← Back to Guides
AI & Data Analysis
Original Research

Keywords are a metal detector. AI is a geologist.

Keywords found 400 tenders.Humans rejected 67.5% of them.

What 400 real procurement decisions reveal about the five dimensions keyword search can't measure.

18 min read
400 tenders analysed

The keyword trap

400 tenders matched supplier keywords. Only 130 were worth pursuing. The other 270 were rejected - not because the search was broken, but because keywords are one-dimensional.

Search "software development" and you get SaaS builds, embedded firmware, legacy maintenance, and mobile apps. All legitimate matches. Only one might be relevant. The words match. The opportunities don't.

"Management" means nothing

28% approval rate. Same keyword, completely different worlds:

"IT project management services" → Relevant for tech consultancies
"Waste management collection services" → Completely different industry
"Cloud infrastructure management" → Relevant for cloud providers
"Building management maintenance" → Facilities, not IT

Even when the product is right, the tender can be wrong. Already awarded. Deadline passed. Certifications you don't hold. A planning notice, not a live opportunity. Keywords find volume. You need relevance.

5 dimensions keywords can't see

Keywords measure one axis: word overlap. AI evaluates five at once. Any single one can kill an opportunity that keywords said was perfect.

1. Keyword Relevance - the word matches, the product doesn't

Not "does the keyword appear" but "does the thing being procured match what you deliver." Keywords are proxies for relevance, not relevance itself.

Relevance ZoneWhat it meansObserved approval rate
Core productExactly what your company manufactures or delivers65-75%
Adjacent productClose to your offering but not primary40-55%
Niche variantSame category but specialised or outside your range20-35%
Mixed procurementYour product is one lot in a multi-lot tender15-30%
Adjacent categorySame broad sector but fundamentally different product10-20%
IrrelevantKeyword matched but product is entirely different0-2%

Key insight: Subcategory words predict outcomes far better than the keyword itself. "Framework," "maintenance," "transport," or "services" within an otherwise matching tender dramatically lower the approval rate. AI reads these semantic signals. Keyword search doesn't.

2. Strategic Fit - walls you can't see from a search result

A perfect product match can still be impossible to win. Six structural barriers keywords can't detect:

Registration walls

One country in our dataset had 47 tenders with 0% approval. Not because the product was wrong. Because the supplier wasn't registered in the local system.

Certification walls

German, Polish, and French procurement systems each require specific national certifications that take months to obtain. No cert, no bid.

Security clearance walls

Some tenders require industrial security clearance just to access the documents. If you can't read the RFP, you can't bid.

Incumbent advantage

When the existing supplier is entrenched, your bid preparation costs may never justify the low probability of displacement.

Language barriers

As our EU Procurement Paradox guide shows, 96% of European tenders are published in non-English languages. Many require response in that language too.

Local presence requirements

"Local shop within 30 km" or similar constraints that make remote delivery impossible. No amount of keyword matching fixes geography.

3. Timeline - is this even a live opportunity?

Not everything in a tender aggregator is live. Some are planning notices. Some are already awarded. Keywords treat them all the same.

Tender StageApproval RateWhat Keywords Show
Active RFP / Tender Notice~47%Looks identical to other results
Live Contract Notice~32%Looks identical to other results
Planning / Prior Information Notice~27%Looks identical to other results
RFI / Pre-Market Engagement~25%Looks identical to other results
Contract Award Notice (already won)~12%Looks identical to other results

Read that last row. A "Contract Award Notice" with a relevance score of 100 is still 88% likely to be rejected. Someone already won it. Keywords surface it anyway because the words are identical. Only the document type differs.

4. Scope Clarity - are you the headline or lot 6 of 12?

A 7-lot tender mentions your product in lot 4. Is it worth the effort?

You're the headline

  • • Single-lot tender specifically for your product
  • • Title leads with your category
  • • Quantity and value clearly defined
  • • Specific CPV code matches your offering

You're the footnote

  • • Multi-lot tender, your product is 1 of 7+ lots
  • • Generic title: "Supply of equipment and clothing"
  • • Your category mentioned as "includes" or "one lot for"
  • • Dozens of CPV codes covering diverse categories

5. Win Probability - can you actually compete?

Some buyers are better fits than others, regardless of product match. Competition level, customer alignment, and procedure type all affect whether you have a real shot.

