
| Illumination | 11,644 | 1.76% |
| Nonconforming signs | 10,861 | 1.65% |
| Sign measurement | 8,200 | 1.24% |
| Signs (generic) | 7,964 | 1.21% |
The shape of that table is a fair description of what municipal sign codes are actually about. Three categories — ground signs, temporary signs, and building-mounted signs — are more than half the corpus on their own, because those are the three things a sign ordinance spends most of its text regulating. Permit process is the fourth largest not because permits are complicated in any one town, but because almost every town writes its own procedure for them.
The long tail is 476 categories and 50,167 rules
Subtract the top sixteen and you are left with 476 categories holding 50,167 rules — 7.6% of the corpus spread very thinly. 445 of the 492 categories carry fewer than 100 rules each, and 154 of them appear on exactly one rule in the entire country.
That tail is not noise to be swept up. It is a direct consequence of the fact that municipalities do not share a vocabulary: one town writes a chapter on “changeable copy signs,” the next calls the same object a “reader board,” a third files it under “message centers,” and a fourth invents a category that exists in exactly one code in America. We wrote about that vocabulary problem in more depth in forty names for the same sign, and it is the single biggest reason a keyword search of a code library is a poor substitute for reading the code.
Why an earlier version of this page reported a smaller number
An earlier version of this article reported 567,913 rules across “16 categories” and flagged the difference from a larger figure used elsewhere on the site as an unresolved discrepancy. That was our error, and the correction is boring rather than dramatic: 567,913 was the sum of a hand-picked set of sixteen category counts, not a count of the table. The missing rules were simply the other 476 categories. There is no reconciliation mystery — there was a subtotal mistaken for a total, and both the number and the “organized into 16 categories” framing were wrong. The figures above come from a single COUNT over the whole table rather than a sum of parts, which is the check that would have caught it.
The general method, top to bottom
Every row in that table started as raw ordinance text — a municipal code chapter, often buried dozens or hundreds of pages into a much larger zoning code, that a human clerk would otherwise have to locate and read manually. Getting from that raw document to a structured row in our database follows roughly the same four stages regardless of which category the rule lands in:
- Acquisition. The ordinance text is retrieved from the municipality’s published code — whichever platform hosts it — and registered before anything else happens to it.
- AI extraction into category / attribute / value facts. The sign-code chapter (once isolated from the surrounding zoning code) is processed to identify discrete, structured facts — a category (like “awning signs”), a specific attribute within that category (like “projection clearance”), and the value the ordinance actually specifies for it. This is the step that turns unstructured legal prose into the rows counted in the table above, and the step where the 492-label vocabulary comes from: the category is read out of the ordinance rather than forced into a fixed list.
- Confidence scoring. Each extracted fact is scored for how confidently it was pulled from the source text — a numeric value with an explicit unit stated in the ordinance scores very differently than a value inferred from ambiguous language like “as determined by the zoning administrator.” Across the whole corpus, 88.1% of rules carry high confidence.
- Human-review flagging for ambiguous cases. Facts that don’t clear the bar aren’t published silently — they carry a review flag. 9.65% of all 660,173 rules are flagged that way today.
Those last two numbers are the ones worth arguing with, and we do argue with them in a companion piece, what a sign code doesn’t tell you, which is about how much to trust any single extracted rule. This article is about the pipeline and the corpus; that one is about the boundary where automated extraction should hand off to a person.
One further caveat that the corpus size can obscure: a large number of extracted rules for a jurisdiction is not the same as a complete picture of that jurisdiction. Codes differ enormously in how much they write down, and the authority question — which government’s code even applies to an address — is answered separately from the extraction described here. We cover that in who holds sign authority in America.
See the extraction for your project
The structure described above — category, attribute, value, citation, confidence — is exactly what a report shows you for one address, minus the 660,172 rules that belong to somebody else’s jurisdiction. Creating an account is free.