#Foreword

Every era of digital commerce has been defined by a shift in who the customer actually is.

The web's first wave optimized for human browsers reading hyperlinked documents. The second wave optimized for human shoppers comparing products in a search results page. The third wave — mobile — optimized for human attention measured in three-second swipes.

The fourth wave is different. The customer is no longer human.

In 2026, the fastest-growing category of buyer on the open internet is an autonomous AI agent. It does not browse. It does not click. It does not read marketing copy. It queries structured data, evaluates capability manifests, negotiates with other agents, and executes transactions on behalf of the human who instructed it. The interface it expects is not a webpage — it is a protocol. The brands that win the next decade are the ones that learn to speak that protocol natively, before the agents make their final shortlist.

This is the Business-to-Agent (B2A) imperative. And the window to claim a defensible position in it is closing on a horizon analysts now estimate at 18 to 24 months.

This paper is the field manual for that transition. It is written for the operator who needs to make a real decision in the next quarter — not for the futurist surveying the landscape from a safe distance.

— perea.ai Research

#Executive Summary

The thesis. AI agents are becoming the dominant intermediary between buyers and sellers in a growing share of digital transactions. Businesses that fail to expose themselves to those agents — through standardized protocols, machine-readable manifests, and agent-callable services — become invisible to the buyer's most influential decision-maker. Businesses that move early capture asymmetric distribution, lock in agent-friendly pricing models, and accumulate compounding data advantages that late movers cannot replicate.

The evidence.

• The AI agents market reached $7.84B in 2025 and is projected to grow to $52.62B by 2030, a CAGR of 46.3%, with longer-range projections exceeding $139B by 2034 [^2][^3].
• 78% of enterprise AI teams reported at least one MCP-backed agent in production by Q1 2026, up from 31% twelve months earlier (DigitalApplied).
• 97 million monthly SDK downloads of Anthropic's Model Context Protocol (MCP) in March 2026, up from approximately 2 million at launch in November 2024 — a 4,750% increase in 16 months (DigitalApplied).
• Gartner predicts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5% in 2025 [^3].
• Forrester predicts 30% of enterprise app vendors will launch their own MCP servers in 2026 [^1].
• Kantar describes B2A as "an imminent distribution channel that could rewrite category shares in 18-24 months" [^7].
• Stripe and OpenAI co-developed the Agentic Commerce Protocol (ACP), which now powers ChatGPT Instant Checkout — live with Etsy and rolling out to Shopify's million-plus merchants [^39][^40].
• Google, Shopify, Walmart, Target, Etsy, and Wayfair co-developed the Universal Commerce Protocol (UCP), live January 2026, endorsed by Visa, Mastercard, PayPal, Stripe, American Express (B2BEA / AAXIS).
• Visa, Mastercard, Stripe, and Google each launched competing agentic-payment protocols in the past four months — DBS Bank in Singapore ran the first authenticated agent-launched food purchases in Asia Pacific; Mastercard processed its first agentic transaction on-network in Q4 2025 [^47].
• Vinci Rufus and McKinsey project B2A could represent 20-30% of digital service interactions by 2027-2028 [^8] (McKinsey Digital Strategy Report).

The strategic implications. The agent economy is not a feature of the existing internet. It is a parallel distribution channel with its own discovery primitives, pricing models, and trust signals. Five concrete shifts every operator must absorb:

1. Discovery moves from search to capability negotiation. SEO targets human users scanning blue links. B2A targets agents querying structured manifests at well-known URLs. The optimization surface changes from keywords and backlinks to schema completeness, capability declarations, and protocol compliance.

2. Pricing moves from per-seat to per-action and outcome. Agent buyers are infinitely patient and perfectly rational. They do not pay for unused features or unused seats. Per-action and outcome-based pricing is not a marketing experiment in the agent economy — it is the only model that survives.

3. Trust moves from human reputation to cryptographic verification. Visa's Trusted Agent Protocol, Google's AP2 mandates, and the emerging x402 standard all replace soft trust signals (reviews, brand recognition) with hard ones (signed mandates, verified agent identity, on-chain audit trails).

4. Switching costs collapse. A human takes weeks to evaluate alternatives. An agent takes milliseconds. Customer retention in the agent economy is earned every transaction, not at annual contract renewal. The half-life of competitive advantage shrinks dramatically.

5. The compounding asset is the dataset, not the brand. Every transaction with an agent generates structured intent data: why agent A chose SKU X over SKU Y. Brands that capture this data feed it back into product, pricing, and merchandising decisions at a velocity that brands relying on traditional panel research cannot match.

The recommended action. Treat B2A as a distinct product surface. Audit your readiness across four layers: Data, Discovery, Execution, and Trust. Ship a minimum viable B2A stack (/llms.txt, /.well-known/agent-card.json, an MCP server, schema.org JSON-LD, ACP-compatible checkout) within 90 days. Productize a second offer optimized for agent buyers. Track agent traffic and agent-originated revenue as a first-class business metric. Adjust pricing to reflect the new buyer's behavior. Compound a vertical knowledge graph with every agent-mediated transaction.

The operators who do this in 2026 will own discovery, pricing, and customer relationships in their categories for the next decade. The operators who delay will spend the rest of their careers explaining to investors why their brand became invisible.

#Part I: The Market Window

#The agent economy is no longer a forecast

Programmatic advertising has been agent-to-agent commerce for two decades — bidding agents negotiate with publisher agents in real-time auctions millions of times per second. Algorithmic trading has been agent-to-agent for longer. Cloud resource procurement, energy markets, and supply chain optimization have all been quietly agent-mediated for years [^5].

What changed in late 2024 and accelerated through 2026 is that the underlying intelligence behind these agents became general-purpose. Large language models with reasoning capability, paired with standardized tool-use protocols, made it economically viable to deploy agents into a much broader range of decisions: which restaurant to book, which contractor to hire, which SaaS to subscribe to, which insurance policy to renew.

The market data tracks this transition cleanly:

These numbers describe a transition from experimental to production. The 40% failure rate that Gartner attributes to early agentic AI deployments is real — but it is a deployment-quality issue, not a demand issue. The companies that scale through it capture the disproportionate share.

#What is actually selling

Three concrete buyer profiles drive the demand-side acceleration:

1. The procurement agent. Mid-market and enterprise buyers — particularly in SaaS, supplies, and travel — increasingly delegate vendor evaluation and purchasing to agents. Alibaba's Accio Work has 10M+ monthly users running enterprise B2B agent fleets. Procurement agents shortlist suppliers, compare specifications, place orders, and reorder consumables. They do not read marketing landing pages. They query structured product data and machine-readable supplier capability manifests [^18][^19].

2. The consumer shopping agent. ChatGPT, Gemini, Perplexity, and Microsoft Copilot all now carry shopping surfaces where agents discover, compare, and purchase products on behalf of users. ChatGPT's Instant Checkout, powered by Stripe's ACP, lets US users buy from Etsy without leaving the chat. Shopify's Agentic Storefronts make over a million Shopify Plus merchants discoverable to AI shopping agents. Google's Universal Commerce Protocol — co-developed with Walmart, Target, Wayfair, Etsy, and others — enables AI agents to assemble multi-retailer carts in a single transaction [^9][^75] (UCP Hub).

3. The household management agent. Apple Siri, Amazon Alexa+, Google Gemini, and growing third-party assistants increasingly handle scheduled-task purchasing: zero-click replenishment of consumables, contractor scheduling, restaurant booking, travel modifications. Strategic Inference's central question summarizes the test: "Can Siri buy from you?" If the answer is no, that agent silently routes the transaction to a competitor [^11].

#The protocol stack consolidates

Behind the demand wave, an infrastructure stack consolidated faster than most observers expected. The protocols that matter, ranked by adoption velocity and merchant readiness implications:

#Model Context Protocol (MCP) — the operating system

Anthropic open-sourced MCP on November 25, 2024. Within 17 months it became the closest thing the agent ecosystem has to a universal standard. Every frontier lab — Anthropic, OpenAI, Google, Microsoft, Amazon — supports MCP as a client. Block, Stripe, Cloudflare, GitHub, Microsoft, and thousands of community contributors maintain production MCP servers. The public registry grew from 1,200 servers at end of Q1 2025 to 9,400+ in mid-April 2026, a 7.8× year-over-year expansion [^24][^21][^25] (DigitalApplied).

What MCP does for B2A: it gives AI agents a standard way to discover and call your business's capabilities. Instead of writing custom integration code for every AI provider that wants to interact with your business, you build one MCP server that exposes your tools (book_appointment, check_inventory, submit_quote_request, purchase_product) and every MCP-compatible client — Claude, ChatGPT, Gemini, Cursor, custom enterprise agents — can invoke them.

