← Notes SMEs

Agentic Web and Business to Agent: Getting Your Company Ready for AI Agents

Pasquale Caiazzo / / 12 min read

In brief We're moving from the Generative Web, where people use AI to search, to the Agentic Web, where autonomous AI agents search, select and buy on their behalf. For a company the counterpart changes: no longer just a customer to convince, but an agent to be chosen by. This is the Business to Agent model - B2A, not to be confused with the older Business-to-Administration. Getting ready doesn't mean rebuilding the business, but adding a machine-readability layer on top of it: exposed data, verifiable identity, frictionless transactions. Whoever moves first lowers their acquisition cost and becomes the default supplier; whoever doesn't risks invisibility.

From the Generative Web to the Agentic Web

Over the past two years we've grown used to one idea: people ask AI instead of searching Google. That's the Generative Web, and it has already changed how a customer discovers a supplier. But it's a transitional phase. What's coming is different: AI agents that don't just answer, but act - they search, compare, select and complete a purchase on a person's behalf. That's the Agentic Web, and it moves the point at which a customer is won or lost.

The difference is sharp. An answer engine retrieves and synthesizes: it names you in a response. An agent decides and acts: it selects you, checks the terms and, increasingly, buys. The first level requires being well described; the second requires being operationally ready. For a company this introduces a model I call Business to Agent. I've written the underlying thesis elsewhere; here I look at the operational side: what concretely changes in the funnel, in costs and in technical requirements when there's no longer a human eye on the other side, but a machine that decides.

B2A: a necessary disambiguation

A note first, because it's exactly the kind of confusion that is my craft. The acronym B2A is ambiguous. It has historically meant Business-to-Administration, the relationship between a company and public administration; you'll also find it used for Business-to-Affiliate or, more recently, Business-to-Algorithm. When I write B2A on this page I mean one precise thing: Business-to-Agent, the company facing the AI agent that buys on a person's behalf.

This isn't pedantry. It's precisely the mistake an AI system makes when an entity isn't declared with precision: it takes an ambiguous label and attaches it to the most frequent meaning, not the correct one. Disambiguating a term - stating what it refers to, and what it doesn't - is the first move of anyone who wants to be understood by machines. I'm doing it here on my own vocabulary.

From B2C/B2B to B2A: the counterpart changes

In traditional marketing you optimise channels to capture a person's attention: design, persuasive messaging, urgency. In B2A a new step inserts itself, and the recipient of that step doesn't feel emotion. The person remains, and the final word is still theirs; but between them and you sits an agent acting as a filter, and that filter doesn't read your landing page: it reads your data.

Model Counterpart Main lever Trust signal Metric
B2C Person (consumer) Persuasion, emotion, brand Social proof, aesthetics Conversions, clicks
B2B Person in an organization Relationship, ROI, references Case studies, references Leads, pipeline
B2A AI agent (on a person's behalf) Readability, declared data Verifiability (sameAs, registries, citations) Citation and selection rate

The row that matters is the last one: the lever, the trust signal and the metric all change. You no longer win with the best message, but with the most readable and verifiable data; and you no longer measure clicks, but how often you get chosen.

A B2A transaction, in concrete terms

To grasp the scale of the shift, it helps to follow a transaction in three stages. A procurement lead tells their assistant: «Buy 20 licences of a project-management tool for a team of developers in Italy, max €15 per user per month, with GDPR-compliant servers in Europe.»

The agent doesn't open suppliers' graphic homepages. It scans the web for declared facts: it reads llms.txt files, queries structured data, extracts price, legal compliance and server location. It narrows the field to who is readable, and discards who isn't - not because they're worse, but because they're opaque.

Having found the best option, it executes. It connects to the supplier's systems, checks real availability and completes the purchase through payment protocols designed for agents - like the Agentic Commerce Protocol, the open standard co-developed by OpenAI and Stripe, in which a shared payment token lets the agent close the deal without exposing card details. The human never opened a browser.

The infrastructure is already here

It's worth pausing on one point, because it's what separates a forecast from a fact: the technical stack of the Agentic Web isn't on its way, it has already been built, in little over a year.

Anthropic released the Model Context Protocol (MCP) in November 2024, the open standard connecting agents to a company's data and tools; in 2025 it was adopted by OpenAI and Google too. Google introduced Agent2Agent (A2A) to let agents from different vendors talk to each other, and then the Agent Payments Protocol (AP2) for agent-to-agent payments. On the card networks, Visa launched Intelligent Commerce and Mastercard unveiled Agent Pay, both in April 2025. And in September 2025 OpenAI turned on Instant Checkout inside ChatGPT, built with Stripe on the Agentic Commerce Protocol. With over 700 million people using ChatGPT every week, the agent that buys isn't a distant promise: it already has rails to run on.

