How to make your business agent-ready in 2026: a practical checklist

A business is agent-ready when AI assistants can discover it, quote it, and transact with it programmatically. Getting there is a ladder, not a leap. Here is the ladder, ordered by effort.

Step 1: Let the crawlers in (one hour)

AI assistants cannot recommend what they cannot read. Audit your robots.txt and allow the retrieval bots: OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot, and Google-Extended. The evidence favors opening up: a Rutgers/Wharton study (December 2025) found publishers that blocked AI crawlers lost 23.1% of total traffic without reducing how often they were cited.

While you are in there, confirm your site serves real server-rendered HTML. Most AI crawlers do not execute JavaScript, so a client-rendered product page is invisible to them.

Step 2: Publish structured data (one day)

Answer engines lift structure. The Princeton GEO study (KDD 2024) measured which content changes improve AI citation: adding statistics raised visibility up to 41%, quoting sources up to 115% for lower-ranked content.

• Add schema.org JSON-LD: Organization, Product or Service with price, FAQPage, and LocalBusiness with hours if you have locations
• Answer the obvious questions directly on the page in 40 to 60 words, then elaborate
• Add an llms.txt file: a concise, markdown summary of what you sell and how to reach you. Adoption is around 10% of domains and Lighthouse now audits it, so it doubles as a readiness signal

Step 3: Expose your catalog as tools (one week)

This is where readiness becomes transactable. Publish an MCP server with a handful of strict, well-described tools:

search_catalog   find products or services
get_quote        price a cart including tax and fees
check_slots      list bookable availability

Read-only tools first. You learn what agents ask for, with zero risk, before you let them buy anything.

Step 4: Turn on checkout with controls (one to two weeks)

The final rung is letting agents complete purchases and bookings:

• Serve ACP and UCP checkout endpoints from your existing catalog and processor
• Require a signed AP2-style mandate on every order, so each purchase carries proof the customer approved it
• Set policy before you set it live: order value caps, booking windows, rate limits per agent platform, and a refusal log you actually review

The control layer is what separates “open to agents” from “exposed to agents.” Every capability should be a rule you wrote down, not a default you inherited.

What results should you expect?

Set expectations honestly. Agentic orders are a small share of commerce today, roughly 1.5% of US retail e-commerce. But the growth curve is steep (Adobe measured 693% growth in AI-driven retail traffic in the 2025 holidays; Shopify saw AI-originated orders up 13x in Q1 2026), and early listings compound: assistants tend to keep recommending businesses that reliably complete transactions.

The build-or-buy question

Everything above is buildable in-house if you have platform engineers to spare and appetite to track four evolving specs. If you do not, this ladder is exactly what Errand ships as one integration: crawler-friendly structured surfaces, a hosted MCP server, protocol-compliant checkout, and the policy layer, connected to the store or booking system you already run.