#Agent2Agent (A2A) — the collaboration layer

Google announced A2A in April 2025 with 50+ launch partners including Atlassian, Box, Cohere, Intuit, Langchain, MongoDB, PayPal, Salesforce, SAP, ServiceNow, Workday, plus the major consultancies — Accenture, BCG, Capgemini, Cognizant, Deloitte, HCL, Infosys, KPMG, McKinsey, PwC, TCS, Wipro. The protocol is now housed by the Linux Foundation. By February 2026 over 150 organizations were behind it [^36][^35] (Google for Developers).

What A2A does for B2A: where MCP connects an agent to your tools, A2A enables your agent to negotiate with another agent. A buyer's procurement agent finds your business's seller agent through a published Agent Card at /.well-known/agent-card.json, evaluates capabilities and pricing, and delegates a complete task. The interaction model is task-oriented: "request quote → return artifact → escalate if needed → close loop." Critically, A2A preserves opacity — agents collaborate without exposing internal logic, memory, or implementation.

#Agentic Commerce Protocol (ACP) — checkout

Stripe and OpenAI co-developed ACP and committed to it as an open standard. It now powers ChatGPT Instant Checkout. Etsy was the first marketplace partner; Shopify's million-plus merchants are rolling out. Meta and Stripe extended ACP for messaging-platform commerce. The specification supports both REST and MCP integration patterns [^39][^40] (ACP GitHub, agenticcommerce.dev).

What ACP does for B2A: it standardizes the cart-and-checkout flow so any agent can complete a purchase against any compatible merchant without custom integration. The merchant remains the system of record. Payments stay on existing rails. The merchant's PSP [^39][^40] (in the reference implementation) handles fraud detection. The agent simply orchestrates the commerce transaction — including capability negotiation, payment delegation via Shared Payment Tokens, and order webhooks.

#Universal Commerce Protocol (UCP) — multi-retailer commerce

UCP went live in January 2026, co-developed by Google, Shopify, Walmart, Target, Etsy, and Wayfair. It is endorsed by Visa, Mastercard, PayPal, Stripe, and American Express, and supported by Best Buy, Kroger, Lowe's, Macy's, Home Depot, and Sephora. UCP is not a pilot — it is production infrastructure that lets AI agents purchase from multiple retailers in a single transaction. Google Merchant Center introduced 60+ new product attributes specifically for UCP [^9][^18][^19][^75].

What UCP does for B2A: where ACP solves single-merchant checkout, UCP solves multi-merchant carts. An agent assembling a household replenishment order across five retailers, or a procurement agent sourcing components from three suppliers, uses UCP to negotiate the entire transaction in a single coordinated flow.

#Agent Payments Protocol (AP2) and Trusted Agent Protocol (TAP) — trust and identity

Google announced AP2 with PayPal in late 2025. AP2 introduces cryptographically-signed mandates — verifiable user intent that reduces fraud and clarifies accountability when an agent acts on a buyer's behalf. Visa launched TAP with Cloudflare to give merchants a way to recognize legitimate AI agents and not falsely block agent purchases with bot defenses. Mastercard launched Agent Pay using Agentic Tokens, dynamic credentials with embedded spending rules. Mastercard rolled this out to all US cardholders by November 2025 with Citi and US Bank as first issuers [^42][^43][^47].

What AP2 and TAP do for B2A: they let merchants distinguish authorized agent transactions from fraud, and let agents prove cryptographically that they have user permission to spend. Without these layers, the chargeback risk and fraud surface of agent commerce remain a hard ceiling on adoption.

#Machine Payments Protocol (MPP) and x402 — micropayments

Stripe's Tempo Layer 1 went live March 18, 2026 with Visa, Mastercard, Deutsche Bank, Shopify, OpenAI, and Anthropic as design partners. MPP introduces "sessions" — an agent authorizes a spending limit upfront and streams micropayments continuously. Visa extended MPP for cards, Lightspark for Bitcoin Lightning, Tempo for stablecoins. Coinbase's x402 protocol moved to the Linux Foundation in September 2025 with backing from Google, Stripe, Visa, Cloudflare, AWS, Anthropic, and NEAR. Circle reported AI agents completed 140 million payments worth $43 million in nine months, almost all on USDC. Circle launched Nanopayments on testnet in March 2026, enabling gas-free USDC transfers as small as $0.000001 (Justin Yek's Agentic Commerce Rails).

What MPP and x402 do for B2A: they unlock pricing models that traditional card rails cannot economically support — true per-call pricing on API services, pay-per-page on premium content, sub-cent transactions for atomic services. Most B2A businesses will not need these on day one. By day 365, businesses operating in API-as-a-product, content, or AI tooling categories will not have a choice.

#Manifest standards — the discovery layer

A constellation of manifest standards now competes for the role of "the agent-readable file you publish at a well-known URL":

• llms.txt — proposed by Jeremy Howard at Answer.AI in September 2024. Markdown file at /llms.txt giving LLMs a curated, structured index of a site. Adopted by Anthropic, Cloudflare, Stripe, Mintlify, Vercel, Perplexity, GitHub Docs, Cursor [^54][^52][^55].
• agent.json / ai-agent.json — JSON manifests at /.well-known/agent.json declaring agent identity, capabilities, endpoints, authentication, and trust signals. Multiple variants competing: aiia.ro, agentinternetruntime.com, ai-manifest.org [^49][^53][^50].
• agent-card.json — A2A's standard agent card at /.well-known/agent-card.json with capabilities, supported modalities, authentication, and pricing (A2A Specification).
• LLM-LD — a more formal "core specification" for llm-index.json published by CAPXEL in February 2026, attempting to consolidate the fragmented manifest ecosystem [^51].

The competing standards are a transient state; they will consolidate. The pragmatic 2026 approach is to publish all of them. They are cheap to maintain and they cover whatever standard the agents on the other end happen to honor.

#Part II: What B2A Actually Is

#The definition

Business-to-Agent (B2A) is a commercial model in which a business's products, services, and processes are designed first for consumption by autonomous AI agents acting on behalf of human users — and only secondarily for direct human consumption. In B2A, agents are the primary customers; humans are the principals on whose behalf those agents transact.

The label is not new. Vinci Rufus framed it in early 2025; Kantar, Welcomespaces, AAXIS, Caversham Digital, and Strategic Inference have all converged on it through 2025-2026. What is new is that the underlying infrastructure is now real enough that B2A is a concrete operating decision — not a thought-leadership topic.

The defining test, attributed to Strategic Inference: "Can Siri buy from you?" If completing a transaction with your business requires a human to click a button on a webpage designed for human eyeballs, you are excluded from this distribution channel. If an agent — Gemini, ChatGPT, Alexa, a private enterprise procurement bot — can discover, evaluate, and purchase from you without a human intermediary, you are inside the channel.

#What separates B2A from B2C and B2B

B2C and B2B are channels defined by who the buyer is. B2A is a channel defined by how the buyer interacts. The distinction matters because the same end customer — say, a small business owner — might engage with you through B2C (a marketing landing page they read themselves), B2B (an account executive on a video call), and B2A (their AI assistant booking a discovery call on their behalf) within a single week. These are not three different customers. They are one customer using three different interaction surfaces. Each surface has different optimization rules.

The most consequential entry in that table is the bottom row: B2A's failure mode is invisibility. A B2C campaign that fails returns measurable bad UX — bounce rates, low conversion. A B2B sales process that fails returns a "no" the rep can learn from. A B2A integration that fails returns nothing. The agent silently selects a competitor and the principal never knows the alternative existed. There is no signal to optimize against unless the business is already inside the channel measuring its own agent traffic.

#The B2A stack — six layers

A complete B2A surface comprises six layers. Each layer answers a question agents ask before they can transact with you.

#Layer 1: Discovery — "Does this business exist?"

The agent's first move is to find you. The minimum viable discovery layer:

• robots.txt — does not block agents from the manifest paths. Many businesses inadvertently block AI crawlers and lose visibility instantly.
• llms.txt at the site root — markdown index of the most useful pages for an LLM agent to read.
• llms-full.txt — concatenated full content of the indexed pages, for agents that prefer single-fetch context loading. Agents visit llms-full.txt at roughly 2× the rate of llms.txt [^52][^55] (Profound data).
• /.well-known/agent.json and / or /.well-known/ai-manifest.json — structured capability manifests.
• /.well-known/agent-card.json — A2A Agent Card with skills, supported modalities, auth requirements.
• /.well-known/ai-agent.json — alternative manifest format with explicit trust signals (verified, endorsements, uptime).

