Checking identity is not enough.
An authenticated agent can still act against the merchant's interest. Knowing who it is without checking its track record does not guarantee the operation ends well.
The score measures readiness, not identity.Trust Score v4.1 · public surface live
See how ready your store is for AI-assisted purchases.
Trust Score v4.1 combines three weighted dimensions with an optional external-assurance bonus and confidence-tier caps. Each response identifies the engine used; some responses fall back to legacy v2. A legacy v2 component breakdown does not explain an overall score produced by v4.1.
This page describes how it is calculated. What that number decides (visibility in agent listings and, below a certain cut-off, automatic suspension of the store) is published with its exact thresholds and your GDPR Article 22 rights in Privacy.
It is a prioritisation signal with context: not a seal, not a customer review, not a guarantee of outcome.
In just over a minute
How the Trust Score helps an agent trust your store
Rafi explains it without jargon: the trust problem between agents and merchants, how a signed, verifiable score works, and what it doesn't do.
What does the Trust Score do between a merchant and an AI agent?
The Trust Score is a signed score that summarizes verifiable signals about a merchant, so an AI agent can decide whether to trust it. Its methodology is public, and any agent can verify the signature without asking our permission. It combines several weighted dimensions and is recalculated periodically. It is a signal that helps decide: it does not guarantee the outcome or replace other checks.
What the video says
Illustrative example · the score on the card is not a real merchant's. The presenter and his voice are AI-generated.
Hi, I'm Rafi, from Trusteed. Commerce with AI agents: an emerging problem, trust between agents and merchants. That's what this video is about.
The problem: an AI agent buys from your store on someone's behalf. It acts for another person, but it can't trust you the way that person would.
What the agent can't do today: call you on the phone, read reviews with human eyes or sense whether you're trustworthy. All it has is what you say about yourself. Or nothing.
The solution: enter the Trust Score.
Step 1 · A signed score. It carries a digital signature: if anyone alters it, it shows.
Step 2 · With a published formula. The full methodology is public. No black boxes.
Step 3 · That any agent can check, on its own, without asking our permission or taking our word for it.
Step 4 · Not made up by the merchant. It combines several weighted dimensions (merchant reliability, agent readiness and evidence from agent purchases), plus an external-assurance bonus of up to 10 points, and states the source of each signal: verified, measured or declared. It is recalculated periodically.
What matters: it is a signal that helps decide. What it does: it summarizes verifiable signals about the merchant and helps the agent decide with more information. What it doesn't do: it doesn't promise everything will go well, and it doesn't replace the other checks a responsible agent makes. We tell you exactly what it proves, and what it doesn't.
Trusteed: trust you can verify, between your store and the agents that buy from it.
Why it exists
In the human channel trust is built with a brand. In the agentic channel, with checkable data.
An agent decides in milliseconds, without seeing your site, reading your reviews or having room to interpret. It needs to know up front which actions it may attempt, which conditions will apply and what proof will remain afterwards. The Trust Score is that upfront reading: the same one for the merchant who publishes it and the agent who queries it.
Reviews are not machine readable.
Stars and opinions describe a past human experience, not the technical ability to answer a delegated action today.
A structured, versioned, queryable signal.With no shared reference, both sides carry the risk.
The merchant assumes the agent will behave; the agent assumes the merchant will respond. Both find out in the same transaction.
One reference both read the same way.How it is measured
Three aspects that help explain how ready your store is.
v4.1 redistributes weight over applicable dimensions. If it cannot produce a value, the public route can fall back to v2. The displayed v4.1 number requires successful signature verification; that does not certify merchant reliability.
PILLARS IN THE RESPONSE
Merchant reliabilityTrack record, continuity and real operation
merchant_reliability_baseAgentic readinessManifests, authentication, limits and policies
agentic_readinessAgentic evidenceOnly when there is a real sample of operations
agentic_evidenceLive Trust Score
demo-store — demo store, synthetic data
57.2/ 100
EdDSA signature verified
PER-SIGNAL QUALITYconfidence: cold_start- measured
- observed on the real system
- verified
- also checked against a third party
- not_applicable
- no sample: excluded from the calculation
The response is signed: the EdDSA signature and the key id travel in the headers (x-jws-signature, x-jws-kid), alongside the methodology version.
