Public procurement data quality: the August 2026 benchmark.

A zero is not automatically an outage, and a null is not automatically bad source data. In WebTruffle's public schema 2.0 procurement snapshot, 102 of 104 scheduled source-runs succeeded. Eight successful runs returned zero records. Across 91,952 latest normalized records, buyer and classification presence exceeded 99.6%, while supplier, deadline, and value presence depended heavily on the source, lifecycle stage, and current mapper.

This is a benchmark of the published normalized data product and its collection evidence. It is not a grade for SAM.gov, TED, Find a Tender, or Contracts Finder. A missing normalized value can mean the field does not apply to that notice, the publisher did not supply it, the selected source route did not expose it, or the normalizer did not map it. The report keeps those explanations separate wherever the evidence allows.

The measured edition is dated August 8, 2026, generated August 9, and evaluated for this report on August 11. The exact public release, manifest, source-health evidence, and field-presence rows remain available for reproduction.

Public procurement data quality benchmark: the findings

Five findings matter more than a league table:

1. Collection was operationally reliable, not perfect. Of 104 source-days from July 14 through August 8, 102 succeeded: 98.08%. The two failed or incomplete source-days were both in the SAM.gov series.
2. Eight zero-record runs were successful. Seven were TED weekend runs inside a window in which TED's nominal publication schedule was Monday through Friday. The eighth was Contracts Finder. Treating all eight as outages would have created false incidents.
3. Buyer and classification presence were the strongest normalized fields. Buyer was present on 91,943 of 91,952 records, or 99.99%. Classification was present on 91,625, or 99.64%. Presence still does not prove a resolved legal identity or a correct category.
4. The pooled supplier, deadline, and value rates were composition-weighted, not universal quality scores. Supplier appeared on 3,612 records, or 3.93%; deadline on 31,254, or 33.99%; and value on 30,850, or 33.55%. Those rates mostly reflect the dataset's source and stage mix.
5. The largest gaps identify pipeline work as well as source limits. TED supplier presence was zero even among 33,421 normalized award-stage rows, although the official eForms model supports winner information in result notices. Find a Tender source-link presence was zero in the public schema 2.0 snapshot. Those are normalization or retrieval gaps to investigate, not evidence that the official portals contain no winner or notice links.

There is no honest “best source” in these figures. Contracts Finder was 92.01% award-stage in the measured population; Find a Tender was 98.43% contract-stage; SAM.gov contained only planning and opportunity stages; and TED was split almost evenly between award and opportunity rows. The denominator changed before the percentage did.

Scope, cutoff, and what this report does not measure

The benchmark covers four official publication systems represented in one public normalized edition:

• SAM.gov Contract Opportunities;
• Tenders Electronic Daily (TED);
• Find a Tender; and
• Contracts Finder.

The collection-health window spans 26 observed calendar days, July 14 through August 8, 2026. Each source has 26 recorded run-days, giving 104 source-days. Health is based on the final stored outcome for each source and date.

The field-presence population contains 91,952 latest normalized rows as of the August 8 cutoff. In this young retained history, those rows were observed within the same July 14–August 8 period. Each stable record contributes its latest observed representation, rather than one row for every day it appeared.

The source denominators are unequal:

That distribution is why a pooled percentage is not a neutral cross-source average. TED dominates it. A record-weighted aggregate answers “what share of this exact normalized snapshot has the field?” It does not answer “how complete is the typical portal?”

This report does not measure:

• recall against an external universe of every procurement that should legally have been published;
• the accuracy of a non-null buyer, deadline, value, supplier, or classification;
• whether all versions, corrections, attachments, lots, awards, or contracts were captured;
• comparable national procurement-market coverage;
• publication compliance by individual buyers;
• the freshness lag between a real-world event and its first public appearance; or
• a full year, despite the report's 2026 date.