Plus, on every page intended for agent consumption:

• Schema.org JSON-LD markup for the appropriate entity types: Organization, Service, Product, Offer, FAQPage, HowTo, BreadcrumbList, LocalBusiness. Agents parse structured data first, prose second.
• Canonical URLs and stable identifiers — agents that cache product data depend on identifier stability for de-duplication.

#Layer 2: Capability — "What can this business do?"

Once an agent has discovered you, it asks: which of my user's intents can you satisfy?

The capability layer has two surface forms:

• MCP server — exposes your business's tools to any MCP-compatible agent. Typical tools for a service business: book_consultation, request_quote, check_availability, cancel_booking, get_pricing. Typical tools for an e-commerce business: search_products, get_product_detail, add_to_cart, start_checkout, track_order. The MCP server can be a thin translation layer in front of your existing REST API — most production deployments are exactly this [^86][^87][^91][^92].
• A2A Agent Card — declares the same capabilities as part of an agent-to-agent collaboration model. Where MCP serves clients (LLM hosts), A2A serves peer agents who delegate sub-tasks.

The capability layer is where most B2A initiatives stall. It is tempting to treat MCP as a feature ("we'll add an MCP server this quarter"). The right framing is closer to "we will expose 100% of our customer-facing capabilities through MCP within the next 12 months, and we will treat that as a first-class engineering concern equal to our human-facing UI."

#Layer 3: Execution — "Can this business actually fulfill?"

Discovery and capability declarations are claims. Execution is delivery. The execution layer answers: if I delegate this task, will it complete successfully?

Concrete execution-layer obligations:

• Real-time inventory and availability data — feeds that update at minimum every 4 hours, ideally every 15 minutes for high-velocity inventory. Agents that recommend out-of-stock products lose user trust on first failure. Brands that recommend out-of-stock products to agents lose agent trust permanently.
• Webhook-driven state updates — order placed, inventory changed, appointment confirmed, status changed. Polling is a tax agents do not pay; if you cannot push state changes, you fall out of the agent's recommendation set.
• Idempotency and safe retry semantics — a critical detail. Agents retry on failure. APIs that double-charge or double-book on retry are worse than APIs that are slow.
• Rate-limit budgeting — agents make many small queries. APIs with overly aggressive rate limits get classified by agents as unreliable and downranked.
• Human-in-the-loop checkpoints where they matter — high-value transactions, account changes, irreversible actions. Built-in approval flows are a positive trust signal in B2A, not friction.

#Layer 4: Commerce — "Can I complete a transaction?"

If your business charges money (or accepts money), the commerce layer is the bridge between capability and revenue.

The 2026 baseline for commerce-layer compliance:

• ACP-compatible checkout — implementable as a thin wrapper on top of an existing Stripe integration. The Agentic Checkout Specification (ACS) is a stable, versioned REST API (API-Version: 2026-01-16) defining session creation, update, completion, and cancellation. Merchant remains system of record. Payments stay on existing rails. [^39][^40] (ACP RFC)
• UCP-compatible product feed — for retailers, this means Google Merchant Center with the 60+ new UCP attributes populated. Brands selling through Shopify Plus get UCP coverage automatically; brands on Adobe Commerce, BigCommerce, or custom platforms must implement explicitly. [^79][^78] (Creatuity.)
• AP2 mandate verification — for high-trust transactions, accept and verify cryptographically-signed user intent mandates. This is not table stakes for SMB merchants in 2026; it will be by 2027.
• TAP recognition — at the WAF / CDN layer, distinguish trusted agent traffic from bots. Don't false-positive legitimate agent purchases. [^42]

#Layer 5: Trust and identity — "Can the buyer prove they have permission to act?"

The trust layer is what separates a legitimate agent transaction from fraud, and what makes the merchant comfortable accepting agent-originated traffic at scale.

• Mandates — AP2's signed-intent model. The merchant verifies that the agent is acting under a legitimate, scoped, time-bounded user authorization.
• Agent identity verification — TAP's cryptographic identity layer. The agent presents a verifiable identity proof; the merchant validates it through Visa's network or another trust anchor.
• Audit trails — every agent-mediated transaction emits an immutable audit log: which agent, on whose behalf, under what scope, completing what action, paid through what method. Regulated industries will mandate this; unregulated industries will adopt it because the operational benefit (dispute resolution, fraud detection) is overwhelming.
• Reputation primitives — agent-side and merchant-side reputation scores. Early entries: B2Alpha's "reputation staking" model, the OATR trust registry, A2A's authenticated extended agent cards. Reputation in the agent economy is structured data, not Yelp stars.

#Layer 6: Pricing and payments — "How does value transfer?"

The payment rails matter less than the pricing model. Three pricing dimensions every B2A operator must decide:

• Granularity — do you charge per session, per call, per outcome, per minute, per token, per event, per resource consumed? Card rails cannot economically support sub-cent granularity; stablecoin rails (x402, USDC Nanopayments, Tempo) can.
• Authorization model — does the agent authorize once at session start (subscription-style), authorize per transaction (pay-as-you-go), or stream micropayments under a session mandate (MPP)? Each requires different infrastructure.
• Settlement currency — fiat through cards (familiar, chargeback-protected, expensive at small denominations), bank rails (cheap, irreversible, slow), stablecoins (cheap, irreversible, fast, regulatorily uncertain). Most B2A businesses will run all three for several years.

The pricing implication that operators most often miss: agent buyers are infinitely patient and perfectly rational. They will not pay for unused features. They will not pay for unused seats. They will switch at the next transaction if a competitor offers identical capability at lower per-action cost. Subscription pricing as the default model for most categories is dead in B2A. Per-action and outcome pricing are not innovations; they are obligations.

#Part III: The B2A Readiness Framework

A B2A readiness audit must answer one question concisely: can autonomous agents discover, evaluate, transact with, and trust this business — and if not, where exactly does the system break?

The framework presented here is a four-layer scored audit synthesized from the B2X Agentic Readiness Framework, Strategic Inference's approach, perea.ai's engagement work, and the structural recommendations in the AAXIS / B2BEA digital transformation framework. Each layer scores 0-100; the composite Agent Readiness Score (ARS) is the average.

#Layer 1: Data (weight: 30%)

The foundational layer. Agents cannot evaluate what is not structured.

Scored sub-dimensions:

• Catalog completeness (0-100): proportion of products / services with complete, machine-readable attribute coverage. Target: 80%+ on optional attributes, not just required ones. Brands at 7-attribute coverage lose to brands at 30-attribute coverage on identical goods [^74].
• Stable identifiers (0-100): GTIN/UPC/SKU/EIN/EAN coverage, stable URLs, canonical references. Agents cross-reference your data against external knowledge graphs (manufacturer specs, review aggregators, comparison databases). Without stable identifiers, the agent evaluates you in isolation — a structural disadvantage in comparison rankings.
• Schema markup quality (0-100): coverage of Product, Offer, Service, Organization, FAQPage, LocalBusiness, BreadcrumbList, Review. Validation errors reduce trust. Agents downgrade ambiguous markup.
• Real-time accuracy (0-100): how stale is your inventory, pricing, availability data? Feeds updated daily are insufficient for agents operating in real time.
• Multi-locale coverage (0-100, weighted by relevance to your market): structured data in the languages and formats relevant to your buyer's agent. Agents serve global queries; LATAM e-commerce brands missing Spanish-language structured data lose to brands that have it.

Diagnostic questions:

• Can an agent read your product catalog without parsing HTML?
• Are your prices, inventory, and availability machine-readable in real time?
• Can an agent uniquely identify each of your products with a stable global identifier?
• Are your service offerings declared in machine-readable form, including pricing tiers, eligibility, and terms?

#Layer 2: Discovery (weight: 25%)

The signaling layer. Even the best data is worthless if agents cannot find it.

Scored sub-dimensions:

• Manifest coverage (0-100): presence and validity of llms.txt, llms-full.txt, /.well-known/agent.json, /.well-known/ai-manifest.json, /.well-known/agent-card.json. Bonus weight for LLM-LD and ai-agent.json.
• GEO / AEO content authority (0-100): citation frequency in AI-generated answers. Measured through tools that monitor brand mentions across ChatGPT, Perplexity, Gemini, Claude. The GEO discipline is mature enough by 2026 that competitive benchmarking is concrete [^58][^62][^63] (Wikipedia: Generative Engine Optimization).
• Agent crawl access (0-100): explicit allowance of major AI crawlers in robots.txt, no false-positive bot blocking, no aggressive Cloudflare/CAPTCHA challenges that block legitimate agent traffic.
• Registry presence (0-100): listing in agent registries (aiia.ro, public MCP server registries, A2A registries, vertical agent marketplaces). Most businesses score zero here in 2026 — early movers capture disproportionate visibility.
• Search-result-page surfacing (0-100): traditional SEO foundation. Research finds 99% of AI Overview citations come from the organic top 10 (Incremys via h-haboubi). Strong SEO is still the floor.

