Measuring small-business web presence without Google Places.

Three claims on this page did not survive the scale-up.

This pilot published on 2026-08-14. A day later I ran the same method across 11 verticals and 1,916 firms in the same two counties, and it contradicted three things stated here. The corrections are below, dated, with what was claimed, what is actually true, and how it was found. Nothing was quietly edited out. The full study is at Small-business web presence across eleven verticals.

Five numbers, with their error bars attached.

At N = 119 the interesting quantity is not any single rate. It is how far each rate moved when I fixed a measurement defect. The table below is the corrected position; the sections after it are how each correction was found and what it was worth.

Every interval on this page is a Wilson score interval at 95%. Wilson rather than the textbook normal approximation because several of these proportions sit close enough to a boundary that the normal interval runs below zero, which is not a defensible thing to print.

21.8% of HVAC contractors have no website. Read that as a floor.

Overture Maps is assembled from business-listing feeds. A company enters this frame by having been listed somewhere first. That is a selection rule, and it selects in exactly the direction that matters here: a firm with no website, no listing, and no online footprint of any kind cannot appear in the denominator, because nothing put it there.

So the 26 of 119 firms with no site are the ones who are listed somewhere and still have no site. The firms invisible to the frame are disproportionately the ones with the least digital presence. 21.8% is a lower bound on the true rate among all operating HVAC contractors in these two counties, and the gap between the two is a quantity this method cannot estimate.

That last sentence has since been improved on. The scale-up bounded the frame against Census County Business Patterns, BLS QCEW and Census Nonemployer Statistics, and it also measured a second bias running the other way — a 30.5% recall miss over the no-presence stratum, mostly booking-platform pages the listing feeds never carry. Two biases, opposite directions, one measured and one only bounded. That is why the full study refuses to publish a single point estimate for local business as a whole.

Fold in the sites that exist but do not load and 35 of 119 firms — 29.4% [22–38.1] — have no working web presence at all. That figure inherits the same floor property, for the same reason.

The interval — 21.8% [15.4–30.1] — does not help with this. Wilson intervals describe how much a number would move under resampling from the same frame. They are silent about whether the frame is the right one. On this study the frame bias is almost certainly the larger error, and it points in a known direction, which is the most useful thing I can say about it.

Place data is one record per listing, not one per business.

The query returned 175 candidate records. After exclusions, 145 records remained. Those 145 records are 119 businesses. One firm appeared 9 times — one record per service-area pin across six localities.

Counting records instead of firms inflates N by 47%. That alone would be a citation-integrity problem. The worse part is that it also biases the headline in a specific direction, because duplication is not random.

Every one of the 14 multi-listing firms in this frame has a website. 75.2% of the 105 single-listing firms do. Multi-location operators duplicate the most and are the most likely to have a site, so record-weighting quietly overstates web presence: the no-website rate comes out at 17.9% per record versus 21.8% per firm. Nearly four points, in the flattering direction, from a decision most people never notice they made.

Getting the category filter right is harder than it looks.

The intuitive way to enumerate a trade is a keyword filter over the category field. It fails in both directions, and both failures are measurable here.

False positives. A regex including duct — a reasonable token for an HVAC filter — matches 61 records inside the bounding box. 2 of them are HVAC. The rest are broadcasting_media_production (27), music_production (17), beauty_product_supplier (6), and so on down. “Duct” is a substring of “production” and “products.” That is a 96.7% false positive rate on one token.

False negatives. Filtering on the primary category alone is the other trap, because the source data miscategorises real contractors. 10 of the 119 firms in the final frame — 8.4% — carry a primary category that is not hvac_services. One HVAC contractor is filed under electrician, another under plumbing, one under truck_dealer. A primary-category-only filter drops all of them.

What I did instead: select on the primary category, the alternate-category list, and a short set of adjacent primaries, then adjudicate every record from the wider tiers by hand with a written reason. Thirty records were excluded that way, and eleven were affirmatively kept. The reason codes ship in the dataset so anyone who would draw the line differently can say exactly where.

The recall check found nothing. That is worth reporting.

