How agents experience your business: a manifesto for business leaders

1. The short version

Customers are starting to hand tasks to software. They ask an assistant to find a hotel for two nights, to compare three project management tools, to reorder printer ink, or to book a table for Friday. The assistant does not stop at an answer. It sends an agent to do the job: to visit sites, read what is there, fill in forms and, in some cases, pay.

For a business, this creates two separate questions. The first is whether the assistant mentions you at all. The second is what happens when the customer's agent arrives on your site. Most of the attention, and most of the tools sold under the AI banner, focus on the first question. This manifesto is about the second, because that is where a visit turns into a booking, a sign-up or a sale, or does not.

We call the second question agent experience: how agents experience your business once they arrive. It covers five plain areas that run through this whole site: can agents find you, read you, act on your site, pay, and can they be trusted, and trust you. None of these needs you to become an engineer. Each comes down to a decision someone in your business makes, or has already made without noticing.

Our position is simple. Being surfaced in an AI answer is useful. Being converted, meaning the task is finished, is the point. You can do well on the first and still lose the second, and the second is the one you can check, fix and check again.

2. What is a customer's agent, and what does it do on your site?

An AI agent is software that carries out a task for a person on the web. It is different from the crawlers that collect pages to train a model, and different from the bots that build a search index. A customer's agent works for one person, on one task, right now. Our page on what agents do sets out the kinds of AI visitor side by side.

It helps to follow one visit from start to finish. Picture a customer who asks their assistant to find and book a quiet hotel room near a conference venue, under a set budget, with free cancellation.

First, the agent looks you up. It may start from an AI answer, a search result or a list it already knows. To reach you at all, it has to be allowed in by your rules for automated visitors and has to find the pages that matter: the rooms, the prices, the location, the policies.

Next, it reads your prices and policies. It needs the nightly rate, the taxes and fees, what is included, and the cancellation terms, as text it can read. If the rate only appears after scripts run in a browser, or the cancellation policy is a PDF linked from the footer, the agent may come back with gaps, or with nothing.

Then it tries to book. It chooses dates, a room type and a guest count, and works through your booking form. If a step only makes sense to a person looking at the screen, or a challenge appears that only a person can pass, the task stops there.

Maybe it pays. Some agents hand the final step back to the customer. Others can pay within limits the customer has set, using newer payment approaches built for agents. Either way, your checkout has to let the task reach the point where payment can happen.

Finally, it has to be trusted. Your fraud and bot rules decide whether this visitor looks like a real agent acting for a real customer or like an unknown bot. And the customer has to trust what the agent reports back about you, which depends on your policies being public, current and clear.

Every one of those steps is a place where the visit can end. The customer will usually never tell you. The agent simply reports that it could not complete the booking here and offers another option.

The same pattern holds outside hotels. A customer who asks for the best plan for a team of five sends an agent to read pricing pages and sign up for a trial. A customer who wants a replacement part sends an agent to find the right model, check stock and delivery, and place the order. A customer who wants a table on Friday sends an agent to a booking page. The details differ. The five steps do not.

3. How AI agents read a website

This site groups agent readiness into five areas. Each one has a few standards behind it. You do not need to know how they work, only what each one is for and who in your team owns it. Here is each area in plain words, with the standard behind it named once.

Find. Can agents find you? The oldest tool here is robots.txt, a public file on your site that tells automated visitors which parts they may visit. Alongside it, your bot protection, firewall or content delivery network may have its own rules for AI bots. A newer proposal, llms.txt, is a plain text file that points AI tools to your most useful pages. Our briefing on which AI bots to let in covers the choices.

Read. Can agents read you? Agents read text. If your main content, prices and policies arrive in the first response from your server, an agent can read them. If they only appear after scripts run, or sit inside images, PDFs or pop-ups, many agents will miss them. Some services can also give agents a markdown version of a page, a simpler text format, when an agent asks for one. Our lesson on making pages readable for agents explains what to look for.