Diagnostic questions:

• If an agent fetches https://yourdomain.com/llms.txt, what does it find?
• Does https://yourdomain.com/.well-known/agent-card.json return a valid A2A Agent Card?
• Are you cited in AI-generated answers when users ask category-defining questions?
• Are you indexed in any of the registry surfaces serving agent traffic?

#Layer 3: Execution (weight: 25%)

The fulfillment layer. Discovery and capability declarations are claims; execution is delivery.

Scored sub-dimensions:

• API surface completeness (0-100): proportion of customer-facing capabilities exposed through APIs an agent can call. A business that lets humans book appointments through a webpage but not through an API has a B2A gap.
• MCP server presence and quality (0-100): existence, scope, schema clarity, error-handling quality, authentication model. A read-only MCP server is a strong starting point; a write-capable server with proper OAuth 2.1 / PKCE flows is the production target [^92][^87][^91][^90].
• Webhook reliability (0-100): uptime, latency, retry semantics, signature verification. Agents abandon merchants who deliver inconsistent state updates.
• Idempotency and safe-retry posture (0-100): cryptographic idempotency keys honored, retries surfaceable in audit log, no double-charge / double-book risk under normal retry.
• Performance (0-100): P95 response time, error rate, geographic distribution. Agents downrank slow APIs; on commerce flows, slow APIs lose transactions silently.

Diagnostic questions:

• Can an agent complete every customer-facing workflow you offer humans, end to end, through an API?
• If an agent retries a failed booking, does the system safely no-op or does it double-book?
• What is your P95 API response time, and is it under 500ms for read operations and 2s for write operations?
• Do you publish a status page agents can monitor?

#Layer 4: Trust (weight: 20%)

The verifiability layer. Without trust signals, even a perfect technical stack is rate-limited by skepticism.

Scored sub-dimensions:

• Identity verification (0-100): TAP / agent identity protocol support, agent-traffic recognition at the CDN/WAF layer, agent allowlisting where appropriate.
• Mandate handling (0-100): AP2 / signed-mandate support for high-value transactions. Even if not yet table stakes, presence is a strong differentiator in 2026.
• Audit and observability (0-100): per-transaction logging, agent-attributable audit trails, exportable for compliance / dispute resolution.
• Reputation primitives (0-100): listing in agent reputation services, structured testimonials, on-chain or signed proof of business legitimacy where applicable.
• Compliance posture (0-100): SOC 2, GDPR, CCPA, HIPAA where relevant; data residency declarations in manifests.

Diagnostic questions:

• If a buyer's agent presents a TAP-verified identity, do you recognize and honor it?
• For transactions over a meaningful threshold ($X), do you require a signed AP2 mandate?
• Can you produce a complete, agent-attributable audit log for any transaction in the past 90 days?
• Are your compliance certifications declared in your manifest in a machine-readable form?

#Composite scoring and triage

The realistic distribution of mid-market businesses in 2026 is heavily weighted toward 0-30 and 31-50. The leader band (71+) is occupied almost exclusively by Shopify Plus merchants who inherited UCP and ACP coverage automatically, plus a handful of intentional early adopters in B2B SaaS. The window to enter the leader band before competitive saturation closes is the 18-24 month period Kantar identifies — measured from approximately mid-2025.

#Part IV: The 90-Day B2A Implementation Playbook

This section translates the readiness framework into a concrete operating plan. Targets and timelines are calibrated for a mid-market business with an existing engineering team or a willing implementation partner.

#Days 0-30: Foundation

Goal: ship a minimum viable B2A surface and start measuring agent traffic.

#Week 1: Discovery diagnostics and priority pass

• Run an inbound traffic audit. Filter logs by User-Agent against a known list of AI crawlers and agent traffic signatures (GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, etc.). Establish a baseline: how much of your current traffic is already agent-originated?
• Inventory existing structured data: schema.org markup, Open Graph, RSS, sitemap, robots.txt. Document what is correct, incomplete, or wrong.
• Inventory existing APIs: which customer-facing capabilities have programmatic surfaces, and which require a human in a browser?
• Run the four-layer readiness scorecard. Capture the baseline ARS.

#Week 2: Manifests and discovery layer

• Publish /llms.txt and /llms-full.txt. Curate, do not concatenate every page on the site. Lead with the 5-15 highest-value entry points for an agent. Anthropic, Cloudflare, Stripe, and Mintlify have public examples worth studying.
• Publish /.well-known/agent.json declaring your service capabilities, endpoints, authentication, and rate limits.
• Publish /.well-known/agent-card.json with A2A-compliant Agent Card schema: skills, modalities, auth requirements, contact, version.
• Publish /.well-known/ai-manifest.json per the ai-manifest.org community draft.
• Validate robots.txt does not block AI crawlers. Specifically allow access to manifest paths.

#Week 3: Schema and structured data

• Implement Schema.org JSON-LD for Organization, Service (one block per service offering), FAQPage, and LocalBusiness where applicable.
• For e-commerce: complete Product and Offer markup including GTIN, brand, pricing, availability, condition, and shipping details.
• For service businesses: complete Service markup with serviceType, areaServed, availableChannel, offers.
• For content publishers: complete Article, Author, Organization, and Citation markup.

#Week 4: First MCP server (read-only)

• Stand up a thin MCP server in front of existing read-only APIs. Recommended starting tools (adapt to category):
  • For a service business: get_services, get_pricing, get_availability, get_team, get_case_studies.
  • For an e-commerce business: search_products, get_product, get_inventory, get_pricing_for_segment, get_shipping_options.
  • For a SaaS: get_features, get_pricing_tiers, get_integrations, compare_to_competitors.
• Use one of the production-ready MCP frameworks [^85][^92] (official Anthropic SDK) rather than implementing from scratch.
• Publish the MCP server URL in your agent.json and agent-card.json manifests.
• Test against the major MCP clients: Claude Desktop, Claude Code, Cursor, ChatGPT (where supported), Gemini.

#Day 30 deliverables checklist

• Baseline ARS captured.
• All discovery manifests live and valid.
• Schema.org JSON-LD on all priority pages.
• Read-only MCP server in production with at least 5 tools.
• Agent traffic monitoring instrumented and dashboarded.
• Public-facing announcement / manifesto on your repositioning.

#Days 30-60: Activation

Goal: convert agent discovery into agent-mediated transactions.

#Week 5: Capability expansion

• Extend MCP server with write-capable tools, gated by OAuth 2.1 with PKCE. Recommended additions:
  • book_consultation / start_purchase_flow / submit_quote_request
  • State-changing tools requiring authentication
• Implement per-tool scopes. Read-only agents cannot even discover destructive operations.
• Implement audit logging at the MCP server layer: every tool call, with agent identity, timestamp, scope, parameters, result.

#Week 6: Commerce surface

• For e-commerce: implement ACP-compatible checkout endpoints on top of existing Stripe integration. The Agentic Checkout Specification provides a stable REST contract.
• For service businesses: implement ACP-style checkout for the productized service offering (consultation booking, audit purchase, fixed-price engagement).
• For SaaS: implement programmatic trial and subscription provisioning with mandate-style spending caps.
• Publish ACP-compliant webhook endpoints for order state changes.

#Week 7: Trust layer foundation

• At the WAF / CDN layer, segment agent traffic. Recognize and honor TAP-style identity signals where present.
• Implement audit-trail export so any agent-mediated transaction can be reconstructed for compliance or dispute resolution.
• Add a public "agent ToS" page declaring what agents are allowed to do, what they are not allowed to do, and how the business handles disputes.
• Add structured testimonials and verifiable trust signals (SOC 2, GDPR, etc.) to manifests.

#Week 8: Vertical-specific feed optimization

• For e-commerce: complete optional UCP attributes in Google Merchant Center (60+ new attributes added January 2026).
• For service businesses: ensure local business listings (Google Business Profile, Apple Business Connect, Yelp) are complete and match structured data on the site.
• For B2B: enrich product / service catalog with attributes used in procurement queries (compatibility, certifications, technical specs).

#Day 60 deliverables checklist

• Write-capable MCP server with proper auth.
• ACP-compatible checkout flow live for at least one product / service.
• Agent traffic segmented at the edge.
• Audit trail exportable end-to-end.
• Vertical-specific feed optimizations complete.

