Stop missing 71% of your opportunities

Why Activity Search is 3x better

Tender portal searches missed 71% of relevant opportunities in our five-country test. We found 95% of them.

more relevant opportunities vs keyword searches on live portals

fewer missed tenders: portal search missed 70.8%, ours missed 4.7%

relevant open notices across 5 portals and 100 searches

of your opportunities show up on our first page vs 19% on other portals

Understanding and fixing the problem

Keywords sometimes don't align.

Keyword searches WILL misses most of what is relevant to you.
In our study keywords miss 71% of the relevant opportunities.
This doesn't just happen occasionally either, it happens most of the time.
Not because you're searching for the wrong thing, just the wrong words.
When you search for the activity rather than the word, you find more.

Recall and ranking (the science bit).

We don't just look at one notice, we look at every notice on each search.
If one notice's information is "adjacent" to another, we can spot patterns.
So when you search for "training", we also know you could mean "skills".
And so we build a search for what you need and some adjacent terms too.
Finally we filter out as much noise as we can, in the blink of an eye.

The proof

5 national portals, 100 real business needs, identical document universes on both sides, independent relevance grading, every search result auditable.

Portal	Universe (open notices)	Found: ours	Found: portal search	Advantage	Ranking: ours	Ranking: portal search	Ranking improvement
TED (EU) ¹	8,816	93.2%	9.0%	10.4×	77.7%	5.4%	14.4×
SAM.gov (US) ²	2,485	97.7%	17.4%	5.6×	51.1%	7.8%	6.6×
bund.de (DE) ³	11,813	96.1%	20.5%	4.7×	55.5%	7.3%	7.6×
Doffin (NO) ⁴	619	99.8%	39.0%	2.6×	72.2%	15.2%	4.8×
Find a Tender (UK) ⁵	817	89.8%	59.9%	1.5×	55.8%	21.6%	2.6×
Average, 5 portals, equal weight		95.3%	29.2%	3.3×	62.5%	11.5%	5.5×

"Found" is the share of all relevant open tenders each search returned within its first 500 results, on an identical set of notices for both sides. "Ranking" is average precision, a 0 to 100% score of how high the relevant tenders sit in the results. Advantage and ranking improvement are ours divided by the portal's. The average row divides average by average; it is not an average of the per-portal ratios. Tested 2 Sep 2026. Download the data (CSV).

1. TED: our figure carries a known 19% store-coverage handicap. 1,677 of the 8,816 notices in the comparable universe are known to be absent from our store (documented data loss, deliberately kept in the universe). Fixing it would raise our number, not the portal's.
2. SAM.gov: the universe is Solicitation-type notices with an open response date, the class the connector is designed to ingest. Pre-solicitation notice types were excluded on both sides.
3. bund.de: the universe is notices with an offer deadline (Angebotsfrist) on or after the test date. Deadline-less pre-announcements were excluded on both sides.
4. Doffin: a possible Norwegian query-localisation bias may understate what a Norwegian-language user would find with the portal's own search. Our number is unaffected.
5. Find a Tender: figures use notice-id identity. Our store holds some Find a Tender notices twice under two OCID prefixes, so matching on the notice id credits both copies (88.4% on exact-OCID identity, 89.8% on notice-id identity). Ranking is quoted on the exact-OCID basis and slightly understates ours.

You can only bid on what you find

Activity Search looks deeper across our full dataset to find real patterns.

Find items that don't match your keywords
Search using a single URL
More relevant results on the first page
Finds tenders in any language

Try it now

How we ran the test

The steps we took to complete this research, broken down across portals.

1. 100 search inputs, for real business needs

Plain descriptions of what a supplier does: "we install solar PV", "we provide penetration testing", and 96 more.

2. Built each portal's census of open notices

Every notice with a deadline on or after the test date, 2 September 2026, taken from the portal's own API.

3. Restricted both sides to the same universe

Only notice classes both systems ingest. Ingestion gaps count against neither side.

4. Ran both searches for every need

Our activity search (top 500) and the portal's own keyword search, using the best query a diligent user of that portal would type.

5. Had an independent judge grade every candidate

A large language model graded each notice's title, buyer, CPV codes and description for relevance. Only top-grade notices count. The judge does not know which system found what.

6. Scored recall, precision and ranking quality

At depths 20 to 500. Every per-need result is an auditable JSON record: which notices were relevant, and where each side ranked them.

Portal by portal

The real numbers for each portal, including the one where we win by least.

TED (EU): 10.4×

Activity Search found 93.2% of relevant open notices.
TED's own search found 9.0%.
Universe: 8,816 open notices.
Ranking quality 0.777 vs 0.054.

TED's expert search found 9% of relevant notices and stopped. We found 93%, despite a documented 19% data handicap on our side.

Caveat: 1,677 of the 8,816 notices are known to be absent from our store and were kept in the universe. Fixing that raises our number, not the portal's.

SAM.gov (US): 5.6×

Activity Search found 97.7% of relevant open solicitations. SAM.gov's own search found 17.4%. Universe: 2,485 open notices. Ranking quality 0.511 vs 0.078.

Caveat: the universe is Solicitation-type notices with an open response date. Pre-solicitation notice types were excluded on both sides.

bund.de (DE): 4.7×

Activity Search found 96.1% of relevant open notices. bund.de's own search found 20.5%. Universe: 11,813 open notices, the largest in the test. Ranking quality 0.555 vs 0.073.

Caveat: the universe is notices with an offer deadline on or after the test date. Deadline-less pre-announcements were excluded on both sides.

Doffin (NO): 2.6×

Activity Search found 99.8% of relevant open notices. Doffin's own search found 39.0%. Universe: 619 open notices. Ranking quality 0.722 vs 0.152.

Caveat: a possible Norwegian query-localisation bias may understate what a Norwegian-language user would find with the portal's own search. Our number is unaffected.

Find a Tender (UK): 1.5×

Activity Search found 89.8% of relevant open notices. Find a Tender's own search found 59.9%. Universe: 817 open notices. Ranking quality 0.558 vs 0.216.

Find a Tender has genuinely good keyword search, the best we tested. We still found half again as many relevant tenders (90% vs 60%) and ranked them 2.6× better.

Caveat: figures use notice-id identity, because our store holds some notices twice under two OCID prefixes (88.4% on exact-OCID identity, 89.8% on notice-id identity).

Questions

Activity Search FAQ

This is just the starting point, we're going to keep looking at different search engines and adding their data to our analysis, you can expect major sources in Europe and Latin America to be added soon.

A relevant open tender is one an independent judge graded top marks for the business need, out of everything either search returned. The truth set is built from the union of both systems' results, so each side is measured against the same target. A tender neither side returned is not counted, and that limits both sides equally.

A pinned large language model, run at zero temperature with a fixed rubric, graded every candidate on the notice's own title and description. It never saw which system found a notice, and each notice was graded once and reused wherever it appeared.

Yes. Every per-need result is a JSON record showing which notices were judged relevant and where each side ranked them, alongside the fairness conditions and the universe definitions. Ask us and we will send the bundle.

Because the portals' own search varies a lot. Find a Tender's keyword search found 59.9% of relevant open tenders, so our lead there is 1.5×. TED's expert search found 9.0%, so our lead there is 10.4×. The headline 3.3× divides our average recall by the portals' average recall across all five; it is not an average of those per-portal ratios.

Try Activity Search

See the difference now

Try for free Book a demo

Activity Search - finding the missing opportunities.