AI Agents and the Next Phase of Search: How to Prepare Your Business Now

By acezhuo@gmail.com | August 22, 2026

AI agents complete multi-step tasks on a user’s behalf, including research, comparison, and increasingly transaction. Preparing for them means making your business machine-readable rather than integrating against any single standard, because the protocol landscape is unsettled and the data work required is identical across all of them.

What Changes When the Buyer Is Software

An agent does not browse, skim, or respond to persuasion. It parses.

Human visitor AI agent
Reads Selectively, visually Structurally, completely
Persuaded by Design, copy, social proof Verifiable data
Handles ambiguity Infers from context Fails or skips
Needs A reason to choose Machine-readable terms
Abandons when Confused or unconvinced Data is missing or inconsistent

The practical consequence is that everything a human infers must be stated. Pricing implied by a demo request, availability implied by a contact form, and compatibility implied by a case study are all invisible to a system that needs the value rather than the impression.

Ambiguity is the failure mode. Agents do not compensate for missing information the way people do; they select an option where the information is present.

This inverts a long-standing commercial instinct. Withholding pricing to force a sales conversation worked because a human who wanted the product would make contact. An agent comparing twelve options against a stated budget will simply exclude the one that did not publish a number, and no inquiry, no bounce, and no signal of any kind will record that it happened.

How Agents Differ From Chat Assistants

The distinction matters for planning, because the two require overlapping but not identical preparation.

  • Assistants answer. They retrieve sources and compose a response, and the human acts on it.
  • Agents act. They complete a sequence of steps, evaluating options and executing decisions against criteria the user set.
  • Assistants need citable passages. Agents need structured data they can compare and transact against.
  • Assistants operate in one turn. Agents operate across many, holding state between steps.

Everything covering artificial intelligence optimization for assistants remains a prerequisite. An agent that cannot find you during research will not transact with you afterward.

The two also share the same failure points. Crawler access, indexation, entity clarity, and structured data determine whether either can use you, which means a business already doing assistant visibility work has completed most of the agent preparation without labeling it that way.

What an Agent Needs That a Human Does Not

Six things, in rough order of how often they are missing:

  1. Explicit pricing. Bands, tiers, or a stated model. Contact-for-pricing removes you from any comparison that includes cost.
  2. Availability and lead times as data rather than as reassurance.
  3. Specifications as attributes. Dimensions, compatibility, capacity, and constraints stated in fields, not prose or images.
  4. Terms stated plainly. Returns, cancellation, minimum commitments, and jurisdictional limits.
  5. Stable identifiers. Consistent product and service names across your site, feeds, and third-party listings.
  6. Accessible interfaces. Content and functionality that do not require interaction, login, or script execution to reveal.

The pattern connecting them: an agent can only compare what it can extract, and it can only extract what has been stated.

Explicit pricing is the item that generates the most internal resistance and deserves the most consideration. The objection is legitimate in categories where pricing genuinely depends on scope, and the workable answer is rarely full transparency. A stated model, a starting figure, a typical range, or the variables that determine cost all give a system something to compare against, and all of them beat a contact form in any prompt that mentions budget.

Where Standards Currently Stand

Several protocols emerged in quick succession, addressing different layers of the same problem. They compose rather than compete.

Standard Origin Layer
MCP, Model Context Protocol Anthropic, November 2024 Connecting systems to data and tools
A2A, Agent2Agent Google, April 2025 Agent-to-agent discovery and delegation
AP2, Agent Payments Protocol Google, September 2025, donated to the FIDO Alliance April 2026 Cryptographic proof a user authorized a purchase
ACP, Agentic Commerce Protocol OpenAI and Stripe Checkout handshake between agent and merchant
UCP, Universal Commerce Protocol Google and Shopify Full journey: discovery, cart, checkout, orders

AP2 launched with more than sixty partners spanning card networks, processors, and wallets. Card networks have since built protocol-agnostic on-ramps on top of the whole set, which suggests merchants will work through their existing payment service provider rather than picking a winner.

The division of labor is worth understanding even if you never implement any of it. One layer establishes what is being sold and how a cart is assembled. Another executes the checkout inside an AI surface. A third proves cryptographically that a specific user authorized a specific amount, which is the problem that made autonomous agent purchasing legally difficult in the first place. Underneath all of it sits the connectivity layer that lets a system reach product data at all.

The authorization layer is the genuinely novel piece. Before it existed, agent purchases happened through stored credentials with no way for a merchant or network to distinguish a human-initiated transaction from an automated one, or to verify that an agent had acted within the scope its user set. That gap was a liability problem more than a technical one, and it is the reason institutional backing arrived quickly.

The Adoption Reality

The infrastructure is ahead of the behavior, and any honest readiness guide has to say so.

OpenAI’s Instant Checkout, the flagship consumer surface for ACP, launched with Etsy US sellers and a small group of Shopify brands. It was shut down in March 2026 after roughly five months, reportedly with minimal sales. The protocol continues; the in-chat checkout product did not.

Three conclusions follow:

  • Standards existing does not mean buyers are using them. Consumer behavior is moving more slowly than the specifications.
  • Betting engineering time on one protocol is premature. One flagship implementation has already been withdrawn.
  • The underlying data work is not premature at all. Complete attributes, explicit pricing, and clean structured data pay for themselves in conventional and assistant visibility today.