Open the demo store's signed responseMETHODOLOGY 4.1.0Trust with context
- Three weighted dimensions: merchant reliability, agentic readiness and agentic evidence
- Confidence-dependent weights, renormalized over applicable dimensions
- Optional external-assurance bonus, subject to evidence gates and tier caps
- Read engines and data_quality; optional statistical extensions are not guaranteed in every response
Read engines.overall and engines.breakdown before interpreting a score. overall uses 0–1; trust_score_v41 uses 0–100 when available. With engines.mixed=true, the breakdown does not explain overall. insufficient_data is missing usable evidence, not a measured zero; the public route can return a degraded v2 fallback.
From signal to score
Four steps between what your store exposes and what an agent reads.
01
Collectllms.txt, mcp.json, .well-known, policies and declared surfaces
available02
CheckEvery signal is verified against the real system, not against the declaration
available03
ContextualiseSample, age and confidence adjust the result
in pilot04
Publish signedScore, range, provenance and version in a single response
availableThe merchant's view
For a merchant it is not just a grade: it is what lets you judge who you open the channel to.
The score is calculated during the diagnosis and reads backwards: every point you miss names a concrete capability that is not operational yet.
01available
HOW IT IS USEDAs a diagnosis before opening the channel.
You learn which control and evidence signals exist in your systems today and which ones you assumed existed. The score does not reward declared intentions.
- Priority by impact, not a filed report
- The difference between declared and checked
- A measurable starting point for the pilot
02in pilot
WHAT IT GIVES YOUVisibility in the channel where nobody looks for your brand.
When an agent compares stores that solve the same need, it picks by complete and current signals. Being measured and signed is the condition for entering that comparison.
- Eligibility and ordering in the agentic channel
- Fewer agent rejections and retries
- Fewer integration hours spent on trial and error
03in pilot
WHAT IT PROTECTSA defensible position when something is questioned.
The score does not win a dispute, but it records which controls were active and with what provenance at the time of the operation.
- Control status with date and version
- Gaps declared as gaps, not hidden
- Read alongside the receipt for the operation
The agent's view
For an agent, picking the wrong store means failing the user who delegated the purchase.
Every tool call costs time, tokens and one chance to fail in front of the user. The Trust Score is queried during planning, before the first attempt is spent.
You prioritise instead of probing
Among several stores selling the same thing, you start with the one whose signals are most complete and recent, not with the first in the index.
Fewer retries after an unexpected rejection.You know which condition applies
Thresholds, human review and supported surfaces are visible before attempting the action, not after the error.
Planning with the condition known.You can justify the choice
The response arrives signed and versioned, so your agent's decision is explainable to the user who delegated it.
An auditable choice, not an opaque heuristic.One signal, two readings
A number the merchant reads as a work plan and the agent as an order of priority.
Merchant and agent views must retain the engine, scale and evidence quality of the response they use. A signed response can mix v4.1 overall and legacy v2 breakdown.
The merchant reads“These pillars are measured, these do not apply yet. This is the next one worth solving.”
The agent reads“This store has complete, recent signals. I start here and expect this condition at checkout.”
Explicit limits
What the Trust Score is not.
It does not certify the merchant, does not guarantee a transaction goes well and does not replace the evidence of a specific operation. It is a prioritisation signal with visible provenance and sample. Presenting it as anything more would make it useless for both sides.
Review limits and evidenceavailable
A signal is always read with its context.GET /trust-score- methodology_version
- 4.1.0
- confidence_level
- by sample and age
- data_quality
- one per pillar
in pilot
The pillars that depend on real operations switch on once there is enough sample in your pilot.
Frequently asked
What people ask before trusting a number.
Can I raise my score without changing anything real?
No. Every pillar is checked against the system, not against what a manifest declares. Declaring a capability that does not answer does not move the score; at best it leaves it marked not applicable.
What if my store is new and has no track record?
With a small sample, Bayesian shrinkage pulls the result towards the mean instead of rewarding a few favourable cases. The score rises as real operations accumulate, not all at once.
Does the score guarantee the transaction will go well?
No. It is a prioritisation signal (catalogue, policies, checkout, evidence), not a guarantee of outcome. The proof of a specific operation is its receipt, not the score.
How often is it recalculated?
Signals decay over time, so an old capture weighs less even if nothing changes. Every signed response states its date and methodology version so you can decide whether it still holds.
Can I compare my score with another store's?
Only by reading the quality of each pillar. Two stores with the same figure may have different pillars measured, and that is explicit in the response.
A number you can show
Calculate your store's Trust Score
The assessment is scoped to one operation, some systems and one metric. You leave with a figure, its provenance and the list of what is worth solving first.