← BlogAI & Technology

Semantic Search vs Keyword Search: Why It Matters for Tenders

July 25, 2026 · 7 min read

A medical-gas company sets up tender alerts on TED, the EU's Tenders Electronic Daily portal. They use the obvious keywords: "medical gas", "oxygen plant", "PSA generator". For six months the alerts are quiet, so they assume the market is quiet too. Then a competitor wins a hospital contract worth several hundred thousand euros. The buyer had published it as "central pipeline system for respiratory therapy". Three words the company would never have thought to monitor, and the keyword filter let it sail straight past.

This is not a rare failure. It is the defining weakness of how most tender alerts still work. And it is the single clearest reason semantic search beats keyword search for anyone serious about winning public contracts.

How keyword search actually fails

Keyword matching is literal. It looks for the exact strings you typed, or close stems of them, inside a tender's title and description. If the buyer used a different word, you get nothing. Public procurement is the worst possible environment for this, for three reasons:

Buyers describe the same thing in dozens of ways. "Oxygen plant" might appear as "oxygen generation unit", "PSA system", "gas production skid", or "central supply source for medical air". A hospital in Lisbon, a clinic in Nairobi, and a ministry in Riyadh will each phrase it differently.

Language and translation drift. TED publishes in 24 languages, Brazil's PNCP in Portuguese, most Gulf portals in Arabic and English side by side. A keyword that works in one language returns silence in another, even after machine translation mangles the technical term.

CPV and classification codes are inconsistently applied. The EU's Common Procurement Vocabulary is supposed to standardise this, but buyers routinely pick a code that is close enough rather than exact. Filtering on a single CPV code drops real opportunities; filtering on a broad parent code drowns you in noise.

So you face a brutal trade-off. Narrow keywords mean false negatives: real tenders you never see. Broad keywords mean false positives: a daily flood of irrelevant notices you stop reading after a week. Either way, you miss deals.

What semantic search does differently

Semantic search matches on meaning, not spelling. Instead of asking "does this tender contain my exact words", it asks "is this tender about the same thing I care about".

The mechanism is embeddings. An AI model reads a piece of text, your company profile, or a tender notice, and converts it into a long list of numbers (a vector) that represents what the text is actually about. Texts with similar meaning end up close together in this mathematical space, even when they share no words at all.

In practice that means a profile describing "medical oxygen supply systems for hospitals" lands right next to a tender titled "central pipeline system for respiratory therapy", because the model understands the two phrases point at the same real-world thing. No shared keyword required. The same model also knows that "Apple supplier contract" in a fruit-procurement tender is not about computers, something a keyword filter on "Apple" can never figure out.

The result is matching that survives:

Synonyms and jargon: "PSA generator", "pressure swing adsorption unit", and "on-site oxygen production" all match the same profile.

Translation and multilingual sources: the meaning is preserved across languages because the vector represents the concept, not the literal string.

Loose or wrong classification codes: the model reads the description itself, so a mis-tagged CPV code no longer hides the tender from you.

A concrete before-and-after

Take a mid-sized civil-engineering firm that builds water-treatment infrastructure. With keyword alerts on "water treatment plant" across TED, the World Bank's procurement notices, and a few national portals, they catch the obvious tenders and miss a long tail:

"Wastewater reticulation upgrade" in a South African eTenders notice

"Estação de tratamento de águas residuais" published on Portugal's BASE portal

"Effluent management works" in a World Bank project in East Africa

"Reabilitação de ETAR" in Brazil's PNCP

None of those contain the phrase "water treatment plant". A keyword system shows the firm none of them. A semantic system scores all four as strong matches, because each describes work the firm does, in language the firm simply did not predict. Over a year, that gap is the difference between a healthy pipeline and a thin one.

Semantic does not mean keywords are useless

The honest answer is that the best systems combine both, an approach often called hybrid search. Keywords still matter when you need precision on a hard constraint: a specific buyer name, a registration number, a mandatory standard like ISO 13485 or CE marking, a country you are licensed to operate in. These are facts, not concepts, and exact matching is the right tool for facts.

Where semantic search wins is the fuzzy, high-value question of relevance: out of 3,000 notices published on TED today, plus thousands more across Africa, Latin America, the Middle East, and the multilateral banks, which handful is actually about what my company does. Use embeddings to rank by meaning, then keyword filters to enforce the non-negotiables. You get both recall (you stop missing tenders) and precision (you stop drowning in noise).

A few practical signs you have outgrown keyword alerts:

You only find tenders that use your exact internal vocabulary.

You either get almost nothing or get hundreds of irrelevant emails.

You keep hearing about contracts after they closed, from people who saw them described differently.

You operate across more than one language or country.

What this means for your monitoring

The portals are not going to standardise their wording, and buyers are not going to start using your vocabulary. The fragmentation is permanent. The only thing you control is how you read across it.

This is exactly the problem a meaning-based tender feed is built for. Instead of you guessing keywords, the platform learns what your company does from your profile, embeds every new notice from hundreds of sources as it is published, and scores each one against you by meaning. What lands in your inbox is a short, ranked daily list of the tenders that genuinely fit, including the ones described in words you would never have searched for. Less searching, fewer misses, and far less chance that the next "central pipeline system for respiratory therapy" goes to someone else.

Frequently asked questions

What is the difference between semantic and keyword search for tenders?

Keyword search matches literal strings, so it finds notices containing the exact words you predicted and misses those describing the same requirement differently. Semantic search compares meaning using embeddings, so a notice specifying your product in unfamiliar terminology still matches. The practical difference shows up most on technical specifications written without product names.

Does semantic search make keywords useless?

No. Keywords remain useful as hard filters where precision matters: a specific standard, a named platform, a CPV code, a buyer's name. The effective combination is semantic matching to decide relevance and keywords to constrain or verify it, rather than choosing one approach and accepting the other's blind spots.

How much do keyword alerts actually miss?

It depends on how consistently your market describes what you sell, and the more technical the category, the worse keyword coverage gets. The pattern to watch for is a tender you lost or never saw that used none of your terms while describing exactly your product, since that indicates systematic rather than occasional loss.

Share this article

Reference