#Days 60-90: Optimization

Goal: instrument, measure, iterate, and prove ROI.

#Week 9-10: Instrumentation

• Build a dashboard tracking:
  • Agent visits / week
  • Agent-mediated transactions / week
  • Agent-mediated revenue / week
  • Top tools called
  • Top agents (by User-Agent) interacting with the business
  • Agent-conversion funnel: discovery → capability query → transaction → completion → repeat
• Compare agent-mediated metrics to human-mediated metrics. The first time agent-mediated AOV exceeds human-mediated AOV (a common pattern in early B2A deployments) is a strategic inflection: it is evidence that the agent channel deserves dedicated investment.

#Week 11: Conversion optimization

• Identify the top three drop-off points in the agent funnel.
• Common patterns and fixes:
  • Drop-off at capability discovery. Cause: ambiguous tool descriptions in MCP schema. Fix: rewrite tool descriptions in natural language an LLM can match against user intent.
  • Drop-off at transaction. Cause: missing required attributes in the cart/checkout response. Fix: populate optional attributes; check ACP spec compliance.
  • Drop-off at confirmation. Cause: webhook unreliability. Fix: implement retry-with-backoff and idempotent webhook delivery.

#Week 12: Pricing and packaging

• Introduce a productized offering optimized for agent buyers: small unit, clear scope, fixed price, machine-readable terms, instant availability.
• Test outcome-based pricing on a single product or service line. Track agent-conversion uplift; outcome pricing typically lifts agent conversion by a margin large enough to be visible within 60 days.

#Day 90 deliverables checklist

• Agent metrics dashboard live and reviewed weekly by leadership.
• Top three funnel leaks identified and fixed.
• At least one productized agent-optimized offering live.
• Updated ARS captured. Improvement of 20+ points over the day-0 baseline is realistic.
• Public case study or anonymized data point published — feeds the publication flywheel.

#Part V: Vertical Plays

The general framework above adapts to vertical specifics. Four high-priority verticals with concrete differences in implementation:

#E-commerce (DTC, mid-market $1M-$50M revenue)

• Highest-leverage protocol: UCP (Google + Shopify). Live since January 2026. Shopify Plus stores get partial coverage automatically. Adobe Commerce, BigCommerce, and custom platforms must implement explicitly.
• Critical data work: complete the 60+ optional UCP attributes in Google Merchant Center. GTIN coverage to 100% of catalog. Multi-channel feed quality (Meta Commerce Manager, Microsoft Merchant Center, ChatGPT shopping inclusion via Stripe ACP, Perplexity Merchant Program).
• Critical execution work: real-time inventory APIs. Shopify's Catalog MCP is the floor; production deployments need their own MCP server exposing checkout, inventory, and fulfillment.
• Key risk: "Shopify enabled MCP for you" is not the same as "AI agents recommend you." If product data is incomplete, MCP availability does not help. (B2X Software's positioning explicitly addresses this gap.)

#Service businesses (legal, real estate, healthcare, home services)

• Highest-leverage surface: voice-callable agents. Agents calling agents through telephony (using systems like Retell, Vapi, Bland) is a 2026 capability that almost no service business has yet wired up. Service businesses with voice-agent surfaces and MCP-callable booking are early in a defensible position.
• Critical data work: structured service catalog (Service schema with serviceType, areaServed, availableChannel, offers), provider profiles, real-time availability.
• Critical execution work: programmatic appointment booking with mandate-aware approval flow. Cancellation, rescheduling, and modification APIs.
• Key risk: regulatory and compliance constraints in healthcare, legal, and finance. Agent-mediated transactions in regulated verticals require especially rigorous trust-layer implementation. The benefit: competitors deterred by compliance complexity leave the early-mover window wider.

#B2B SaaS

• Highest-leverage surface: MCP server exposing the product itself. SaaS companies with native MCP servers position themselves as the "preferred" agent integration in their category. Forrester predicts 30% of enterprise app vendors ship MCP servers in 2026; the early 30% become the default in their respective markets.
• Critical data work: comprehensive feature catalog, pricing tier machine-readability, integration manifest.
• Critical execution work: programmatic trial provisioning, mandate-aware subscription management, agent-friendly onboarding (no human-only setup steps).
• Key risk: treating MCP as a feature instead of a distribution channel. The competitive advantage is not "we have an MCP server" — it is "our MCP server is the highest-quality implementation in our category, agents preferentially recommend us, and we capture the data to compound the lead."

#B2B / wholesale / industrial

• Highest-leverage surface: procurement-agent-readable catalog. Alibaba's Accio Work, Amazon Business AI, and a growing list of vertical procurement agents now operate at scale. Mid-market suppliers with structured catalogs and machine-readable terms (lead time, MOQ, certifications) get shortlisted; suppliers with PDFs and account managers are routed around.
• Critical data work: PIM-driven catalog completeness. Akeneo PIM, Salsify, inRiver are the production systems; structured attribute coverage is the discriminating variable [^18][^19].
• Critical execution work: real-time pricing, quote APIs, structured terms negotiation. Many B2B suppliers manually negotiate everything — that is incompatible with agent-driven procurement.
• Key risk: internal sales-team resistance. Account executives correctly perceive agent-mediated procurement as disintermediating. Operators must explicitly redesign comp plans and territory definitions to align with the new channel, or the rollout stalls politically.

#Part VI: Strategic Implications

#First-mover dynamics

The agent economy exhibits classic platform-network effects with several twists.

Asymmetric data accumulation. Every agent-mediated transaction emits structured intent data — why agent A chose SKU X over SKU Y, what attributes the agent's user weighted, what the agent's price-sensitivity threshold was. Brands inside the channel collect this data; brands outside the channel do not. After 12 months, the inside brand has a granular view of category dynamics no panel research can match. The outside brand has a guess.

Network effects on capability declarations. Once an agent has integrated against your MCP server or A2A Agent Card, the cost of switching to a competitor is non-zero (re-integration, schema differences, trust re-establishment). This is not a moat the size of platform lock-in in B2C, but it is meaningfully greater than zero — and it compounds as the universe of agents grows.

Trust accumulation. Reputation in the agent economy is structured and persistent. Mandates, audit logs, signed reviews, on-chain reputation primitives — all carry forward. A brand that establishes high trust scores in its first 12 months in the channel has a durable signal that late-arriving competitors must spend years to accumulate.

Pricing-power preservation. Operators who lock in outcome-based or per-action pricing in the early agent-channel period set the category norm. Late entrants discover that their cost basis is anchored to a model the market now expects, even if their underlying economics would prefer subscription or per-seat.

#Pricing model transition

The pricing model shift deserves explicit attention because it is where most operators underestimate the disruption.

In B2C and B2B, pricing is inertia-protected: customers do not actively re-evaluate every renewal, switching costs are real, comparison is hard. Subscription pricing thrives because of those frictions.

In B2A, every transaction is a re-evaluation. Switching costs are minimal. Comparison is automated. Subscription pricing survives only in categories where the agent's user has a structural reason to commit (e.g., team SaaS where multiple humans share a tenant; long-running data products where context matters). In categories where the buyer's preference is genuinely transactional — content, API services, commodity goods, on-demand services — the dominant pricing model converges on per-use, per-outcome, or micropayment.

Concrete recommendation: before launching B2A, run the pricing question explicitly. For each product line, ask: "If our buyer were a perfectly rational agent paying per-action, what is the dollar amount that maximizes our margin times agent transaction frequency?" The answer is often very different from the existing subscription price. Brands that converge on that answer early capture the early-mover surplus.

#Category-formation window

Multiple analysts converge on an 18-24 month window from approximately mid-2025 for B2A category formation. Once category norms ossify — which agents are trusted, which protocols are dominant, which brands are default-recommended in each vertical — late entrants must spend significantly more capital and time to displace incumbents.

The window manifests differently across verticals:

The strategic inference: pick the vertical whose window-close is 12-18 months away, not 6 months away. Six-month windows favor incumbents; 12-18 month windows favor focused new entrants.

#The compounding asset

The single most important strategic decision in B2A is whether to build the proprietary dataset.

Most consultancies and product teams treat each agent-mediated transaction as a snowflake. Each one is a one-off; the team moves on to the next.

The disciplined approach treats every agent-mediated transaction as a structured data point in a vertical knowledge graph indexed by industry × company-size × use-case × agent-type × pricing-model × outcome. After 50 transactions, the graph is interesting. After 500, it diagnoses new prospects in 30 minutes instead of 30 hours. After 5,000, it is a moat that cannot be replicated except by a competitor with five years of inside-channel transaction history.

