What Trust Signals Do AI Agents Use?

AI agents ranking merchants do not “feel” brand prestige the way humans do; they approximate trust from checkable signals: identity verification, official licences, data freshness, corroborating evidence, review quality, and composite scores. In the emerging agent-to-agent (A2A) economy—where a buyer’s assistant negotiates with a merchant’s callable profile—Generative Engine Optimization (GEO) still matters for discovery, but trust signals decide who gets recommended and invoked. ChatGPT, Perplexity, Gemini, DeepSeek, and custom tool-using agents all benefit from the same underlying pattern: prefer entities that are verifiable, consistent, and operationally ready.

This article maps the trust stack merchants can actually publish, how it relates to E-E-A-T, and how Trust Score fits without becoming a black-box vanity metric.

From Star Ratings to Machine-Checkable Trust

Human marketplaces trained merchants to chase average stars. Agents need more:

| Human heuristic | Agent-oriented signal | Why it differs |

| --- | --- | --- |

| “4.8 stars” | Review provenance + proof | Averages hide fraud and staleness |

| “Looks premium” | Verification level | UI polish is not identity |

| “Been around forever” | Registry / licence history | Tenure needs identifiers |

| “Influencer said so” | Corroborated citations | Mentions help GEO; proof helps trust |

| “Site feels legit” | Freshness + consistent NAP | Agents fetch fields, not vibes |

E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) remains a useful human-rater vocabulary. For agents, translate it into fields and proofs.

The Trust Stack Agents Implicitly Score

1. Identity signals

•Stable public ID (`/t/{id}`)

•Consistent legal / trade names across languages

•Claimed ownership of domain / phone / registry record

•Organization vs personal operator clarity

Weak identity produces duplicate entities and split reputation—fatal for ranking.

2. Licence and official registry signals

Especially for food, healthcare-adjacent, finance, travel, and professional services:

•Licence number + issuing authority

•Status (valid / suspended / expired)

•Premises address matching operations

•Claim linkage from official list → merchant-controlled profile

Hong Kong FEHD-style open data is a strong provenance layer when claimed through [/registry](/registry). Unclaimed registry rows help GEO slightly; claimed + enriched rows help trust and invoke.

3. Freshness signals

•Last updated timestamps on hours, menu, fees

•Recent verification checks

•Change logs for material fields

•Absence of obvious staleness (holiday hours from last year)

RAG systems and tool agents both punish stale chunks. Freshness is a trust signal *and* an accuracy signal.

4. Completeness signals

Sparse profiles force inference. Complete profiles reduce hallucination risk:

•Hours, channels, service area

•Offer list with constraints

•Booking rules

•Languages

•Explicit “we do not offer X”

Completeness is not keyword stuffing; it is decision-relevant coverage.

5. Proof-based review signals

Agents should prefer reviews that can be evidenced (receipt, booking ID, photo proof policies) over anonymous five-star bursts. Industry reports often suggest review authenticity problems inflate marketplace scores; cautious platforms weight proof and recency more than raw averages.

Publish:

•Policies for eligible reviews

•Links to verifiable transactions when privacy allows

•Responses to disputes (trustworthy operators engage)

6. Corroboration / GEO citation signals

Being mentioned by reputable sources helps discovery. Cross-checking the same licence ID and address across registry, Agent Card, and website increases confidence. Brand mentions without corroboration are softer.

7. Operational readiness (invoke) signals

Can the agent complete the job?

•Structured booking / inquiry endpoints

•Clear lead times and capacity

•MCP/API availability ([/developers](/developers))

•Human escalation path when automation fails

A high-trust but non-callable merchant loses A2A contests to a slightly less famous but invokable rival.

8. Composite Trust Score

A transparent Trust Score aggregates the above into a digestible metric for directories and agents. Ideal properties:

•Explainable breakdown

•Resistant to single-signal gaming

•Updated as verification and freshness change

•Separate from paid placement

Browse scored, agent-readable profiles in [/directory](/directory). Conceptual overview of product positioning: Trust Layer for AI—not a review spam network.

How Ranking Differs Across Surfaces

| Surface | What “rank” means | Dominant signals |

| --- | --- | --- |

| Perplexity / AI Overviews | Citation order / inclusion | Relevance, authority pages, clarity |

| ChatGPT (browse / tools) | Mention + tool choice | Retrieval fit, tool success, clarity |

| Gemini | Overview + Maps-like local cues | Entity consistency, local facts |

| DeepSeek / other LLMs | Mention accuracy | Training + retrieval mix; clean text helps |

| Agent directories | Shortlist for A2A | Verification, Trust Score, invoke readiness |

| Custom procurement agents | Supplier selection | Compliance fields, SLAs, APIs |

Do not optimize for one screenshot. Optimize for portable signals.

A Practical Scoring Rubric Merchants Can Self-Audit

Score yourself 0–2 on each row (0 missing, 1 partial, 2 strong):

| Signal | 0 | 1 | 2 |

| --- | --- | --- | --- |

| Stable Agent Card ID | None | Temporary page | Canonical `/t` ID |

| NAP consistency | Conflicts | Mostly aligned | Exact match everywhere |

| Licence / registry | Claim only | Linked ID | Claimed + verified |

| Hours structured | Image/bio only | Text on site | Fields + JSON + timezone |

| Offer structure | Vague | Partial menu/services | Priced, constrained list |

| Verification | None | Soft checks | Identity / org / licence |

| Freshness | Unknown | Occasional edits | Visible last-updated |

| Proof reviews | Stars only | Some detail | Evidence-oriented |

| Invoke path | “DM us” | Web form | API/MCP/booking rules |

| Citeable definition | Slogan | Short blurb | Quotable 2–4 sentence block |

Interpretation: below ~8 total → agents will struggle. 8–14 → discoverable but fragile. 15+ → competitive for recommendation *if* category relevance is clear.