It measures collector outcomes and normalized field presence in one dated public data product. Use the source comparison for jurisdiction, threshold, and portal scope, and the government contract analysis method for market measures and analytical denominators.

How the benchmark was calculated

The release publishes explicit tests for six fields. A value counted as present when:

For each source and field:

presence rate = present normalized records / latest normalized records for that source

For collection health:

run success = successful final source-day outcomes / scheduled source-days

The field calculation treats null and blank as unavailable. It does not turn them into zero, false, or an inferred value. The Open Contracting Partnership's quality guidance similarly separates completeness, timeliness, accessibility, retrievability, correctness, and use-case relevance. It also makes a useful point for this benchmark: the fields required depend on the question. Supplier data matters for competition analysis; payment dates matter for late-payment analysis.

The stage analysis in this report was recomputed from the release's 207 MB rolling-observed-history.csv, grouped by source and record_stage. Because the public history begins within the measured window, that file contains the same 91,952 latest stable records as the field-presence population. Artifact hashes were checked against the manifest before calculation.

Three rules prevent false precision:

1. Per-source numerators and denominators are primary; pooled figures are supporting context only.
2. Supplier, deadline, and value rates are re-read at applicable lifecycle stages.
3. No run, record, and field measures are averaged into a composite score.

Collector health across 104 source-days

The aggregate run-success rate was 102/104, or 98.08%.

The August 8 daily edition contained 270 rows: 142 Contracts Finder, 51 Find a Tender, 77 SAM.gov, and zero TED. That daily count is not the 91,952-row field denominator. It is one target-day publication or update slice.

A successful collector outcome establishes that the planned request, parse, and storage path completed under its run contract. It does not prove that the source returned every applicable record. It also does not prove that every returned field was mapped correctly.

The SAM.gov series had two failed or incomplete days. The public health summary intentionally combines those states, so this report does not invent a root cause. The correct production response is to retain the failed partition, replay it, and withhold any “complete” downstream claim until the rerun and reconciliation pass.

Why zero records did not mean outage

The four sources do not share one definition of “today.”

TED runs on the OJ S issue calendar

TED says daily editions are nominally published Monday through Friday, with the website available during the publication morning and the daily XML package available by its stated 09:30 latest time. The TED release calendar is the authority for holidays and exceptional non-publication days. (TED data reuse)

There were seven weekend dates inside the measured window. TED recorded seven successful zeroes. The final Saturday, August 8, was explicitly assessed as an expected quiet day, and the last non-zero publication was August 7. Those zeroes are evidence that the collector ran on non-edition days, not evidence of seven outages.

SAM.gov exposes a posted-date slice, not a documented change stream

The SAM.gov Opportunities API requires postedFrom and postedTo, returns the latest active opportunity version, and directs users to Data Services for all versions. It says active notices are updated daily and archived notices weekly, but its public request contract does not document a modified-since filter.

It follows that a successful posted-date query is not proof that every later revision to an older notice was captured. That is an inference from the documented filter and version behavior, not a claim about an unpublished backend. A change-aware pipeline needs overlapping replay, versioned bulk data, or another reconciliation route.

Find a Tender and Contracts Finder expose update semantics

The Find a Tender OCDS release API defines updatedFrom and updatedTo as the times a record was last updated. Edited notices can therefore reappear in a later interval. Use overlapping windows, cursor pagination, OCID plus release identity, and content hashes.

The Find a Tender OCDS API implementation guide turns that benchmark boundary into a cursor-safe collector with record-package reconciliation and lifecycle-aware validation.

The Contracts Finder OCDS search defines its publication interval as the latest edit and the first publication in the current stage. A later window can contain an edit or stage transition rather than a new procurement.

Contracts Finder also has a changing policy boundary. Current UK guidance says procurements beginning under the Procurement Act 2023 publish notices on Find a Tender from February 24, 2025, while Contracts Finder continues to matter for earlier-regime activity. Lower or intermittent 2026 Contracts Finder yield therefore needs policy context before it becomes an incident. (Contracts Finder transparency guidance)

The operational rule is simple: classify zero only after checking HTTP status, authentication, pagination, response schema, the source's native publication clock, and an expected-edition or update-window signal.