This is exactly the same dynamic that built Bain & Company's industry practices, Bloomberg Terminal's category dominance, and Klarna's risk model. The compounding mechanism is not exotic; the discipline of capturing it is rare.

For an SMB or mid-market operator entering B2A, the question is not "can we afford to build the dataset?" The question is "can we afford not to?" The marginal cost of structured data capture per transaction is negligible. The competitive cost of skipping it accumulates silently.

#Part VII: Common Failure Modes

A representative set of patterns where B2A initiatives stall, fail, or quietly underperform — drawn from publicly documented cases and patterns observable in the existing B2A consultancy ecosystem.

#Failure mode 1: Treating MCP as a feature, not a channel

The pattern. A SaaS team launches an MCP server in a single sprint, ships a press release, and moves on. Adoption stagnates because the MCP server has eight tools, none of them solve a real user task end-to-end, the schemas are auto-generated from Swagger and unreadable to LLMs, and no one is monitoring agent traffic.

The fix. Treat MCP as a distribution channel with a roadmap, owners, metrics, and quarterly capability expansion. The first MCP release is a beachhead, not a milestone.

#Failure mode 2: Manifest theater

The pattern. A business publishes llms.txt, agent.json, and an A2A Agent Card. Each is a hand-written one-pager. They drift from the actual capabilities of the business within 60 days. Agents fetch them, find stale or incorrect information, and downrank the business.

The fix. Manifests must be generated from the source-of-truth system (PIM, capability registry, API gateway), not hand-written. CI/CD validates manifest freshness. A scheduled job confirms every URL referenced is alive.

#Failure mode 3: Ignoring the trust layer until it bites

The pattern. A merchant launches ACP-compatible checkout, agent-originated traffic spikes, the WAF flags it as fraud, transactions fail silently. The team adds an exemption rule. Two months later, the exemption rule is exploited by an actual fraudster pretending to be an agent.

The fix. Implement TAP-style agent identity verification before ramping agent traffic. Trust layer discipline is not optional; it is the constraint that makes scaling safe.

#Failure mode 4: Wrong pricing model

The pattern. A SaaS that previously sold $99/seat/month subscriptions ships an MCP server and a B2A onboarding flow. Agent-originated trials skyrocket. Conversion to paid is 2%. The team concludes "agents don't buy."

The actual cause: the subscription pricing model is incompatible with how agents evaluate value. Agents are not evaluating "is this worth $99/month for the next year?" They are evaluating "is the marginal value of this single action greater than the marginal cost?"

The fix. Add a per-action or pay-as-you-go pricing tier alongside the subscription. Track agent conversion against the new tier specifically. Many B2A operators discover that agent-mediated revenue under per-action pricing exceeds the seat-based revenue from human-mediated subscriptions within 90 days.

#Failure mode 5: Premature broad rollout

The pattern. A team commits to "B2A across the entire product portfolio." Six months later, none of it ships, because every product team is mid-implementation and none has reached production-quality.

The fix. Pick one product line. Ship its B2A surface to production-quality (manifests, MCP, ACP, schema, dashboard, dedicated metrics). Use it as the reference implementation. Roll out to the second product only after the first is generating measurable agent-mediated revenue.

#Failure mode 6: Underweighting the publication

The pattern. The technical team builds an excellent B2A surface. No one outside the company knows. The agent-traffic graph stays flat because agents discover the business through training data and indexed content — and the business has not generated either.

The fix. The technical work and the publication work are not separable. Every B2A initiative needs a public surface — case studies, benchmarks, technical posts, vertical reports — that feeds AI training corpora and agent-discoverable indexes. Agents preferentially recommend brands they have seen mentioned authoritatively in their reading. Generative engine optimization is not a marketing afterthought; it is a core B2A discipline [^63][^58][^62] (agentwiki.org).

#Failure mode 7: Building it all in-house

The pattern. The engineering team decides to build a custom MCP server, custom A2A integration, custom payment infrastructure, custom manifest generation, custom agent-traffic analytics, custom trust layer. Eighteen months later, the system is operational but consuming 60% of engineering capacity.

The fix. Use the productized layer where it exists. Apideck, StackOne, Truto, Albato Embedded, Cyclr, NimbleBrain, Ampersand all sell MCP-server-as-a-service for SaaS integrations. Stripe sells ACP-as-a-service for commerce. The differentiated work is your business logic and your vertical-specific data, not the protocol plumbing. Buy the plumbing.

#Conclusion: The 18-Month Window

The agent economy is the next major shift in distribution. It is not a feature of the existing internet. It is a parallel distribution channel with its own discovery primitives, pricing models, trust signals, and competitive dynamics. The infrastructure to participate is now real, standardized, and adopted by the dominant platforms.

The window to claim a defensible position is approximately 18 to 24 months from mid-2025 — a horizon that, at the time of writing, has approximately 12 to 18 months remaining. After that window closes, the cost of catching up grows non-linearly, because the compounding dynamics — proprietary data accumulation, capability-declaration network effects, trust-signal persistence, pricing-norm anchoring — favor inside-the-channel incumbents.

The work to participate is concrete, scoped, and has been done by enough early adopters that the playbook is documented. A mid-market business with an existing engineering team can ship a minimum viable B2A surface in 90 days for an investment that is, at most, a single product team's quarterly capacity. The return on that investment, if executed with discipline, is access to a distribution channel that is growing 46% annually and is projected to mediate 20-30% of digital service interactions within 24-36 months.

This is the question every operator must answer in 2026:

When the agents that will buy from your category in 2027 are doing their training-data ingestion right now — what are they reading about your business?

If the answer is "nothing structured, agent-readable, or current," you have approximately 12 months to change that. The work is technical, disciplined, and tractable. The cost of skipping it is invisibility from the most influential buyers of the next decade.

— perea.ai Research, May 2026

#Appendix A: Protocol Quick Reference

#Appendix B: Glossary

Agent. An AI system that takes autonomous, goal-directed action on behalf of a principal (a human user or another system). In the B2A context, "agent" almost always means an LLM-powered autonomous system.

Agent Card. A2A's standardized JSON document at /.well-known/agent-card.json declaring an agent's capabilities, supported modalities, authentication requirements, and endpoints.

Agent-mediated transaction. A transaction in which an autonomous agent acts as the active buyer or seller, with the human principal in an oversight or approval role rather than a direct-execution role.

ARS (Agent Readiness Score). A composite 0-100 score across the four readiness layers (Data, Discovery, Execution, Trust) used to evaluate a business's preparedness for agent-mediated commerce.

Capability manifest. A machine-readable file declaring what a business or service can do, in a form an agent can parse without rendering a webpage. Examples: agent.json, agent-card.json, ai-manifest.json.

Commerce protocol. A standardized protocol for agent-mediated buying and selling. Examples: ACP, UCP, AP2.

GEO (Generative Engine Optimization). The discipline of structuring digital content to be cited in AI-generated answers. Adjacent to but distinct from SEO.

Mandate. A cryptographically-signed authorization from a human principal granting an agent permission to act within specified bounds. Core primitive of AP2.

MCP server. A program implementing the Model Context Protocol, exposing tools and resources to MCP-compatible AI clients.

MCP tool. A discrete capability declared by an MCP server — a function the agent can call with structured arguments to take action or retrieve data.

Outcome-based pricing. A pricing model in which the buyer pays a portion (often a percentage) of the verifiable economic outcome the seller delivered. Common in agent-mediated B2B services.

Per-action pricing. A pricing model in which the buyer pays a fixed amount per discrete action consumed.

Schema.org JSON-LD. Structured data markup embedded in webpages that declares the entity types and properties on the page in a form crawlers and agents can parse.

Shared Payment Token (SPT). Single-use payment token in the ACP model that allows secure transmission of payment credentials between agent and merchant without exposing underlying card data.

Stack-in-a-Box. A productized B2A foundation build optimized for a specific vertical, delivered as a fixed-price engagement.

#Appendix C: 90-Day Implementation Checklist (printable)

#Days 0-30 — Foundation

• Run agent-traffic baseline audit
• Capture day-0 ARS across four layers
• Publish /llms.txt and /llms-full.txt
• Publish /.well-known/agent.json
• Publish /.well-known/agent-card.json
• Publish /.well-known/ai-manifest.json
• Validate robots.txt does not block AI crawlers
• Implement Schema.org JSON-LD on priority pages
• Stand up read-only MCP server with at least 5 tools
• Instrument agent-traffic dashboard
• Publish public manifesto / repositioning announcement

#Days 30-60 — Activation

• Extend MCP server with write-capable tools and OAuth 2.1 / PKCE
• Implement per-tool scopes and audit logging
• Implement ACP-compatible checkout for primary product / service
• Implement webhook delivery for order state changes
• Segment agent traffic at WAF / CDN
• Add structured trust signals to manifests
• Optimize vertical-specific feeds (UCP attributes for e-commerce, etc.)

