A useful tender search needs a test set, not just a reassuring result count. These three public records expose different ways a shortlist can go wrong: an unexpected classification, a procurement stage that is not an open tender, and geography attached to the wrong organisation.
This is an original, hand-checked source-reading exercise as of 13 September 2026. The records were deliberately selected to illustrate failure modes. They are not a random sample, a recall benchmark, a market-size estimate or evidence of Duke's retrieval performance.
1. The subject is useful, but the classification is broader than expected
Maastricht's Learn@Maastricht notice, TN-532939, 19 June 2025, includes a data-driven working lot involving data and AI training, under education and training classification 80000000.
Test to run: would your AI-services search retrieve a training requirement, or does it only search software classifications? Whether this is a positive example depends on your offer. An AI implementation supplier and a training provider should not automatically make the same decision.
Read the lot's deliverables before adding keywords. This historical notice is a test record, not a claim that the competition is still open.
2. A relevant buyer signal is not a confirmed competition
IVO Rechtspraak's knowledge-management market consultation, TN555193, document dated 3 November 2025, explores a possible procurement. The document does not commit the buyer to launching a tender.
Test to run: can your workflow retain the buyer signal without counting it as an open bid? Put it in research, record the source date and check for later publications before changing its status. Use the Dutch buyer-watchlist guide to define the next action.
3. The country appears, but not as the delivery market
TED 163844-2026 concerns Lithuanian snow-clearing services. Its electronic sender has an address in Athens.
Test to run: does a full-text Greece search admit the record solely because of the sender? Inspect buyer country and place of performance separately. Retain this as a negative example for a Greece-only delivery search, not as evidence that the source is wrong.
The Greece monitoring guide develops the location and stage checks.
Build your own small regression set
- Select known relevant notices from your actual offer and delivery area. Include mixed lots and unusually broad classifications.
- Add tempting but unsuitable records: historical awards, wrong service models, wrong locations and closed or restricted routes.
- Record the expected decision and its source evidence before testing a revised query.
- Run the saved query, note which known records appear and inspect any changed decisions.
- Keep the query version and exclusion reasons in the monitoring log.
A small set can reveal a specific mistake. It cannot establish completeness across a country or measure recall unless the relevant universe is independently known. Count unique notice-and-lot identities rather than duplicate publications as separate opportunities.
What this changes in the sales workflow
Keep discovery, qualification and buyer research separate. A broad search can be useful when the review explains what was rejected. A narrow search can look clean while silently missing relevant work.
Bring your positive and negative examples to a Duke workflow demonstration, using the platform evaluation worksheet. Ask to see the source record and the resulting decision. A convincing demo should explain the misses and uncertainties as well as the matches.