Field presence across 91,952 normalized records

The exact public numerators show what the percentages hide:

Buyer presence was nearly universal, but identity was not solved

Buyer passed when either an official buyer identifier or a published buyer name existed. That makes 99.99% useful for search and display, but it is not a 99.99% legal-entity resolution rate. Names can vary across releases, languages, departments, and reorganizations. Supplier and buyer analytics still need scheme-aware identifiers and a disclosed unresolved share.

Classification presence was high across incompatible schemes

Classification reached 99.64% overall. SAM.gov uses NAICS and PSC; TED and the UK sources commonly use CPV. A populated code makes category analysis possible inside its source vocabulary. It does not make NAICS, PSC, and CPV interchangeable. Keep scheme, code, source label, version, and normalized hierarchy fields together.

Source-link gaps were concentrated in the public Find a Tender mapping

Source-link presence was 99.09% overall. All 829 missing Find a Tender links and nine Contracts Finder gaps account for the 838 missing URLs. Find a Tender publishes notice and OCID identifiers and provides OCDS release and record routes, so the zero rate is a schema 2.0 normalization gap to repair and reconcile—not evidence that its notices cannot be located.

Low pooled supplier, deadline, and value rates need stage context

The 3.93% supplier rate is dominated by SAM.gov and TED rows with no normalized supplier. The 33.99% deadline rate combines open competitions with award, contract, planning, cancellation, and other notices. The 33.55% value rate mixes estimated, award, and contract bases across currencies and grains.

Those numbers are useful as a warning about the published normalized product. They are not comparable portal scores.

Use stage-conditioned denominators

The measured stage mix was:

• Contracts Finder: 2,810 award, 208 opportunity, 27 amendment, and 9 planning rows;
• Find a Tender: 816 contract and 13 cancellation rows;
• SAM.gov: 17,721 opportunity and 3,756 planning rows; and
• TED: 33,421 award, 31,406 opportunity, 1,097 planning, and 668 other rows.

Together, opportunity and award stages represented 93.05% of the population, but their expected facts differ.

Measure supplier where a winner can exist

Contracts Finder supplier presence was 99.96% among its 2,810 award rows. Find a Tender supplier presence was 98.41% among its 816 contract rows. Zero supplier presence on SAM.gov planning and opportunity records is structurally unsurprising: a pre-award notice normally has no winner.

TED requires a different conclusion. Its official eForms model distinguishes consultation, planning, competition, result, modification, completion, and change notice families. Result notices can carry award decisions, contracts, values, and winners. (TED notice families; TED competition results)

Zero normalized supplier presence across 33,421 TED award rows therefore identifies a current pipeline limitation. A robust TED award parser must follow the eForms XML relationships among lot results, tenders, winners, and contracts instead of assuming a flat search field can represent them.

Measure deadline on actionable competition records

Opportunity-stage deadline presence was:

• Contracts Finder: 208/208, or 100%;
• SAM.gov: 17,602/17,721, or 99.33%; and
• TED: 6,872/31,406, or 21.88%.

Find a Tender had no opportunity-stage rows in this public schema 2.0 population, so its overall zero deadline rate is not a valid test of whether Find a Tender publishes tender deadlines. The official notice-type guide says UK4 advertises tenders or requests to participate, UK6 follows an award decision, and UK7 follows contract signing. A contract-heavy observed slice should not be scored on bid-deadline coverage.

TED deadlines can also be lot-level and can distinguish tender submission from a request to participate. A correct denominator needs actionable competition lots and the applicable deadline type, not every TED notice. (TED procedure, lot, and deadline fields)

Keep value basis and object grain visible

Value presence was 100% on Contracts Finder award rows and 71.15% on its opportunity rows. It was 24.63% on Find a Tender contract rows, zero on the measured SAM.gov planning and opportunity rows, 41.54% on TED award rows, and 42.51% on TED opportunity rows.