Gaming Risks (and Why Agents Will Discount Them)

•Fake review floods

•Keyword-stuffed categories

•Schema that lies

•Copy-pasted licence numbers

•Purchased “AI citation” spam posts

•Trust badges with no verifier

As agent ecosystems mature, cryptographic or platform-mediated verification and registry provenance beat DIY badge PNGs. Soft sell reality: platforms that bind claims to checks outcompete brochureware.

Scenario Walkthrough: How an Agent Might Shortlist

User intent: “Find a verified, licensed dinner spot in Wan Chai for four people at 19:30; prefer places with clear allergen info.”

A competent agent path:

1. Discover candidates from directory search, registry-backed profiles, and retrieved pages

2. Filter by licence status and open hours at 19:30

3. Rank by verification level, completeness of allergen fields, freshness of menu, and Trust Score

4. Recommend two to three options with citations

5. Invoke booking on the first option that exposes party-size rules and a callable action

Who loses?

•Beautiful sites with PDF-only menus (no allergen fields)

•High star averages with no proof policy and conflicting addresses

•Unclaimed registry rows that never gained hours/booking enrichment

•Merchants who are easy to mention in GEO blogs but impossible to book by tool

This is why trust signals are not a branding exercise. They are selection features in agent workflows.

Separating Marketing Trust From Machine Trust

| Marketing trust artifact | Machine trust artifact |

| --- | --- |

| Lifestyle photography | Timestamped field updates |

| Press quotes without identifiers | Licence ID + authority link |

| “As seen in…” logo walls | Corroborating URLs agents can fetch |

| Influencer reels | Proof-based reviews tied to visits |

| Soft “verified” sticker | Platform verification state in JSON |

Keep the marketing. Wire the machine artifacts into your Agent Card so ChatGPT, Perplexity, Gemini, and DeepSeek—and the tools they call—do not have to infer.

Playbook: Strengthening Trust Signals in 30 Days

Days 1–7 — Identity & provenance

•[ ] Publish or clean Agent Card ([/dashboard/create-trusgent](/dashboard/create-trusgent))

•[ ] Claim official registry record if applicable

•[ ] Align legal name, trade name, address

•[ ] Add licence identifiers to profile and site footer

Days 8–15 — Completeness & freshness

•[ ] Structure hours and exceptions

•[ ] Convert offers/menus to fields

•[ ] Set update cadence (weekly for F&B; monthly for B2B)

•[ ] Add last-updated display

Days 16–23 — Verification & proof

•[ ] Complete identity / business verification flows

•[ ] Attach certificate evidence where allowed

•[ ] Define proof-based review policy

•[ ] Respond to existing public complaints

Days 24–30 — Invoke & monitor

•[ ] Enable booking/inquiry actions on `/b`

•[ ] Confirm JSON parity via [/developers](/developers) docs

•[ ] Run AI visibility prompts ([/tools/ai-visibility](/tools/ai-visibility))

•[ ] Fix any hallucinated attributes models still emit

Where Trust Score Fits in GEO

GEO gets you into the candidate set via content and mentions. Trust signals reorder that set when the user asks for “reliable,” “licensed,” “verified,” or when an agent must minimize downside risk (wrong restaurant, unlicensed operator, stale phone).

Think of GEO and trust as orthogonal axes:

•High GEO, low trust: frequent mentions, frequent hedges and errors

•Low GEO, high trust: accurate when found, rarely discovered

•High GEO, high trust: the compounding zone—mentions resolve to verified, fresh, invokable entities

Trusgent’s Trust Layer for AI is built for that compounding zone: Agent Cards, registry claims, verification, Trust Score, and developer-callable surfaces—not star inflation.

Related reading:

•[Beyond GEO: Why Being Mentioned by AI Is Not Enough](/blog/beyond-geo-trust-layer-for-ai)

•[What Is an Agent-Ready Business Profile?](/blog/agent-ready-profile)

•[How to Make Your Business Discoverable by AI Agents](/blog/how-to-make-your-business-discoverable-by-ai-agents)

FAQ

Is Trust Score the same as SEO Domain Authority?

No. Domain Authority-style metrics estimate link equity for pages. Trust Score (in a Trust Layer sense) estimates entity reliability from verification, completeness, freshness, and proof—aimed at agents and directories.

Can I rank in AI answers with weak verification if my content is great?

Sometimes for informational queries. For transactional or safety-sensitive queries (food, medical-adjacent, finance), weak verification is a liability. Agents will prefer corroborated operators when tools allow.

Do proof-based reviews replace star ratings?

They complement and discipline them. Averages remain intuitive for humans; proof and recency help machines avoid fraud.

How often should I refresh profile data?

Whenever operational facts change—and on a minimum cadence even if nothing major changed, so timestamps stay honest. F&B menus may need weekly attention; professional services may be monthly.

Will paying for ads create trust signals?

Ads can buy attention; they rarely substitute identity, licence, or freshness checks. Keep paid acquisition separate from trust architecture.

What is the single highest-leverage trust upgrade for HK F&B?

Claim the correct FEHD (or relevant) registry record, align premises data, structure hours/menu, and verify ownership—then expose JSON. Details in the discoverability playbook.

How do MCP tools affect trust?

Successful, well-specified tools increase invoke readiness. Flaky tools that error under agent calls can *hurt* selection even if marketing trust looks strong. Reliability is part of trust.

---

*Updated September 2026*