How this was measured

In plain English

What we did and what we found

We drew 2,500 businesses at random from the 43,469 active business registrations in Richmond, Henrico, Chesterfield and Hanover, then audited the web presence of the 574 we could confirm are customer-facing operations rather than holding entities. About a third of those — 33.1%, give or take 3.9 points — have no website at all. Among the ones that do, most are missing the basic technical signals search engines and AI assistants use to identify a business: 40.6% carry no structured data and 73.7% have no llms.txt file.

Everything below is how that was measured and where it can be wrong. If you only read one more section, read known limitations.

Population

What "Richmond small business" means here

The population is every business entity with ACTIVE status in the Virginia State Corporation Commission's corporation and LLC files whose registered address falls in Richmond city, Henrico, Chesterfield, or Hanover county. As of this run, that is 46,543 unique entities.

3,074 were excluded before sampling as entities that would not be expected to have a customer-facing website: holding companies, single-property real estate vehicles, homeowner and condo associations, churches, and foundations. The exclusions are name-pattern based and listed in full below. That leaves an eligible population of 43,469.

Sampling

A random draw, not a lead list

2500 businesses were drawn using Python's random.sample with seed 20260825 against the eligible population sorted by entity type and ID. The seed is published so the identical sample can be redrawn by anyone.

This matters more than it may sound. The obvious shortcut would have been to reuse a prospecting database — but a prospecting database is deliberately weighted toward businesses that look underserved, which would inflate every "missing X" figure here. A random draw from the full registry is the only version of this worth citing.

Measurement

How each business was audited

• Website presence. Two independent sources. Google Places is authoritative when a business has a profile. For businesses without one, a domain probe builds candidate hostnames from the registered name and accepts a match only when the page proves it belongs to that business — its street address, its phone number, or its full name plus a Virginia signal. A business is recorded as having no website only after both sources looked and neither found one.
• Mobile and speed. The homepage is fetched directly. Mobile readiness is the presence of a viewport meta declaration; speed is time to first byte and transferred page weight.
• AI readiness. Schema.org JSON-LD blocks are parsed from the homepage and their @type values recorded; the site's /llms.txt and /robots.txt are fetched, and robots.txt is checked for rules blocking GPTBot, ClaudeBot, PerplexityBot, Google-Extended, CCBot, and OAI-SearchBot.
• Google Business Profile. A Places text search on the registered name and address, accepted when at least 60% of the significant name tokens match.

Sectors

How businesses were grouped

Google assigns each business a category, and across a randomly drawn registry sample those fragment badly — 105 distinct categories across the confirmed businesses, only three of which clear 30. Reporting per category is therefore not possible from this sample; reporting per sector is.

Categories are grouped by keyword into nine sectors, of which seven cleared the 30-business threshold. The groupings are deliberately coarse, and the keyword lists live in sectors.py in the published pipeline so they can be disagreed with specifically rather than in general. Categories that carry no sector meaning — "Services", "Building", "Point of interest" — are left unclassified rather than forced into a bucket; 189 businesses fall there. Every business's raw Google category is in the CSV, so a different grouping can be applied to the same data.

Honesty

Known limitations

• A business trading under a name unlike its registered name may be recorded as having no Google Business Profile when it has one. This is the largest known source of error and it biases the "no profile" figure upward.
• The domain probe cannot find a site whose domain bears no resemblance to the registered name. Google Places covers most of that gap, but not all of it, so the "no website" figure is an upper bound.
• Only the homepage is audited. A site carrying schema on interior pages but not the homepage is recorded as having none.
• The registry lags. Virginia's published entity files run roughly eighteen months behind the live register, so businesses formed very recently are not in the population at all.
• Sole proprietors who never registered an entity are not in the population. This report describes registered businesses.
• Sector sample sizes are small enough that the no-website intervals overlap between sectors. Differences on that measure should not be read as a ranking. Review counts and llms.txt separate the sectors far more clearly.
• Confidence intervals are Wilson score intervals at 95% and describe sampling error only. They do not account for the measurement limits above.

Reproduce it

Run it yourself

The full pipeline — registry pull, sampling, audit, statistics, and this site — is a small Python project. With the same seed, it produces the same sample. The audited records behind every figure are published as CSV, one row per business, including the evidence for each website match.