The SAM.gov public response contract documents award amount when available but no general structured pre-award estimated-value field. Zero value in this pre-award normalized slice is therefore consistent with the selected route and schema. It should not be “repaired” with guessed dollars.

Even when present, value is not one additive measure. Estimated procedure value, lot value, award value, contract value, maximum framework value, and transaction obligation answer different questions. Store basis, object, currency, tax treatment, and source scope before aggregating.

Read each source through its observed stage mix

SAM.gov: strong pre-award timing, no normalized value population

SAM.gov contributed 21,477 rows: 82.51% opportunity and 17.49% planning. Buyer and source-link presence were 100%; deadline was 99.00%; classification was 98.52%; supplier and value were zero.

That profile is useful for the dataset's pre-award discovery purpose. It is not an award-data quality assessment. The API's notice-type list also includes presolicitations, sources sought, special notices, solicitations, combined notices, justifications, awards, surplus property, and other forms. Each has different applicable fields.

The operational concern is revision history. Because the public route documents posted-date filters, only the latest active version, and no modified-since filter, run success does not establish complete amendment capture. The SAM.gov API implementation guide covers that collector boundary.

TED: predictable issue cadence, mixed lifecycle, relationship gap

TED contributed 66,592 rows: 50.19% award, 47.16% opportunity, 1.65% planning, and 1.00% other. Buyer, classification, and source-link presence were effectively complete; value was present on 41.53%; deadline on 10.43%; supplier on zero.

The seven successful weekend zeroes show why schedule-aware health matters. The mixed stages show why raw null rates mislead. The award-supplier result shows where a flattened normalization strategy stops being sufficient.

TED says optional eForms fields are omitted when unfilled rather than represented as empty XML elements. Absence is therefore meaningful, but its cause still depends on the notice form, applicable business rule, publisher input, and parser. (TED field metadata)

Find a Tender: contract-heavy observed slice, not a portal-wide score

Find a Tender contributed 829 rows: 816 contract and 13 cancellation records. Buyer and classification presence were 100%; supplier was 96.86%; value was 24.25%; deadline and source link were zero.

The population's 98.43% contract share explains why supplier was high and deadline was not applicable. The zero source-link result is a normalization gap. The source's OCDS API itself supports planning, tender, and award stages, update-time filters, cursor pagination, notice IDs, and OCIDs. This measured slice cannot support a claim about full Find a Tender opportunity coverage.

For lifecycle work, retain both release and record views. Find a Tender documents release IDs as unique within an OCID and party IDs as unique within a release, not globally. (Find a Tender release packages; record packages)

Contracts Finder: award-heavy legacy tail with strong mapped fields

Contracts Finder contributed 3,054 rows: 92.01% award, 6.81% opportunity, 0.88% amendment, and 0.29% planning. Buyer and classification were 100%; source link and deadline 99.71%; value 97.97%; supplier 91.98% overall and 99.96% on award rows.

Those figures describe a strongly mapped, award-heavy retained population. They do not make Contracts Finder the best current UK opportunity source. New Procurement Act procurement notices belong on Find a Tender, while Contracts Finder remains relevant to the legacy tail and earlier regimes.

Its one successful-zero day should be investigated against that publication boundary and the requested window, not automatically escalated as an outage.

Which fields are fit for which job

The benchmark supports decisions about the normalized edition, not universal source rankings.

The government contract tracker guide explains how to turn those decisions into monitoring states. The government contract analysis guide explains how to carry the coverage denominator into buyer, category, and supplier measures.