How a buyer that is a machine reasons

Unlike a person, an agent isn't swayed by aesthetics or tone. It chooses on objective criteria, and there are three.

  • Data retrievability. Prices, features, availability and terms must be readable by the machine instantly, not buried in an image or a PDF.
  • Verifiable authority. The agent checks reviews, citations and third-party evidence to estimate a trust score anchored to sources, not promises.
  • Absence of friction. All else being equal, it picks whoever allows an immediate transaction or booking over whoever imposes forms and manual steps.

The commercial impact: why B2A lowers CAC

Getting ready for the Agentic Web isn't only a technical matter, it's a lever on the income statement. When a customer delegates a purchase, they do it at the moment of peak intent: the agent intercepts exactly that demand and skips much of the traditional funnel, with its advertising costs. Being the supplier the agent selects means lowering your acquisition cost, because the AI does the filtering in place of campaigns.

There's also a retention effect: once a customer's agent has registered your company as a reliable, «machine-compatible» supplier, recurring purchases tend to repeat automatically. And the stakes are large: Gartner forecasts that by 2028, 90% of B2B buying will be intermediated by AI agents, with over $15 trillion of spend flowing through agent exchanges - these are projections, but they point in a direction. On the consumer side the phenomenon is already measurable: according to Adobe Analytics, during the 2025 holiday season traffic to e-commerce sites from generative-AI sources grew by nearly 700% year on year.

The flip side is the risk of invisibility. «Zero-click» searches, where the answer or the purchase closes inside the ChatGPT, Gemini or Perplexity interface, are growing fast. Not being ready for B2A isn't a theoretical disadvantage: it's ceding market share to whoever is already readable.

Three requirements to become agent-ready

You don't need to program anything yourself. You need three pieces of infrastructure, to be requested from the technical team with a clear rationale.

A content map for machines (llms.txt). Agents need a schematic summary of the business so as not to waste compute. An llms.txt file at the site root - following the standard documented at llmstxt.org - works like a sitemap for AI: it excludes the graphic code and exposes only the information of value.

A verifiable identity (JSON-LD and Wikidata). For an AI to trust the brand, company data must be structured with the Schema.org vocabulary and linked to recognized nodes like Wikidata, the open knowledge base of the Wikimedia galaxy. This is the work of entity linking: connecting the brand unambiguously to its specialism. I call it linking, not «resolution», on purpose: the goal isn't to merge duplicate records, but to link your entity to the right ones.

An interoperability channel (MCP). The Model Context Protocol, the open standard promoted by Anthropic, lets AI applications connect securely to a company's data. Exposing an MCP channel allows an agent to check availability and terms in real time, and therefore to book or buy without friction.

GEO: influencing AI responses

When someone asks ChatGPT or Perplexity «who are the best business-design consultants in Italy?», the model uses RAG - Retrieval-Augmented Generation - to pull information from the web and compose the answer. Appearing in that answer, with an active citation, is the goal of Generative Engine Optimization.

The study that formalised the discipline, «GEO: Generative Engine Optimization» (researchers from Princeton, Georgia Tech, IIT Delhi and the Allen Institute for AI, presented at KDD 2024), shows that structured optimisation can increase a piece of content's visibility in generative responses by up to 40%. In practice, information density matters: text rich in data, numbers and verifiable statements is preferred over long but generic text. And co-mention matters: if your brand appears often alongside the right concepts on authoritative sources, the model tends to treat you as a reference. The metric is no longer the site visit, but how often you're cited as a source. On that level - getting found and cited by AI - I've gathered the fundamentals and cases in my GEO guide.

An honest checklist to see if you're ready

Five questions tell you in a few minutes how readable your company already is to agents.

  • Does your site allow crawling by AI bots (GPTBot, ClaudeBot, PerplexityBot) or block them outright?
  • Is there an llms.txt, a machine-readable version of the site?
  • Are products and services described with validated Schema.org and linked to authoritative external entities?
  • Can your checkout or booking flow integrate with agentic commerce protocols?
  • Do you measure, even just once a month, how and how often your brand is cited on ChatGPT and Perplexity?

If most answers are «no», it's not an alarm: it's the starting snapshot. From there you decide where to act.

Where to start

Getting a company ready for the Agentic Web doesn't mean redoing everything. It means adding a machine-readability layer on top of the business you already have: declaring the data that matters, anchoring it to verifiable sources, removing friction from transactions. The first step is always the same - ask the main AI systems what they know today about your company and your products, and observe what's correct, what's missing, and what's invented.

I work with companies to build this layer: making data, offer and identity readable and citable by AI systems, and ready for the arrival of agents.

Book a call ↗ For SMEs →

Key references