A no-website finding is only as good as the search that failed to find a website. So all 26 firms in the no-site set went through a second, independent pass: fresh searches on name, city and phone in several formats, domain guesses, BBB and chamber-of-commerce profiles, Facebook and Nextdoor, averaging 3.4 distinct sources per firm.

It recovered 0 additional websites. 25 firms returned a confirmed not-found — several with a BBB profile that explicitly lists no website — and 1 had a domain listed in a chamber directory that no longer resolves. 19 of the 26 verdicts were recorded at high confidence, 7 at medium.

I expected the second pass to claw several back. It clawed back none. Miss rate 0%, 95% interval [0–12.9] — a zero numerator still leaves real room for error at n = 26, and the interval says so. The result makes the no-website finding more robust, and I would have published it either way.

Everything needed to run it again.

The frame cost nothing and required no account. Overture Maps publishes its releases as Parquet on an anonymous S3 bucket; DuckDB reads them in place. The bounding-box pull returned 36,852 places in roughly 8 seconds on a laptop.

The county assignment is worth isolating, because it is the step people skip. Overture records carry a locality string, and using it is tempting. I did a point-in-polygon join against TIGER/Line 2024 county boundaries instead. 2,879 of the 36,852 places in the rectangle — 7.8% — fall outside both county polygons. They are in Mississippi, in Florida, or in the water. A locality-string filter would have kept an unknown share of them and quietly changed the denominator.

-- 1. Pull the bounding box straight from Overture's public S3 bucket.
INSTALL spatial; LOAD spatial;
CREATE TABLE bbox_places AS
SELECT id, names.primary AS name,
       categories.primary AS cat_primary,
       categories.alternate AS cat_alt,
       confidence, websites, phones, addresses,
       operating_status,
       ST_X(ST_GeomFromWKB(geometry)) AS lon,
       ST_Y(ST_GeomFromWKB(geometry)) AS lat
FROM read_parquet(
  's3://overturemaps-us-west-2/release/2026-07-22.0/theme=places/type=*/*',
  hive_partitioning = 1)
-- The southern edge sits offshore, so it does not bind: any ymin at or
-- below 30.22 returns the identical 36,852 places.
WHERE bbox.xmin > -88.50 AND bbox.xmax < -87.35
  AND bbox.ymin >  30.10 AND bbox.ymax <  31.40;

-- 2. Assign county by POINT-IN-POLYGON against TIGER/Line 2024,
--    not by trusting the locality string on the record.
CREATE TABLE bbox_county AS
SELECT p.*, c.county_name AS county_tiger
FROM bbox_places p
LEFT JOIN counties c
  ON ST_Contains(c.geom, ST_Point(p.lon, p.lat));

-- 3. Candidate selection. Three tiers, each adjudicated by hand:
--    A = primary category is hvac_services
--    B = hvac_services appears in the ALTERNATE categories
--    C = an adjacent primary category
SELECT * FROM bbox_county
WHERE county_tiger IS NOT NULL
  AND (operating_status IS NULL OR operating_status <> 'permanently_closed')
  AND confidence >= 0.6
  AND ( cat_primary = 'hvac_services'
        OR list_contains(cat_alt, 'hvac_services')
        OR cat_primary IN ('air_duct_cleaning_service',
                           'hvac_supplier',
                           'commercial_refrigeration') );

Downstream of the query, five scripts do the rest: frame construction and record linkage, two-client liveness measurement, domain verification for search-recovered sites, the PageSpeed Insights and CrUX collection, and the assembly step that computes every interval on this page. A sixth is kept deliberately — the first liveness pass, LibreSSL bug intact, because the artifact it produced is half of section 04. They are not polished. They are the scripts that produced these numbers. If you want them for a replication, email me and I will send them.

Take the data.

The aggregate dataset behind every number on this page is a single CSV: one row per measured quantity, with section, numerator, denominator, value, unit, and both bounds of the 95% Wilson interval where a proportion is involved.

Download cds-hvac-web-presence-pilot-2026-08-14.csv

Measured and published August 14, 2026. Corrected 2026-08-15 — see the notice at the top of the page and the corrections rows in the CSV. Corrections and replications welcome; if a number here is wrong, I would rather know.