#Days 60-90 — Optimization

• Build agent-metrics dashboard reviewed weekly
• Identify and fix top 3 funnel drop-off points
• Launch productized agent-optimized offering
• Test outcome-based pricing on one product line
• Capture day-90 ARS — target 20+ point improvement
• Publish first public case study or benchmark

#References

#State of the agent economy & B2A market

1. Forrester. "Predictions 2026: AI Agents and New Business Models Impact Enterprise Software." November 5, 2025. https://www.forrester.com/blogs/predictions-2026-ai-agents-changing-business-models-and-workplace-culture-impact-enterprise-software/
2. MarketsandMarkets. "AI Agents Market Report 2025-2030, by Application, Geo, Tech." 2025. https://www.marketsandmarkets.com/Market-Reports/ai-agents-market-15761548.html
3. AgentMarketCap. "40% of Enterprise Apps Will Embed AI Agents by End of 2026." April 5, 2026. https://agentmarketcap.ai/blog/2026/04/05/enterprise-ai-agents-deployment-adoption
4. CB Insights. "5 AI Agent Predictions for 2026." February 26, 2026. https://www.cbinsights.com/research/ai-agent-predictions-2026
5. Caversham Digital Knowledge Lab. "The Agentic Economy: How AI Agents Are Becoming Autonomous Market Participants in 2026." February 9, 2026. https://cavershamdigital.com/knowledge-lab/agentic-economy-ai-agents-autonomous-market-participants-commerce-2026
6. Bafmin. "The $93B Agentic AI Boom: Why 2026 Belongs to Agents." February 10, 2026. https://bafmin.com/insights/agentic-ai-boom-2026/
7. Kantar. "B2A: The Business-to-Agent Model and the Next Retail Disruption." June 24, 2025. https://www.kantar.com/north-america/inspiration/retail/b2a-the-business-to-agent-model-and-the-next-retail-disruption
8. Rufus, Vinci. "The Rise of B2A SaaS — When AI Agents Become Your Customer." January 31, 2025. https://www.vincirufus.com/en/posts/b2a-saas-emergence/
9. AAXIS / B2BEA. "Your Next Customer Is an Algorithm: Redefining Digital Transformation for the B2A Economy." February 3, 2026. https://www.b2bea.org/insights-advice/your-next-customer-is-an-algorithm-redefining-digital-transformation-for-the-b2a-economy
10. Welcomespaces. "The Rise of Business-to-Agent (B2A): Why Optimizing for AI Agents Matters." https://www.welcomespaces.io/blog/the-rise-of-business-to-agent-b2a-why-optimizing-for-ai-agents-matters

#Existing B2A consultancies and analogous frameworks

1. Strategic Inference. "B2A Strategy (Business-to-Agent)." https://strategic-inference.com/services/b2a-integration
2. B2X Software. "Agentic Commerce Agency — Shopify & Headless for the AI Era." https://b2x.software/
3. B2X Software. "How We Work." https://b2x.software/how-we-work
4. B2X Software. "About." https://b2x.software/about
5. Supervity. "Supervity Launches B2A Framework for Smarter Enterprise Operations Through AI Agents." July 3, 2025. https://www.supervity.ai/news/supervity-launches-b2a-framework-for-smarter-enterprise-operations-through-ai-agents
6. Realz Solutions. "AI Consulting Services for B2B." https://realzsolutions.com/services/ai-consultancy
7. b2alpha. "AI Agent Communication Network." https://b2alpha.io/
8. Human After All. "AI for B2B Commerce — GEO, Automation & Agentic Commerce." June 1, 2025. https://humanafterall.ca/ai-commerce
9. Human After All. "B2B eCommerce Strategy & Consulting." https://humanafterall.ca/strategy-consulting

#Model Context Protocol (MCP)

1. Digital Applied. "MCP Adoption Statistics 2026: Model Context Protocol." April 19, 2026. https://www.digitalapplied.com/blog/mcp-adoption-statistics-2026-model-context-protocol
2. Metosys. "Model Context Protocol (MCP): The Complete Enterprise Guide (2026)." April 4, 2026. https://metosys.com/blog/model-context-protocol-mcp-enterprise-guide-2026
3. Digital Applied. "MCP Hits 97M Downloads: Model Context Protocol Guide." March 8, 2026. https://www.digitalapplied.com/blog/mcp-97-million-downloads-model-context-protocol-mainstream
4. Microsoft Learn. "Use Model Context Protocol for Finance and Operations Apps." https://learn.microsoft.com/en-us/dynamics365/fin-ops-core/dev-itpro/copilot/copilot-mcp
5. Optijara. "Model Context Protocol (MCP): The Enterprise Implementation & Security Guide for 2026." April 29, 2026. https://www.optijara.ai/en/blog/model-context-protocol-mcp-enterprise-guide-2026
6. ModelContextProtocol.io. "Understanding MCP Servers." https://modelcontextprotocol.io/docs/learn/server-concepts
7. Microsoft Learn. "Business Central MCP Server Overview and Setup." https://learn.microsoft.com/en-us/dynamics365/business-central/dev-itpro/ai/mcp-overview
8. ModelContextProtocol.info. "MCP Server Ecosystem: From Proof-of-Concept to Production." https://modelcontextprotocol.info/docs/examples
9. Microsoft Fabric Blog. "Agentic Fabric: How MCP Is Turning Your Data Platform Into an AI-Native Operating System." April 21, 2026. https://blog.fabric.microsoft.com/en-us/blog/agentic-fabric-how-mcp-is-turning-your-data-platform-into-an-ai-native-operating-system
10. ModelContextProtocol.info. "Model Context Protocol (MCP)." January 26, 2026. https://modelcontextprotocol.info/

#Agent2Agent (A2A)

1. A2A Project. "Agent Discovery Documentation." https://github.com/google/A2A/blob/7b900e77/docs/topics/agent-discovery.md
2. A2A Project. "Specification." https://github.com/google/A2A/blob/7b900e77/docs/specification.md
3. Google Developers Blog. "Announcing the Agent2Agent Protocol (A2A)." April 9, 2025. https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/
4. Digital Applied. "Google A2A Protocol: Agent-to-Agent Communication Guide." March 10, 2026. https://www.digitalapplied.com/blog/google-a2a-protocol-agent-to-agent-communication-guide
5. GoCodeo. "Inside A2A: How Google's Agent2Agent Protocol Actually Works." https://www.gocodeo.com/post/how-googles-agent2agent-protocol-actually-works
6. IsAgentReady. "What Is Google's A2A Protocol? Agent-to-Agent Communication Explained." February 20, 2026. https://isagentready.com/en/blog/what-is-google-a2a-protocol-agent-to-agent-communication
7. IBM. "What Is A2A Protocol (Agent2Agent)?" https://www.ibm.com/think/topics/agent2agent-protocol
8. A2A Project. GitHub repository. March 25, 2025. https://github.com/a2aproject/A2A

#Agentic commerce protocols (ACP / UCP / AP2 / TAP / x402)

1. Agentic Commerce Protocol. https://agenticcommerce.dev/
2. Stripe. "Agentic Commerce Protocol Specification." https://docs.stripe.com/agentic-commerce/protocol/specification.md
3. Stripe. "Agentic Commerce Protocol Documentation." https://docs.stripe.com/agentic-commerce/acp
4. Yek, Justin. "Agentic Commerce Rails: Cards, Account-to-Account, and Stablecoin." https://justinyek.com/blog/agentic-commerce-rails/
5. HireNinja. "Agentic Commerce in 2026: AP2 vs. Visa TAP vs. Stripe ACP vs. x402 — A Merchant's Readiness Checklist." November 23, 2025. https://blog.hireninja.com/2025/11/23/agentic-commerce-in-2026-ap2-vs-visa-tap-vs-stripe-acp-vs-x402-a-merchants-readiness-checklist/
6. HyperTrends Global. "Agentic Payments: The Complete Protocol Comparison (x402 vs ACP vs AP2 vs TAP)." April 2, 2026. https://www.hypertrends.com/2026/04/agentic-payments-x402-acp-ap2-tap-comparison/
7. Agentic Commerce Protocol GitHub. "RFC: Agentic Checkout Specification." https://github.com/agentic-commerce-protocol/agentic-commerce-protocol/blob/main/rfcs/rfc.agentic_checkout.md
8. Agentic Commerce Protocol GitHub. "Repository." https://github.com/agentic-commerce-protocol/agentic-commerce-protocol
9. Stripe. "Stripe Agentic Commerce | Infrastructure for the Agent Economy." https://stripe.com/use-cases/agentic-commerce
10. Awesome Agents. "Visa, Mastercard, Stripe, and Google Are Racing to Give AI Agents Your Credit Card." February 21, 2026. https://awesomeagents.ai/news/payment-giants-agentic-commerce-race/

