The agentic economy, in plain English.
The protocols agents use to reach businesses and pay, and the disciplines that claim to measure or improve how agents treat a brand. What each one is, what it is not, and who provides it.
Protocols
Different protocols govern different handoffs. A specification is not proof of deployment: a working connection still depends on authorization, versions, region and onboarding.
MCP
Agent ↔ tool/dataThe standard way an AI agent plugs into tools and data: a calendar, a CRM, a company's own systems. Created by Anthropic; now stewarded by the Linux Foundation's Agentic AI Foundation.
Key distinction: Connect to tools, services and data; not a peer-agent collaboration protocol.
A2A
Agent ↔ agentA way for one AI agent to find another, hand it a task and get the work back.
Key distinction: Discover other agents, delegate tasks, exchange work.
Poppy — Personal Agent Protocol
Personal agent ↔ businessA draft protocol for a customer's personal agent to identify itself to a business, get the customer's consent and open a session.
Key distinction: Identity, business discovery, sessions, consent, cross-channel access.
UCP — Universal Commerce Protocol
Shopping agent ↔ merchantAn open specification for shopping agents to work with merchants, from browsing a catalog through checkout and order tracking.
Key distinction: Commerce activities including checkout and order lifecycle.
Agentic Commerce Protocol (ACP)
AI commerce ↔ merchantThe Agentic Commerce Protocol from OpenAI and Stripe: how merchants expose product feeds and checkout to AI agents. It powered ChatGPT Instant Checkout until March 2026; it is now used for product feeds and by Microsoft Copilot Checkout.
Key distinction: Merchant product feeds and agent checkout. Not IBM's Agent Communication Protocol (now part of A2A).
AP2 — Agent Payments Protocol
Agent ↔ payment authorityA protocol for proving that a customer authorized an agent to make a specific purchase, using signed mandates. Created by Google; stewarded by the FIDO Alliance since April 2026.
Key distinction: Proof of purchasing intent, signed mandates and payment records.
OAuth / OIDC
User ↔ service permissionDelegated user authentication/authorization used by systems including Poppy.
OpenAPI
Developer ↔ APIMachine-readable descriptions of APIs, not an agent execution layer.
x402
Agent ↔ API / content (payment)An open protocol for agents to pay for APIs and content per request, using the web's HTTP 402 'payment required' code.
Key distinction: Pays per request over HTTP 402. Not a checkout protocol like ACP or UCP.
ACORD data standards (GRLC, P&C)
Insurer ↔ broker ↔ reinsurer (data)Defines insurance messages and data. Agents can use it, but it is not an agent protocol.
Core Data Record (CDR)
London market placement and claims (data)A market data standard, not a personal-agent protocol.
Agent Mediated Insurance (AMI)
Customer agent ↔ insurer (quote, compare, bind)Insurance-specific consent, disclosure and audit rules for agent buying. Not a universally supported standard.
CEN insurance data for MCP (proposed technical report)
Insurance data ↔ MCP toolsGuidance on mapping insurance data to MCP tools. A proposal, not a standard.
Other standards and extensions
OpenAI's checkout specification within the Agentic Commerce Protocol (ACP): how an agent completes a purchase with a merchant on the user's behalf.
A2UI is an open protocol that lets AI agents send interface layouts as declarative JSON, which client apps render natively on web, mobile and desktop.
An open protocol for how an AI agent streams its actions and results into an application's user interface.
RFC 9110 (HTTP Semantics, IETF Internet Standard) defines the architecture, terminology and semantics shared by all HTTP versions.
An open protocol that lets software agents pay for APIs and services over HTTP, using the 402 'payment required' response.
An official MCP extension that lets tools return interactive interfaces, such as forms, charts and dashboards, shown inside the chat; supported by ChatGPT, Claude, Goose and VS Code.
OpenID Foundation work, led by its AI Identity Management Community Group, on how AI agents prove identity and act for users, building on OpenID Connect and OAuth.
Measurement and visibility disciplines
Each works on different phases of a buying decision. One question separates them: does it tell you who wins? See them side by side in the buying journeys.
SEO
Search engine optimizationImproves how a company's pages rank in web search, one of the sources agents draw on during discovery.
Tells you who wins? No Gets you found, strongly on Google's AI Overviews, weakly elsewhere. Doesn't tell you whether you win.
Ahrefs (Aug 2025, 15,000 prompts): Only about 8% of the pages ChatGPT, Gemini and Copilot cite rank in Google's top 10. Source ↗
Show all the evidence (7)
Semrush (Jul 2026, 50,000 brands): SEO metrics predicted which brand ChatGPT names for a topic only about half the time. Source ↗
University of Toronto (Jan 2026, preprint): 4–15% overlap between Google results and the sources AI assistants cite. Source ↗
Peer-reviewed GEO study (KDD 2024): Content changes lifted a fifth-ranked site's AI visibility by 115%; Google rank doesn't lock it in. Source ↗
Counter-finding: Seer Interactive (Jan 2025): Page-one Google rank was the strongest predictor of brand mentions in GPT-4o, with a correlation of about 0.65 (single questions, mentions only). Source ↗
Google AI Overviews: Track rankings more closely: 38–76% of their citations come from the top 10 (Ahrefs, BrightEdge, seoClarity). Source ↗
The final recommendation: No published study links SEO to the brand an assistant finally recommends. AIVO's research is the only work on this so far (internal, Oct 2026, publication planned).
