Two tools monitoring the same procurement portals, covering the same 5,000 daily contract notices, using the same profile of your business. And one of them shows you 40 relevant contracts this week. The other shows you 8. That gap isn't about coverage. It's about matching — specifically whether the tool is doing keyword matching or semantic AI. Here's what the difference actually is.
How Keyword Matching Works
Keyword matching is a word-level operation. You specify a list of terms — say, 'cybersecurity', 'penetration testing', 'information security', 'ISO 27001' — and the system scans contract titles and descriptions for those exact strings. Match found: you get an alert. No match: you don't.
Most keyword tools support partial matching (finding 'cyber' within 'cybersecurity'), boolean operators (AND/OR/NOT), and some phrase matching. The good ones let you search the full specification text rather than just the title. But the fundamental operation is the same: compare your words against the contract's words.
The Four Ways Keyword Matching Fails You
1. Missed synonyms
'Information assurance', 'cyber resilience', 'network security', 'security operations' — all terms that might describe work you do. A keyword system only catches the ones in your list. The contract titled 'Digital Security Improvement Programme' might be exactly your work. If 'Digital Security' isn't in your keyword list, you won't see it.
2. Missed context
Keywords don't understand context. A contract mentioning 'security' might be security guarding, food security, or cybersecurity. If you're a cybersecurity firm and 'security' is in your keywords, you'll get all three. That's noise. The system has no way of distinguishing between them without reading the surrounding context.
3. False positives from keyword overload
The natural response to missing contracts is adding more keywords. But every keyword you add increases false positives. By the time your list has 50 terms, you're getting hundreds of irrelevant alerts per week. Your team stops reading them properly. The alerts become a formality rather than a tool. You've solved the coverage problem by creating a noise problem.
4. Key terms buried in appendices
Many public sector contract notices include a summary description but reference detailed specifications in attached documents. The summary might say 'IT services' while the actual specification makes clear it's a complex data science project. Keyword matching against the summary misses this entirely. You'd need to read every attachment to catch it — which defeats the purpose of automated alerts.
How Semantic Matching Works
Semantic matching uses vector embeddings — a technique from natural language processing that converts text into numerical representations called vectors. Words and phrases that are conceptually similar end up close together in this mathematical space. 'Cybersecurity', 'information security', and 'cyber resilience' are close together. 'Cybersecurity' and 'catering services' are far apart.
When a semantic system evaluates a contract against your profile, it's measuring the distance between vectors — how similar the contract's content is to your business description in semantic space. It doesn't need exact word matches. It understands that 'cloud infrastructure management' is relevant to an AWS partner even if 'AWS' doesn't appear in the notice.
In Plain Terms
The practical difference: keyword matching optimises for recall (don't miss anything with the right words). Semantic AI optimises for both recall and precision — finding more of the right contracts while filtering out more of the wrong ones.
The Practical Difference in Precision and Recall
Recall is the percentage of relevant contracts you actually find. Precision is the percentage of your alerts that are genuinely relevant. Keyword matching can achieve reasonable recall if your keyword list is comprehensive, but precision suffers as the list grows. Semantic AI improves precision without sacrificing recall — you get more of the right contracts and fewer of the wrong ones.
For a business getting 20 relevant contract alerts per day through keyword matching, of which maybe 8 are genuinely worth reading — that's 60% noise. With semantic matching, the same monitoring might surface 25 relevant contracts per day with 18 genuinely worth reading. That's a different conversion rate on your team's time.
Questions to Ask Any Tool Claiming to Use AI
'AI-powered' has become a standard claim. Most tools that make it are doing something that qualifies — but the difference between basic NLP and genuine semantic matching is significant. Ask these questions:
- •Does matching work on the full specification text, or just the title and CPV codes?
- •Can the system identify contracts where my key terms don't appear but the requirements match my work?
- •How is my business profile represented — as a keyword list I define, or as a semantic profile built from my description?
- •What's the false positive rate — what percentage of alerts are genuinely relevant?
- •Can I see a specific example of a contract that was matched semantically rather than by keyword?
If the answers are vague, or if the tool can't show you a concrete example of a semantic match, treat it as keyword matching with marketing language on top. That's not a reason not to use it — keyword matching has its place — but it's worth knowing what you're actually getting.