Skip to content
All articles

Data & Research

How complete is the queue? Measuring our own coverage

4 MIN READ

Shawn Azdam

Founder

Emerald survey grid with filled and hollow markers over dark contours.
On this page

"The queue holds every plugging package we can find" is a claim that invites the obvious question: what fraction is that? Coverage we cannot measure is marketing. So we measure it — against independent state award lists, using the same capture-recapture math ecologists use to count fish. Here is the methodology, the current numbers, and the gaps we already know about.

Capture-recapture, applied to bid boards

The method needs two independent samples of the same population. Sample one: packages in our queue for a state and quarter. Sample two: awards published on the state's own award lists for the same period — compiled independently, from sources our crawler does not use for ingestion. Packages appearing in both samples are the recaptures. The ratio estimates the total population, including packages in neither sample.

Independence is the load-bearing assumption, and we defend it operationally: the award-list compilation uses different sources, different staff, and a different pipeline from ingestion. Where the same county PDF feeds both sides, we exclude it from the comparison set. Imperfect independence biases the estimate upward — shared sources inflate the overlap — so we treat our coverage figures as ceilings and say so in every report.

What the numbers say

Across covered states, the queue currently captures the large majority of biddable packages each quarter — high enough that a contractor searching the queue instead of fifty county sites misses little, low enough that we keep investing in sources. Coverage varies by state in exactly the way you would predict: states with centralized portals and standardized formats score highest, while states where work posts across dozens of county sites with inconsistent formats score lowest. Centralization is a gift to everyone downstream.

The misses cluster. Emergency awards with same-week turnarounds, packages posted only on physical courthouse boards, and work bundled into general construction contracts without plugging keywords account for most of what we miss. Each cluster has a remediation plan: faster cadences for emergency-prone sources, county-courthouse coverage partnerships, and keyword expansion reviewed quarterly.

The gaps we admit

Three known blind spots, stated plainly. First, states where we have not yet built source coverage — the queue labels these explicitly rather than showing an empty result set that implies no work exists. Second, historical depth: the queue is comprehensive going forward from launch, but pre-launch award history is imported opportunistically and should not be treated as complete. Third, private work: operator-direct plugging that never touches a public solicitation is invisible to us by construction, and we make no claims about it.

Coverage by channel, not just by state

The headline coverage number hides as much as it reveals, so we cut it three ways. By source type: state portals score highest (single format, reliable cadence), county sites next (fragmented but watchable), and aggregators last (we treat them as corroboration, never as primary — their errors compound ours). By package stage: solicitations are our strongest channel, awards next, and amendments weakest, because addenda post irregularly and expire from listing pages fastest. By season: winter coverage dips in northern states as county offices close and postings slow, then rebounds in the spring bid rush.

Each cut drives a resourcing decision. Weak amendment coverage justified the versioning work that now diffs every re-fetch. County-site fragmentation justified the per-source scheduler that matches cadence to rhythm. Seasonal dips justified overstaffing ingestion review in March and April. A single coverage number would tell us none of this. Measure in slices, invest where the slices are thin, and re-measure quarterly to confirm the investment worked.

Why we publish this

Coverage methodology is usually proprietary — every data vendor claims comprehensiveness and none shows the math. We publish ours for the same reason we publish confidence bands and risk-model weights: our users make expensive decisions from this screen, and they deserve to know what the screen might be missing. An estimator who knows the queue catches nine in ten packages plans differently — and better — than one who assumes ten in ten.

What we think at Terrabillion

Every data product should publish its coverage methodology. The practice costs little — the measurement work is the same work good operations requires anyway — and it converts the vendor's boldest claim into the buyer's most checkable one. We invite scrutiny of our math, our independence assumptions, and our blind spots; each correction improves the next quarter's number. Comprehensiveness is a direction, not a destination, and we would rather be audited than believed.

Orphan-well intel, monthly

Contact sales