SEO agencies and tools. Not yet cataloged in the Atlas: submissions welcome.
AI visibility (GEO)
First-prompt mentions and citationsTracks and improves how often a brand is mentioned or cited in an assistant's first answer.
Tells you who wins? No Gets you into the conversation. Doesn't decide who wins.
AIVO published research: 87.3% of brands named at the first turn of a buying conversation were displaced by a competitor by the fourth turn (n=1,427 brand probes, 10 industries). Source ↗
Visibility, GEO and SEO tools are outside the Atlas's scope: they measure being seen, not being chosen.
Agent access & content readiness
Feeds, structured data, agent access pointsMakes a company's information easy for agents to retrieve and understand.
Tells you who wins? Partly Makes your facts available. Available is not the same as used.
AIVO published research: 75.7% of the facts models held about a brand were not used when they made a recommendation (n=592 facts, 22 brand-SKUs). This is the Linkage Gap. Source ↗
Providers that meet the published criteria for agent access and content readiness will be listed here.
Evidence for agents
Facts and proof agents can use · e.g. brand.contextA machine-readable file, published on a brand's own website, that gives agents the brand's facts with the evidence behind them. AIVO's remedy for the Linkage Gap.
Tells you who wins? Designed to Designed to carry a brand's facts and evidence through to the decision.
Vendor documented (AIVO): Published as an AIVO specification. Classified like any vendor capability.
Agent journey tracking
Turn-by-turnFollows realistic multi-turn conversations to see where a brand enters the shortlist and where it falls out, and why.
Tells you who wins? Partly Shows where, turn by turn, a brand is dropped.
Vendor documented (AIVO): Measured in controlled multi-turn buying conversations; diagnoses include the Linkage Gap and Evidence Gap.
AI Win Rate · decision research
Who is chosen, and who wins insteadRuns realistic buying conversations and records how each ends: an outright pick, a shared shortlist, a referral elsewhere, or no pick. Win rate is the share won outright; shared listings are reported separately.
Tells you who wins? Yes Measures the final choice directly.
Published methodology (AIVO): AI Win-Rate Methodology v2.0: how outright final recommendations are counted, with shared listings, referrals and no-picks reported separately. Source ↗
Show all the evidence (2)
Vendor documented (AIVO): Controlled research outcomes, not observed sales.
AI traffic attribution
Visits and sales that came from AIAttributes website visits and sales to the AI assistants and agents that referred them, so a business can see what arrived through AI.
Tells you who wins? Partly Shows what AI sent you, after the fact. Not who was chosen instead.
Draft convention (AIVO): ATAC, the AI Traffic Attribution Convention: v0.1 draft for public comment, July 2026. Source ↗
ATAC is an open draft convention published by AIVO. Analytics and attribution vendors are not yet cataloged in the Atlas: submissions welcome.
Providers shown are those included under the published criteria, listed alphabetically. No paid placement. Know a provider with published evidence? Suggest one →
AIVO measures
Terms used in the AIVO Meridian row of the buying journeys. AIVO, Inc. publishes this Atlas; these are its own measures, explained plainly.
Legibility
AIVO MeridianWhether AI agents can read what a business offers and put it on the lists it belongs on, alongside its real competitors. A brand can be well known to an assistant and still not surface: that is a legibility gap, not an awareness gap.
Linkage Gap
AIVO MeridianFacts a business publishes that agents hold but never use when they make the final recommendation. Closed by publishing evidence in a form agents use at the decision, such as brand.context.
Displacement
AIVO MeridianThe turn in a multi-turn conversation at which a brand drops out of the agent's shortlist, who replaces it, and why. Remediation targets that turn.
AIRQ (AI Reputation Quotient)
AIVO MeridianHow AI assistants describe a company across six dimensions (trust, vision, products and services, relevance, ethics, culture) through five lenses (consumer, employee, investor, regulator, impact of AI), against a fixed set of peers.
Revenue at Risk (RaR-AID)
AIVO MeridianAn estimate of the revenue exposed when AI agents recommend someone else: category revenue, multiplied by the share of purchases AI influences, multiplied by the measured rate at which agents displace the brand.
Published method: RaR-AID working paper ↗AIVO explainer: What it's costing you ↗
Poppy design participants · 44
Sierra and Meta, plus 7 launch partners (6 Oct 2026) and 35 design partners (9 Oct 2026). Membership does not establish that a participant's products use Poppy.