Separate health, coverage, presence, and correctness

A production quality report should publish at least five distinct layers:

1. Transport and run health: HTTP status, authentication, retry outcome, pagination completion, parser completion, and accepted artifact state.
2. Expected publication or update coverage: whether the source was expected to publish an issue, return updates, or remain quiet in the requested interval.
3. Observed record coverage: source, stage, date, stable identity, versions, and known collection boundaries for the retained population.
4. Applicable-field presence: numerator and denominator at the object and lifecycle stage where the field can legitimately exist.
5. Correctness and reconciliation: schema validation, source samples, relationship checks, duplicate tests, hashes, and comparison with an independent official route where available.

One measure cannot substitute for the next. A 200 response with a complete page walk can contain incorrect values. A valid zero can be healthy. A populated supplier name can refer to the wrong lot. A missing deadline can be correct on a signed-contract notice.

That is why this report does not create a weighted “quality score.” Weight selection would hide the evidence and reward source-stage mixes with more easily populated fields.

Reproduce and cite the benchmark

The government tender data with Python guide provides a downloadable recipe for pinning this release, verifying its manifest and artifacts, checking the schema/export boundary, and exporting query provenance.

Use the tagged August 8 release as the stable entry point. Its immutable artifact pattern is:

https://github.com/webtruffle/government-tenders-rfps/releases/download/2026-08-08/{artifact}

The manifest declares SHA-256 hashes for every artifact. The hashes independently recalculated for this report were:

source-health.json       79a0448e095fd440d33123c35d36d3891610086a69ef5a54f4d7f99339fe3ec8
field-completeness.csv   acd6ed02332fcec48bfc34f7bb4b70551353348820f1513e573890e75d4083e0
field-completeness.json  d1a64717b44d7cafffc07fc4e98e68f5e30a033923ac493593e3f507ba6175fe

This Python example recalculates the composition-weighted rates directly from the published JSON:

import json
from urllib.request import urlopen

url = (
    "https://github.com/webtruffle/government-tenders-rfps/"
    "releases/download/2026-08-08/field-completeness.json"
)

with urlopen(url) as response:
    rows = json.load(response)

for field in sorted({row["field"] for row in rows}):
    group = [row for row in rows if row["field"] == field]
    records = sum(row["records"] for row in group)
    present = sum(row["present_records"] for row in group)
    print(field, present, records, round(100 * present / records, 4))

For the stage analysis, stream rolling-observed-history.csv with csv.DictReader and count (source, record_stage) pairs. The file is approximately 208 MB, so it need not be loaded into memory.

Suggested citation:

WebTruffle. “Public Procurement Data Quality Benchmark: August 2026.” Published August 11, 2026. Based on the government-tenders-rfps schema 2.0 edition dated August 8, 2026.

Always include the edition date and schema version. A newer mapper, backfill, source correction, or longer retention window should change the result. Reproducibility means a dated claim can be recreated; it does not mean the rate should remain frozen.

What to measure in the next edition

This baseline exposes the next work clearly:

• publish applicable-field denominators by source, notice type, record stage, and lot or contract grain;
• parse and reconcile TED eForms winner, tender, lot-result, and contract relationships;
• populate and validate authoritative Find a Tender source URLs;
• measure buyer and supplier identifier coverage separately from name presence;
• distinguish tender, request-to-participate, clarification, and other deadline types;
• separate estimated, maximum, award, contract, and transaction value coverage;
• add source-to-normalized sample reconciliation with documented failure classes;
• report first-publication latency only where a comparable source timestamp exists;
• separate SAM.gov first publication from revision and archive replay evidence; and
• publish schema-to-schema movements so mapping improvements are not misread as sudden changes in the underlying portals.

The strongest future chart is not a higher overall percentage. It is a smaller unexplained gap at every applicable source-stage denominator.

The August 2026 procurement-demand snapshot is the first worked analysis built on this evidence. It keeps the source-conditioned taxonomies separate and carries the release, grain, window, and denominator into every category claim.

Frequently asked questions