/
Guides
What is agent readiness?
More customers now send an AI agent to research, compare and buy for them. Agent readiness is whether that agent can do business with you.
Ask an AI agent to find a hotel in Lisbon under €200 a night with a late checkout, and to book it. The agent searches the web or its available plugins, opens a handful of hotel and booking options, reads their rates and policies, picks one and tries to reserve a room. The person who asked may never see any of it. Whether your hotel makes the shortlist, and whether the booking goes through, depends on how your business looks to the agent.
Agent readiness is the name for that: how well a business works for customers who arrive through an AI agent. A ready business is one an agent can find, understand and complete a task with, whether the task is booking a room, comparing two CRMs or wiring a payments API into an app.
Until recently, an assistant answered from what it had learned in training and left the rest to you. But today's agents look things up before they answer. In an August 2026 study, GPT-5 and Claude Sonnet 4.6 ran a web search on more than 9 in 10 questions that needed current information. More and more often, they turn not to their training data but to the web, apps, plugins and checkouts built for them.
On recent questions that needed fresh information, all four models searched the web on at least 87.5% of them. Source: Kale, arXiv 2511.18931, August 2026.
This turns your website, docs and APIs into something a machine has to work through on your customer's behalf.
Why it matters
When an agent can't read your pricing or get through your booking flow, it doesn't report an error to anyone. It might struggle for a little while on your website, and then go take what it needs from somewhere else, such as a review site, a reseller or a competitor whose page loaded better information or whose MCP server, the interface agents use to call a service directly, worked better. The customer is satisfied, you never see the visit in your analytics, and the business goes to someone else.
Being visible to an agent is also different from being chosen by it. An agent can name your product in its research and still complete the task with another one, because the other one was the option it could use. Studies of coding agents already show well-known tools mentioned again and again and rarely installed (one showed that Netlify was mentioned 152 times but picked only 6 times. And the same gap applies to any task an agent has to finish, from a software signup to a restaurant booking.
Readiness also depends on the task, not just the site. The same business can be ready for "compare your plans" and not for "start a trial", or ready in one type of agent (a coding agent, a personal agent, a browsing agent…) and not in another, because each agent searches, reads and acts in its own way. That is why it can't be judged from a single page or a single tool.
Three layers: find, access, use
Agent readiness breaks down into the three things an agent has to do, in the order it does them.
Each stage depends on the one before, so a failure early on hides everything after it: an agent that never reaches your pricing page can't weigh your price. Each stage also tends to belong to a different team, which makes the split useful when deciding who fixes what.
An agent has to find a business, access what it publishes and use it, in that order.
Find. The agent has to discover you. Most agents look things up through a search engine, and not always the one you'd expect: Claude's web search, for instance, runs on Brave among other providers, ChatGPT on a mix of OpenAI's own index and Bing and Google results, and Gemini (naturally) on Google. Agents also follow links and read files that tell them where to look, such as a sitemap, an llms.txt file or the /.well-known addresses where a site can advertise its MCP server or API. This is the layer closest to SEO, and it often sits with marketing and content teams.
Access. The agent has to read what you publish. Most AI agents fetch a page's raw HTML with a plain HTTP request and never run its JavaScript, so a site that builds its content in the browser can look empty to them, and a bot filter on your CDN or firewall can answer with a 403 or a CAPTCHA before they read a word. This layer belongs to whoever runs the website and its infrastructure.
Use. The agent has to get the task done, and what that takes depends on the service. For an API it means signing in and calling the right endpoint, for a shop it means choosing options and checking out, and for a hotel it means holding a room. An entry point built for agents gives the agent a direct path: an MCP server exposes your actions as named tools with typed inputs the agent can call, an API with an OpenAPI description lets a coding agent write the integration itself, and an agent checkout lets it pay without filling in a form. Without one, the agent works through screens designed for people, where a login wall or a CAPTCHA can end the task.
How agent readiness relates to GEO
Generative engine optimization (GEO, also called AEO) asks whether AI assistants mention your brand when someone asks about your category. A brand can attempt to reach an assistant through its training data, which takes months or years to change, or through the pages it retrieves while answering. GEO works reliably only on the second, as our article on the subject explains.
Much of that work helps with the find and access layers, since an assistant can't cite a page it can't fetch, and an agent can't use one either. GEO works mainly through clear pages that answer the questions buyers ask, structured data, and mentions on the third-party sites assistants draw on, such as review sites, press and forums.
But whereas GEO optimizes for the mention, agent readiness carries on to whether the agent picks you and whether the task finishes with you. This is the part that turns visibility into revenue. Its tactics are different: being listed where agents look for tools, serving content agents can read, and giving them a way to act, such as an MCP server or a checkout built for agents.
How agents reach businesses beyond search
Not every agent that reaches your business is doing the same job. Some work in the background, crawling and indexing the web for search engines and assistants before anyone has asked a question. Others work at inference time: a customer asks for something, and the agent fetches your pages, calls your services and tries to finish the task on the spot. This second kind is where “use” and transactions tend to happen, and it is changing fastest.
These agents don't only search the web. Many work from a set of tools they can call, such as the MCP apps and connectors a person has available in ChatGPT, Claude, Grok or Meta's Muse. Each tool comes with a name, a plain-language description of what it does and a schema for its inputs. When a request arrives, the agent reads those descriptions, picks the tool that fits the task and calls it, and tends to search the web only when nothing fits.
Those apps come from the directory each assistant runs, and the same directory is where the agent looks for a tool the person hasn't added yet: ChatGPT, for example, can search its plugin directory mid-conversation and suggest one. Being listed there, with a description that matches what customers ask for, is how an agent finds you without a search engine.
Protocols are also appearing for specific kinds of interaction. No list stays complete for long, but most fall into a few families so far. Transaction protocols, such as OpenAI and Stripe's Agentic Commerce Protocol and Google's Agent Payments Protocol, let an agent check out and pay. Identity protocols such as Poppy, the Personal Agent Protocol from Sierra and Meta, let a company verify which personal agent is calling and on whose behalf, so the agent can sign in and act on a customer's account within limits the customer and the company have agreed. Accountability protocols, such as Pact from Decagon and Instinct, give both sides a signed record of deals an agent makes.
Each protocol of this kind gives agents a direct route into one type of service, and being ready increasingly means supporting the ones your customers' agents use.
Where to start
Start by taking stock of what your business exposes to agents today, then check it against the three layers.
Find: where do agents look for a business like yours (search engines, app directories, registries), and do you show up there?
Access: what does an agent receive when it fetches your pricing, product pages and documentation, and do your bot defenses let it through?
Use: for each task your customers hand to an agent, is there a path the agent can take, such as a CLI, an MCP server or a checkout built for agents?
A gap in one layer stops the task even when the other two are solid, so cover all three before polishing any one of them.
If this sounds like a lot at once, Alpic's agent readiness scanner does that review for your website. It runs 83 checks across the three layers, from whether your pages can be read without JavaScript to whether your MCP server and API are declared where agents look for them, and gives you a score out of 100 and a tier.
Try it out today for a first look at how agent-ready you are! And check back here for more details on what each test and each module mean for your agent readiness.