Strong win signals

Open procedures, SME-friendly frameworks, no incumbent flags, buyer history of awarding to new suppliers

Neutral signals

Restricted procedures, medium competition levels, framework call-offs with multiple pre-qualified suppliers

Weak win signals

Known incumbent advantage, bid bonds or advance payment requirements, specialised agencies with established supply chains

They compound. No single dimension decides everything. But a difficult market plus a tight deadline is far worse than either alone.

What 400 tenders taught us

270 rejected. 130 approved. European, North American, and international portals. Not predictions - observed outcomes.

Finding 1: "High Priority" doesn't mean "pursue"

39 tenders flagged "High Priority" by keyword screening. Every surface signal said approve. All 39 rejected. Here's why:

8
Duplicates of already-approved tenders
8
Structural market barriers (registration, certification)
7
Niche products outside the supplier's actual range
5
Planning notices or RFIs, not live tenders
5
Impossible deadlines combined with certification needs
6
Genuinely unpredictable. Required human judgment.

The cost: Without deeper analysis, these 39 tenders would have consumed days of bid prep before being discarded.

Finding 2: A perfect relevance score still fails half the time

Even at a relevance score of 100, only 54% were pursued. The rest failed on dimensions keywords never measured:

Relevance ScoreApproval RateWhat this tells you
90-10052%Best rate, still a coin flip on other dimensions
80-8942%Often rejected on certification, deadline, or niche product
70-7950%True coin flip. Other dimensions decide everything.
50-6919%Some hidden gems, mostly noise
30-494.5%Tangential connection at best
0-29~0%Fundamentally irrelevant

Finding 3: Every tender falls into one of three zones

Score across all five dimensions and tenders naturally cluster:

APPROVE zone (~30% of tenders)

All five dimensions align. Core product match, no market barriers, live tender, clear scope, reasonable win probability.

These move straight to bid preparation.

REJECT zone (~50% of tenders)

One killer flaw. Wrong product, hard market barrier, already awarded, or out of scope. One dimension is enough.

Filtered out before they reach your team.

UNCERTAIN zone (~20% of tenders)

Mixed signals. Two mediocre scores combining, borderline market access, unclear scope. This is where edge cases live.

Flagged for human review with clear reasoning.

Honestly: 15-20% of tenders will always need a human call. Team bandwidth, pipeline commitments, relationships. The job of scoring is to handle the 80% that are predictable, and flag the rest so your team can decide quickly.

From keywords to intelligence

Keyword alerts are flat. "Match" or "no match." AI scores five axes at once, and the shape of that score tells you more than any keyword ever could.

Same search. Two different verdicts.

Both tenders matched the same keywords. Only one is worth pursuing:

Tender A: Keyword Match = 95%

Keywords say "perfect match." AI says "reject."

Already awarded. Wrong market. No chance.

Tender B: Keyword Match = 78%

Keywords say "partial match." AI says "pursue."

Live RFP. Strong fit. Winnable.

The maths

Keywords only:

  • • 400 tenders land on your desk
  • • 7 minutes each to review manually
  • • ~47 hours of analyst time
  • • 270 of those hours wasted

With AI scoring:

  • • 400 tenders scored in seconds
  • • 130 relevant ones surfaced
  • • 20 flagged for human review
  • • ~15 hours of focused bid work

5 questions to ask before every bid

AI or manual, answer these before committing resources. Each maps to a dimension above.

1

Is this actually what we sell?

Not "does our keyword appear" but "is this our core offering, an adjacent product, or a tangential mention?" Check sub-categories, not just titles.

2

Can we even get in the door?

Registration requirements, certification walls, security clearances, language mandates, local presence rules. A perfect product match in an inaccessible market is worthless.

3

Is this still live?

Active RFP? Planning notice? RFI for market research? Award notice for a contract someone else already won? Verify before you invest a single hour.

4

Are we the headline or lot 6 of 12?

Single-lot procurement for your exact product, or a massive tender where your category is buried? The bid effort is the same. The probability isn't.

5

Can we realistically win?

Competition level, incumbent risk, customer alignment, and whether the timeline gives you enough room to prepare a serious bid.

TenderStria does this automatically

Every tender scored across all five dimensions - keyword relevance, strategic fit, timeline, scope clarity, and win probability. Instead of answering these questions for hundreds of tenders yourself, you get a prioritised shortlist with clear reasoning. Talk to our team or read our guide to tender discovery.

Stop reviewing tenders that were never worth your time

400 tenders down to 130. That's what happens when you score on five dimensions instead of one.