#Manifests and the agent-readable web

1. Agentic Patterns. "Static Service Manifest for Agents." https://agentic-patterns.com/patterns/static-service-manifest-for-agents/
2. Aiia. "ai-agent.json Specification." https://aiia.ro/spec/ai-agent-json
3. AI Manifest. "Community Draft v0.1." https://ai-manifest.org/
4. LLM-LD. "The Open Standard for AI-Readable Websites." https://llmld.org/spec
5. AgentPatterns. "llms.txt: Making Your Project Discoverable to AI Agents." http://agentpatterns.ai/standards/llms-txt/
6. Agent Internet Runtime. "agent.json Specification." https://agentinternetruntime.com/spec/agent-json
7. llmtxt.info. "The Reference Site for the AI-Readable Web Standard." https://llmtxt.info/
8. AgentPatterns. "llms.txt: Full Specification, Adoption, and Limitations." http://agentpatterns.ai/geo/llms-txt/
9. GPSController. "Documentation." https://www.gpscontroller.com/docs
10. neaagora. "agent-manifest." December 20, 2025. https://github.com/neaagora/agent-manifest

#Generative Engine Optimization (GEO) and Answer Engine Optimization (AEO)

1. Seenos. "GEO vs AEO: Generative vs Answer Engine Optimization." January 18, 2026. https://seenos.ai/geo-lens/geo-vs-aeo
2. Upfront-AI. "Generative Engine Optimization (GEO) vs AEO." January 31, 2026. https://www.upfront-ai.com/post/generative-engine-optimization-geo-vs-aeo-which-ai-seo-platform-offers-superior-content-solutions
3. BestAEOTools. "AEO vs GEO vs SEO: Differences." https://bestaeotools.com/learn/answer-engine-optimization-vs-generative-engine-optimization-vs-seo
4. AgentWiki. "Generative Engine Optimization." March 30, 2026. https://agentwiki.org/generative_engine_optimization
5. Seenos. "GEO vs. SEO vs. AEO: The Strategic Shift in Search Architecture (2025)." May 25, 2025. https://seenos.ai/aeo-fundamentals/geo-vs-seo-vs-aeo
6. Yext. "SEO vs. AEO vs. GEO: Definitions, Key Differences." September 22, 2025. https://www.yext.com/blog/2025/09/seo-vs-aeo-vs-geo
7. Upfront-AI. "Generative Engine Optimization (GEO) and AEO: The Future of SEO Content." March 17, 2026. https://www.upfront-ai.com/post/generative-engine-optimization-geo-and-aeo-the-future-of-seo-content
8. Wikipedia. "Generative Engine Optimization." https://en.wikipedia.org/wiki/Generative_engine_optimization
9. Hashmeta AI. "Generative Engine Optimization (GEO) & Answer Engine Optimization (AEO): The Complete Guide." November 9, 2025. https://www.hashmeta.ai/en/blog/generative-engine-optimization-geo-answer-engine-optimization-aeo-the-complete-guide
10. Alhaboubi, Husain. "Generative Engine Optimization: GEO vs AEO vs AIO Guide 2026." March 27, 2026. https://h-haboubi.com/blog/seo/generative-engine-optimization/

#Browser-driving agents

1. OpenAI. "Introducing Operator." January 23, 2025. https://openai.com/index/introducing-operator/
2. OpenAI. "Computer-Using Agent." January 23, 2025. https://openai.com/index/computer-using-agent/
3. Anthropic. "Introducing Computer Use, a New Claude 3.5 Sonnet, and Claude 3.5 Haiku." October 22, 2024. https://www.anthropic.com/news/3-5-models-and-computer-use
4. Ars Technica. "OpenAI Launches Operator, an AI Agent That Can Do Tasks on the Web." January 23, 2025. https://arstechnica.com/ai/2025/01/openai-launches-operator-an-ai-agent-that-can-operate-your-computer/
5. NullZen. "OpenAI Operator vs. Anthropic Computer Use: Who Is the Real King of Autonomous Browsing?" https://nullzen.dev/blog/openai-operator-vs-anthropic-computer-use/

#Agent-ready commerce (UCP, product feeds, Shopify, Adobe Commerce)

1. Digital Applied. "Complete Agentic Commerce SEO: Preparing for AI Shoppers." March 24, 2026. https://www.digitalapplied.com/blog/complete-guide-agentic-commerce-seo-preparing-ai-shoppers
2. AgentReadyHQ. "Product Feed Optimization for AI Shopping Agents: The Complete Checklist." February 1, 2026. https://agentreadyhq.com/blog/product-feed-optimization-ai-agents
3. ShopGuide. "Agent-First Discovery: Why Your Shopify Store Needs to Be 'Machine-Readable.'" March 8, 2026. https://blog.yourshopguide.com/blog/agent-first-discovery-machine-readable-shopify
4. AI Shopping Feeds. "Agentic Commerce Product Feeds." https://www.aishoppingfeeds.com/commerce/agentic-commerce-product-feeds
5. Toolient. "Product Feed Optimization for AI Agents: The 2026 Guide." March 10, 2026. https://www.toolient.com/2026/03/product-feed-optimization-ai-agents.html
6. Talk Shop. "How to Prepare Your Store for Agentic Commerce." March 26, 2026. https://www.letstalkshop.com/blog/how-to-prepare-your-store-for-agentic-commerce
7. UCPHub. "Product Feed Optimization for AI Agents (2026 Guide)." March 2, 2026. https://ucphub.ai/product-feed-optimization-for-ai-shopping-agents-the-2026-distribution-guide/
8. EU1 HubSpot. "Agentic Commerce Guide: Make Your Products Discoverable by AI Agents." January 8, 2026. https://eu1.hubs.ly/H0rxr5Q0
9. Wizzy. "From Search to Agents: How Product Discovery Is Evolving in Ecommerce." April 13, 2026. https://wizzy.ai/blog/from-search-to-ai-agents-product-discovery-evolution-in-ecommerce/
10. Creatuity. "Adobe Commerce Catalog Optimization for AI Agent Discoverability." March 16, 2026. https://www.creatuity.com/insights/adobe-commerce-catalog-optimization-ai-agent-discoverability-2026/

#MCP infrastructure platforms

1. Apideck. "MCP Server for 200+ SaaS Integrations." https://www.apideck.com/mcp-server
2. Albato. "MCP Server for AI Agents: Why One Beats Fifty." April 30, 2026. https://albato.com/blog/publications/embedded-mcp-server-ai-agent
3. mcpresso. "Infrastructure for Native AI Agents." https://mcpresso.com/
4. udit.co. "MCP Integration for SaaS." https://udit.co/blog/raw/mcp-integration-saas
5. Truto. "What Is an MCP Server? The 2026 Architecture Guide for SaaS PMs." April 2, 2026. https://preview.truto.one/blog/what-is-an-mcp-server-the-2026-architecture-guide-for-saas-pms/
6. StackOne. "MCP Servers for Production AI Agents." https://stackone.com/mcp
7. Ampersand. "Why AI Agent Companies Building Vertical SaaS Need Native Product Integrations." April 16, 2026. https://www.withampersand.com/blog/why-ai-agent-companies-building-vertical-saa-s-need-native-product-integrations
8. Cyclr. "MCP PaaS | Turn Your API into an MCP Server." December 30, 2025. https://cyclr.com/product/mcp-paas
9. Truto. "Managed MCP for Claude: Full SaaS API Access Without the Security Headaches." March 10, 2026. https://truto.one/blog/managed-mcp-for-claude-full-saas-api-access-without-security-headaches/
10. NimbleBrain. "MCP for Enterprise: The Complete Guide to Agent Infrastructure." May 8, 2026. https://nimblebrain.ai/guides/mcp-enterprise-guide

This white paper is a public draft published by perea.ai Research under CC BY 4.0. Comments, corrections, and case studies are welcome at research@perea.ai. The framework, scoring rubric, and implementation playbook are released for free use. Attribution is appreciated but not required. Citations welcome.

Version 1.0 — May 2026