How AI Relevance Scoring Beats Keyword Alerts for Tender Discovery

The keyword-alert problem

Every tender aggregator lets you set keyword alerts. You type "safety software" and get emailed every tender containing that exact phrase. Simple, cheap to implement, and — for anyone with a non-trivial capability space — completely broken as a filter.

The problems compound:

• Vocabulary mismatch. A buyer procuring the exact thing you sell may write "EHS management system" or "safety compliance platform" instead of "safety software". Your keyword misses.
• False positives. "Safety software training for construction site managers" matches your keyword but is training services, not software. You waste 30 minutes reading it to figure that out.
• Context loss. "Safety software" appears in a facilities-cleaning tender because the FM company mentions their internal safety software as part of company overview. Match, but irrelevant.
• Language fragmentation. To catch French tenders you add "logiciel de sécurité", German "Sicherheitssoftware", etc. Each language multiplies your maintained keyword list.
• Volume overload. A broad keyword returns 200 tenders/week; a narrow one misses relevant ones. There's no dial that solves both.

How AI relevance scoring works

Instead of matching keywords, an LLM (Large Language Model) reads each tender notice's title + description + buyer + category and compares it against a plain-English description of your business — what you sell, who you sell to, what to skip. The output is a relevance score from 0 to 1 with a brief rationale.

Example: your business description says "We sell HSE (Health, Safety, Environment) management software to oil-and-gas and manufacturing companies. Relevant tenders: safety software, permit-to-work systems, incident reporting platforms. NOT relevant: hardware, PPE, training-only, construction services."

• Tender A: "Purchase of contractor safety compliance tracking platform for offshore facilities" → score 0.91, rationale: safety software for oil-and-gas, exact fit.
• Tender B: "Provision of safety training and workshops for site managers" → score 0.18, rationale: training-only, explicitly excluded.
• Tender C: "Supply of high-visibility safety vests for road workers" → score 0.02, rationale: PPE hardware, not software.

The three tenders above would all match a "safety" keyword alert. Only Tender A actually deserves attention.

Real-world impact

From TenderYeti's own customer data (medians across accounts with 6+ months of history):

• Notice volume: ~200/week per account matching keywords.
• AI-scored ≥0.7 (relevant): ~12/week.
• Time saved: the customer reads 12 curated notices instead of 200 keyword hits — roughly 4 hours/week to 20 minutes/week.
• Miss rate: lower than keyword-only, because the AI catches vocabulary variants a static keyword list would miss.

The tradeoffs — honest version

AI scoring is not free of downsides. Anyone selling it as pure upgrade is over-promising. Real tradeoffs:

• Cold-start cost. Your business description needs to be well-written — a couple of paragraphs, not a keyword list. First 15 minutes of setup matter.
• Scoring latency. LLM inference adds seconds per tender. High-volume portals may see 5-30 minute delays between publication and your notification. Keyword matching is instant.
• Compute cost. LLM calls cost money. Aggregators either pass this through in pricing or absorb via smaller models with slightly lower scoring quality.
• Occasional weird scores. LLMs are pattern matchers; a genuinely edge-case tender with unusual phrasing may score 0.5 when a human would say 0.9 or 0.1. Weekly review of your borderline scores catches drift.
• Explainability. A keyword hit is trivially explainable ("matched 'safety software'"). An AI score with a rationale sentence is more useful but harder to audit at scale.

When keyword alerts still win

• Narrow, unambiguous product categories. If your product has a unique brand name or a specific CPV code that maps 1:1, keywords work fine and add no LLM cost.
• Compliance / brand monitoring. Watching for your own brand name mentioned in tenders — pure keyword match.
• Very high-volume auto-processing. If you're routing 10,000+ tenders/day through automated workflows, keyword matching + rules keeps latency and cost predictable.

How TenderYeti implements it

Two-tier model: a fast primary scorer runs on every tender, and a slower more-capable model reviews borderline scores (0.4-0.7) for accuracy. Your business description is stored once in your tenant configuration; every tender across every portal you're subscribed to gets scored against that description independently.

Multilingual handling: tenders in French, German, Spanish, Arabic etc. are auto-translated for the scoring model. You get the original + translated title in your dashboard.

You can adjust your relevance threshold — the default is 0.5, meaning notices scoring 0.5+ make it to your digest. Bid managers who want more coverage set it to 0.3; bid managers who only want high-confidence matches set it to 0.7.

Where TenderYeti fits in

See TenderYeti live — 14-day free trial, no card required, AI scoring runs from your first tender. Or browse portal coverage to see the source scope.

Frequently asked questions