Prepare the data, not the integration. This is the central recommendation, and it is the one that holds regardless of which standard prevails.

There is a reasonable counterargument for platform-hosted merchants specifically. Where a commerce platform builds the integration on your behalf and enabling it is a configuration change rather than an engineering project, the cost of participating is low enough that the option is worth taking. The advice against building applies to bespoke integration work, not to switching something on.

What to Build Now and What to Watch

Build now Watch
Complete, structured product and service data Which protocols your payment provider supports
Explicit pricing and availability Whether consumer adoption follows the infrastructure
Consistent identifiers across all properties How card networks consolidate their agent frameworks
Accessible content without interaction gates Whether discovery manifests become widely consumed
Clean feeds shared across advertising and organic Regulatory treatment of agent-initiated transactions

The left column has immediate returns independent of agents. Structured, accessible, unambiguous data improves conventional search, assistant citation, and internal operations at the same time, which makes it a defensible investment even if agent commerce takes several more years to reach volume.

Sequence it by what is already broken. Most businesses find that explicit pricing and complete attributes are the two items failing hardest, and both are content decisions rather than engineering projects. Identifier consistency across site, feeds, and third-party listings usually comes third and takes longest, because it requires agreement between teams that have never had to agree before.

The Access Question

Agent readiness collides directly with the crawler control decisions many organizations made recently.

Infrastructure providers now categorize AI traffic into search, agent, and training, with defaults tightening on the latter two. A business that blocked broadly to protect its content may find it has also blocked the agents that would transact with it, and the block typically happens at the network layer before any file on your site is consulted.

Three checks worth running:

  • Confirm which categories you are blocking, and whether that decision was made deliberately or inherited from a default.
  • Test what an agent actually receives, not just what your configuration says it should.
  • Re-check after every migration or security policy change, since these settings are frequently reset without anyone noticing.

Sound website structure and clean access configuration are the same work that supports assistant visibility, which is the argument for treating this as one program rather than two.

The decision itself is legitimate either way. A publisher whose content is its product has good reasons to restrict access, and a business whose website exists to generate inquiries has good reasons to permit it. What is not defensible is discovering the decision was made by a default setting during a migration, which is how most businesses currently arrive at their configuration.

A Readiness Checklist

Score yourself honestly. Most businesses fail on the first four.

  1. Is pricing stated explicitly, at least as a band or model?
  2. Are specifications expressed as attributes rather than as prose or images?
  3. Is availability current and machine-readable?
  4. Are terms, returns, and limits stated plainly rather than in linked documents?
  5. Do product or service names match across your site, feeds, and third-party listings?
  6. Is your main content in raw HTML without requiring interaction or scripts?
  7. Is Organization and Product markup complete and accurate?
  8. Can AI agents reach your site at both robots and network layers?
  9. Does your team know which agent categories your infrastructure blocks?
  10. Do your feeds, site, and marketplace listings agree with each other?

Anything scoring poorly here is already costing you assistant visibility. That is the argument for acting now: the work required for agents is the work required for AI search generally, and understanding how search engines and AI Overviews work makes clear that the two share the same foundations.

The Risk of Being Unreadable

The downside is asymmetric, which is what makes a modest investment rational under genuine uncertainty.

If agent commerce grows slowly, you have accurate data, explicit pricing, clean structured markup, and better conventional visibility. If it grows quickly, you are legible to systems that will simply skip businesses they cannot parse, in a category where competitors who prepared are the only options presented.

Neither outcome punishes preparation. Both punish ambiguity.

Timing is the honest uncertainty. Nobody credible can say whether meaningful agent transaction volume arrives next year or in five, and the withdrawal of a flagship implementation after five months is evidence that confident predictions in either direction should be discounted. What can be said is that the preparation has independent value now, which removes the need to forecast at all.

Treat this as data hygiene with an option attached. A large language model optimization program should cover the readiness checklist as part of ordinary AI visibility work rather than as a separate agent project, and reviewing where your current data falls short is a reasonable first step when you get in touch.

Should I integrate with an agent commerce protocol now? 

Probably not, unless your platform provides it with minimal effort. OpenAI’s Instant Checkout, the flagship consumer surface for one of these protocols, was withdrawn in March 2026 after roughly five months. Prepare the underlying data instead, since every proposed standard consumes the same inputs.

Which agent protocol will win? 

The standards address different layers and are designed to compose rather than replace each other, with MCP handling data connectivity, ACP and UCP handling commerce flows, and AP2 handling payment authorization. Card networks have built protocol-agnostic on-ramps, which suggests merchants will work through their payment provider rather than choosing.

Do AI agents affect businesses that do not sell online? 

Yes, at the research and comparison stage. Agents evaluate options against criteria before any transaction, which means explicit pricing, clear service definitions, and machine-readable terms matter for service businesses too, even where the purchase itself still happens through a conversation.

Will blocking AI crawlers protect me from agents? 

It will also remove you from consideration. Infrastructure providers now separate search, agent, and training traffic, and a broad block frequently catches all three. Decide each category deliberately rather than inheriting a default, and verify what agents actually receive rather than what your configuration intends.

Is agentic commerce actually happening yet? 

Infrastructure is well ahead of consumer behavior. The protocols exist and have significant institutional backing, while the highest-profile consumer implementation was shut down after five months with minimal sales. The reasonable position is preparing the data, which has immediate returns, rather than building integrations that may not be used.

Sources