Resources

Understanding the Signal Score

How five LLM responses get compressed into a single 0-100 number, with worked examples and the edge cases the score is built to handle.

The one-paragraph summary

The Reevol Signal Score is a 0-100 measurement of how AI engines perceive a B2B supplier. It is computed by running three buyer-style queries against multiple LLMs in multiple buyer languages, parsing each response into a structured shape, scoring five separate dimensions (AI Recommendation Rate, Identity and Presence, Product Clarity, Buyer Trust, Digital Footprint), and combining those with locked weights into the final number. The public score is a three-run moving average to damp out single-run jitter. Higher is better. The score ranges from Dark Signal (0-39) to Strong Signal (80-100).

Why five dimensions, not one

A single number is easy to read but hides the actionable detail. A supplier with a total of 64 (Moderate Signal) can have wildly different underlying shapes. One 64 might look like 80 / 75 / 70 / 50 / 35: strong AI recall and product clarity, but a poor digital footprint dragging the average down. Another 64 might look like 50 / 60 / 65 / 65 / 80: solid trust signals and a great website, but weak AI recall. The two suppliers face completely different work to improve.

The five-dimension breakdown is what makes the score useful for suppliers (who need to know which lever to pull) and what makes it useful for buyers (who care about Buyer Trust and AI Recommendation more than Digital Footprint per se). The total is for headlines; the breakdown is for decisions.

Why these specific weights

The weights are 30% AI Recommendation Rate, 20% Identity & Presence, 20% Product Clarity, 15% Buyer Trust, 15% Digital Footprint. They were chosen and locked by the Reevol team in May 2026 against a benchmark of which dimensions correlate most strongly with actual procurement-pipeline outcomes.

AI Recommendation Rate carries the most weight because it is the most direct measurement of the commercial outcome we are trying to capture: when buyers ask an AI engine for category recommendations, does this supplier surface? Identity and Product Clarity together account for 40% because those two dimensions form the "AI footprint" that downstream queries depend on: a model that does not recognize a supplier or cannot describe what they make has no basis to recommend them.

Buyer Trust at 15% is intentionally weighted below the AI-surface dimensions because it is most subject to LLM hallucination. Models will sometimes invent certifications or assert export histories with no source. Letting it dominate the score would over-reward suppliers with persuasive marketing copy. The weight is non-zero because real, verified certifications still matter, but it is not the biggest lever. Digital Footprint at 15% closes the loop: it is the deterministic web-presence check that makes everything else possible.

A worked example

Consider a Vietnamese textile supplier with a Tier 2 audit (two LLMs, English only, three prompt types). Six responses come back from the audit. After parsing:

• 5 of 6 responses mentioned the company by name with high confidence.
• 3 of 6 responses listed at least one product matching the Atlas catalog.
• 2 of 6 responses cited a certification (ISO 9001).
• 0 of 6 responses flagged a negative signal.
• The supplier's website resolves, has Open Graph tags, no Organization schema, LinkedIn exists, one trade-directory listing found.

From these inputs the scorers produce roughly:

• AI Recommendation: 60. Found across both LLMs with positive/neutral sentiment, but only one out of six responses passed the high-accuracy threshold.
• Identity & Presence: 70. Strong name recognition, consistent across responses, found in two LLMs.
• Product Clarity: 55. Products are named but only half the time; no export language; no misclassifications.
• Buyer Trust: 55. One certification cited consistently, no negatives, but only two of six responses mentioned trust signals at all.
• Digital Footprint: 80. Website crawlable, OG tags present, LinkedIn, one directory listing.

Weighted total: 60×0.30 + 70×0.20 + 55×0.20 + 55×0.15 + 80×0.15 = 64. Moderate Signal. The improvement priorities are obvious from the breakdown: lift Product Clarity by publishing a structured catalog page, and lift Buyer Trust by adding a certification-proof page. Those two changes alone would push the supplier into Strong Signal territory.

The three-run moving average

LLMs are stochastic. The same query, run twice, often returns different but equally-plausible responses. Run three times, the scores can vary by 5 to 10 points without anything having changed about the supplier. To avoid telling suppliers their score "dropped" when nothing actually happened, the public Signal Score is the average of the three most recent audits.

The trade-off is responsiveness. A supplier who makes a major content change has to wait two or three refresh cycles before the public score fully reflects the improvement. Claimed suppliers see the raw single-run score in their Compass dashboard alongside the smoothed public score, so they can debug short-term changes immediately without waiting for the average to catch up.

What the score does not measure

The Signal Score measures AI perception. It does not measure:

• Financial solvency or creditworthiness.
• Product quality, lead time, or pricing.
• Compliance with specific buyer requirements (REACH, conflict-mineral, ESG).
• Whether the supplier has been sanctioned or appears on a blocklist.
• Whether the supplier is willing or able to ship to your country.

A high Signal Score is a necessary but insufficient condition for a good supplier relationship. A supplier with a Strong Signal Score who fails on any of the above is still the wrong supplier. The Score is best used as a filter at the top of the funnel, then supplemented by direct due diligence on the dimensions it does not cover.

Edge cases the score handles

Name collisions. When parsed responses describe a different entity (a different city, a different product line), the parser detects the mismatch via the accuracy score and the affected response does not boost the dimension scores. Pure name-collision responses score 0 on accuracy and effectively drop out of the dimension averages.

Brand-new suppliers. A supplier whose website went live last month will not be in any model's training data and will not yet have a Signal Score above Dark. This is correct, not a flaw. Suppliers can earn their way out of Dark Signal by adding crawlable content and waiting for the next model refresh cycle (typically 3-9 months between major models).

Suppliers with negative news. If recent web content includes fraud reports, customer disputes, or sanctions, the parser captures those as negative signals and they pull Buyer Trust down. Negative content from competitors (which exists in some categories) is harder for the parser to disambiguate; claimed suppliers can flag specific responses for re-audit.

Multilingual asymmetry. A Chinese supplier whose Chinese-language web presence is strong but whose English presence is weak will see lower scores in English-language audits. The Score does not try to "average across languages" in a way that hides this asymmetry, because the asymmetry is itself useful information for the supplier and for buyers. Tier 1 suppliers run audits in four buyer languages so the asymmetry is visible directly.

How often the score is refreshed

Tier 1 suppliers refresh monthly with the full four-language matrix. Tier 2 suppliers refresh quarterly in English only. Claimed Tier 2 suppliers refresh monthly in English plus one secondary buyer language. Tier 3 (stub) pages refresh only when a visitor triggers an audit via the email-gated refresh button. Anyone can request an ad-hoc refresh, subject to per-tier rate limits.

A note on transparency

Every weight and rule that produces the Signal Score is published on the Methodology page. The prompts, the LLMs, the score bands, the refresh cadence: all visible. We do this because the score is only useful if buyers trust it and suppliers understand it. A hidden algorithm produces ad clicks; a transparent algorithm produces better suppliers and better decisions.