A keyword alert misses 48% of relevant UK public sector contracts

How many contracts does each keyword method miss?

Share of relevant contracts awarded in the 12 months to 16 October 2026 that each method would not have caught. The last column is a first estimate of the contracts the expert list sent that were not relevant.

Misses: observed shares, 12 months to 16 October 2026. False alerts: contracts, 12 months to 27 October 2026.

Which keywords did the study test?

For each service, the one keyword a supplier would type into a portal alert, and an expert list that adds the wording buyers often use. Open a service to see its phrases.

False alerts: contracts, 12 months to 27 October 2026.

What counts as a missed contract?

A contract is missed when it is really the service and none of the alert's words appear in its notice. We searched the title and the first 4,000 characters of the description, the way portal alerts work.

Words had to match whole, ignoring case, with plurals allowed. So "laptop" finds "laptops", but "roofing" does not find "roof repairs".

Why do expert keyword lists still miss contracts?

Buyers describe the same work in many ways, and no list can hold them all. Many notices also name a framework, a lot number, a product or a building rather than the service.

Recent contracts that neither list caught include "Purchase of HP ZBook" (laptops), "New Lead Box Gutters and Reinstating Slates" (roofing) and "CA18094 - CPC /JE/01/2023: ICT Hardware & Peripherals Equipment Framework Agreement" (laptops).

Do keyword alerts send contracts that aren't relevant?

Yes, often. A follow-up review read 15,495 contracts that the alerts on 583 BidSkim service pages sent, and judged whether each was really the service. It is a first estimate, from one reviewer per contract, so we publish a figure only for the 341 pages whose sample passed a check.

Where the figure is published, a one-word alert sent a median 63 contracts in 12 months, about 26 of them not relevant. An expert list sent a median 123, about 55 not relevant.

The word "website" sent 1,686 contracts, about 1,400 of them not website work. The words "cleaning" and "apprenticeship" sent 833 and 575, about 370 and 460 of them not relevant.

Reviewers read only contracts BidSkim had not tagged as the service, and work that was partly the service counted as relevant. A page shows a figure when at least 20 of its contracts were read, at least half of them were not relevant and few were unclear.

How were the contracts judged?

AI reviewers (Claude) read 4,665 awards, drawn as a stratified sample, against written rules for each service. They could not see which keyword method had found each one.

Heavily weighted awards were judged twice by separate reviewers, who agreed 97% of the time, with a third judgement where they disagreed. The results were then weighted back to all 88,365 awards.

Method

• All 88,365 UK public sector contracts awarded in the 12 months to 16 October 2026, from Find a Tender, Contracts Finder, Public Contracts Scotland, Sell2Wales and eTendersNI.
• Two keyword methods per service: the one keyword a supplier would type into a portal alert, and an expert list of 5 to 16 phrases including common buyer wording.
• Each method was matched on the notice title and the first 4,000 characters of its description, whole words, ignoring case and plurals.
• A stratified sample of 4,665 awards was judged by AI reviewers (Claude) for whether it was really the service, blind to which method found it, against written rules per service. Heavily weighted awards were judged twice (97% agreement), with a third judgement on disagreements.
• Shares are observed shares of the relevant contracts, weighted back to the full set of awards. Small services have wide upper bounds.
• False alerts come from a separate review of 84,866 contracts awarded in the 12 months to 27 October 2026. Up to 20 contracts per keyword and 25 per expert list were read, each judged once by one AI reviewer. The figure is the contracts sent outside the service times the share judged not relevant. The CSV gives its 90% upper bound.

Questions