Agentic Search: When AI Agents Browse, Compare and Book for Your Customers

From naming a business to booking it

For two years the worry was whether ChatGPT would mention you. The next worry is whether it can transact with you. An agent is software that takes a goal, works through the steps and returns with the job done or nearly done: the table held, the groomer booked, the lamp bought once it fell to the shopper's price. Google's AI optimization guide defines agents as autonomous systems that perform tasks on behalf of people, and its example is booking a reservation.

Agentic search is the finding half of that errand. The agent gets candidates where answer engines do, from search indexes, map data, feeds and partners. Then it must act, which needs a price it can trust, openings it can read and a booking path a machine can complete. A business easy to name and impossible to book gets mentioned, then skipped.

The GEO hub covers the first row across every engine, and query fan-out explains how Google researches a question. This page covers the other two rows, and the plumbing beneath them, which has collected acronyms at an alarming rate.

Errands the agents already run, as of October 2026

Most of this is live in the US, some behind paid plans or in Google's Labs, and nearly all of it runs through partners rather than your own website. Each item comes from the company's own announcement.

Note the pattern. The agent shortlists from what it can read and books through whoever has a connection. A restaurant on a reservation platform can be booked; one with a PDF menu is only named. Google's calls are the fallback for businesses that publish too little; the AI Mode page covers the opt-out.

Six protocols, sorted by the job each one does

A protocol is an agreement about how machines talk, so an agent needs no custom integration for every store and calendar. The names multiply faster than the substance. Each row below was checked on the owner's own site or documentation in October 2026.

AP2 answers the question owners ask first: what if the robot buys the wrong thing? It records the user's intent, then the exact cart approved, as signed 'mandates' both sides can point to in a dispute. Vendor decks also mention NLWeb, a Microsoft open project that turns a site's Schema.org data into a chat interface that doubles as an MCP server.

Who checks the agent's ID before it pays

Fraud systems were built to stop bots. Now some bots arrive carrying a customer's card. Card networks and the companies that sit in front of websites have built an identity layer for them, and it matters to any business with a firewall.

What to make machine-readable, trade by trade

Agents act on facts they can read and paths they can use. Google's guide sends owners back to Merchant Center and Business Profile, and says structured data is not required for its AI features and no special markup exists for them. The work is duller than the protocol table suggests.

The rule beneath every row comes from entity optimization: one set of facts, everywhere. An agent that finds a Saturday opening on your booking page and 'closed Saturdays' on your profile has no reason to choose you. Prices matter more now. An answer can say 'call for a quote.' An agent filtering by price cannot.

Retailers face the most machinery and have the least reason to build it: platforms and payment processors implement UCP and ACP, and a store owner usually switches support on. See GEO for ecommerce, ChatGPT shopping, restaurants and home services for trade detail.

A booking form a machine can finish

Browser agents read a site three ways at once, says the web.dev guide to agent-friendly websites, which Google's AI guide links: a screenshot, the HTML and the accessibility tree. Anyone who has paid for an accessibility audit will recognize the advice. A form a screen reader can complete is, broadly, a form an agent can complete.

• Use real buttons and links. A styled div with a click handler may never register as something to press.
• Tie every label to its field with the for attribute, so 'Pet weight' attaches to the right box.
• Remove transparent overlays over buttons; agents may discard elements they cannot see.
• Keep layouts stable, so the Book button sits in the same place on every page.
• Put prices, hours and openings in the HTML, not only in a widget a script builds after load.

That last item matters twice. The browser agent runs your scripts, but the crawlers that put you on its shortlist often do not: Vercel's December 2024 analysis found OpenAI's, Anthropic's and Perplexity's crawlers did not execute JavaScript. JavaScript and AI crawlers shows how to test. WebMCP can wait.

Proven, plausible and pitched

Agents are the newest thing to sell, so the pitch runs ahead of the evidence. Between the two lists sits the plausible middle, our practice rather than any engine's rule: published prices and live availability should make a business easier to shortlist, and accessible forms should let more agent bookings finish. Our AI SEO myths and llms.txt pages test more claims like these.

Where to start this month, and what to skip

1. 1

   Ask the agent's question

   Try 'book a dog groomer in Helotes Saturday morning under $90' in AI Mode, then run it through our free AI visibility check, which asks ChatGPT and others through their APIs (answers can differ a little from the apps). Are you shortlisted?

2. 2

   Publish the numbers

   Add price ranges, lead times and service areas to service pages as text, with the date you last checked them.

3. 3

   Make the profiles agree

   Hours, phone and services on Google Business Profile, Apple Business and Bing Places must match the site and booking system, holiday hours included.

4. 4

   Join the partner your trade uses

   A booking platform, a reservation partner or a Merchant Center feed. That connection is often the agent's only way to act. Then try your own booking form using only the keyboard.

5. 5

   Ask your platform first

   Before paying for protocol work, ask Shopify, your booking software or payment processor what they support.

Most small businesses should skip building their own MCP server. It suits a company that sells software or holds a live catalog or booking system worth exposing, which we scope separately as AI and LLM integration. Partner bookings appear in the partner's reports, not your analytics; tracking AI traffic in GA4 covers what GA4 can see.

At DataCram, agent readiness is part of the GEO work in our monthly SEO programs: profiles aligned, prices published, answers tracked monthly. Programs start at $625 a month and most local businesses land between $625 and $2,250, with no long contract. Every custom site includes Athena CMS, whose Hestia agent keeps hours identical everywhere. Scope is set after a free GEO audit; GEO pricing explains the range.