Act. Can agents act on your site? This is where tasks happen: search, booking, sign-up and checkout. MCP, the Model Context Protocol, is a standard way for a business to offer agents a set of actions they can call directly. WebMCP is a proposal that lets a web page describe the actions it offers to an agent in the browser. Without either, an agent has to work your forms the way a person would, and forms built only for people are where many tasks stop. Our lesson on forms, sign-in and consent goes further.

Pay. Can agents pay? Payment approaches built for agents are new. x402 is one example: a way for a site to ask for payment and for an agent to pay as part of the same request. Others exist, and this site does not review them all. For most businesses today, the practical questions are whether the full price is readable before the last step and what your fraud tools do with a purchase that an agent started. Our lesson on agent payments and fraud rules sets out the decisions.

Trust. Can agents be trusted, and trust you? Web Bot Auth is a way for an agent to sign its requests so your site can check who operates it, rather than relying on a name the bot claims. Trust also runs the other way: your returns, cancellation, privacy and terms pages need to be public and readable so the agent, and the customer behind it, can rely on what you say. Our lesson on verifying agents and bots covers both sides.

These five areas are the structure of the self-check: 18 yes, no or not sure questions, scored in your browser, that show where you stand without anyone scanning your site.

You will notice that most of these standards are about making things explicit. Robots.txt makes your rules for visitors explicit. Llms.txt makes your most useful pages explicit. MCP and WebMCP make the actions you offer explicit. Web Bot Auth makes the identity of an agent explicit. Agents do best where a business has written down, in a form software can read, what it already knows about itself. Our glossary explains each term in a sentence or two if you need a reminder later.

4. Why leaders confuse agent experience with AEO

AEO, answer engine optimization, is the practice of tracking and improving how a brand appears in AI answers. It grew out of search marketing, and it answers a real question: when a customer asks an assistant about my category, am I named, and which sources does the answer cite?

Agent experience answers a different question: when the customer's agent arrives on my site, can it read, get in and finish the task? Both sound like AI visibility. Both are sold to the same marketing and digital teams. Both show up in the same meetings, often on the same slide.

The words do not help. AEO tools now use the word agent too. Scrunch, for example, calls its product an Agent Experience Platform, and its Agent Pages feature detects AI visits and serves AI-optimized versions of pages, with the aim of helping a brand appear more often in AI answers. That is a sensible thing to want. It is still mostly about being surfaced, not about whether an agent can complete a booking or a checkout.

So when a vendor, an agency or a board member says agent, it is worth asking one follow-up question: do you mean whether AI answers mention us, or whether an agent can finish a task on our site? The answer tells you which kind of tool, budget and owner you are talking about. Our page on AEO vs agent experience sets the two side by side.

None of this makes AEO the wrong investment. It makes it a partial one. If your plan for AI agents is only an AEO plan, you are paying to be shortlisted without checking that you can be chosen. If your plan is only an agent experience plan, you may be easy to buy from but rarely mentioned. Most businesses will want to know where they stand on both, measured separately, with owners who talk to each other.

5. Being mentioned is not the same as being chosen and completed

Think of the journey in two halves. In the first half, the customer asks a question and the assistant answers. Your brand is either in that answer or it is not. If it is, you have been surfaced.

In the second half, the customer says go ahead, and an agent visits the businesses in the answer to do the work. It reads, compares, chooses and tries to finish. If it completes the task with you, you have been converted: the booking is made, the sign-up done, the order placed.

These halves are measured in different places. Surfacing happens inside the AI answer, on someone else's service, and you see it only through tools that ask the same questions again and again. Conversion happens on your own site and checkout, where you control the pages, the forms, the rules and the policies.

They also fail in different ways. A brand that is not surfaced is missing from the shortlist. A brand that is surfaced but not converted is on the shortlist and then drops out, often at the last step. The second failure is easy to miss because the first half looks fine. Your AEO dashboard shows you in the answers. Your sales do not show the agents that arrived and could not finish.

This is why we keep the two words apart. Surfaced is the find step. Converted is find, read, act, pay and trust, all the way to a finished task.

There is a practical consequence for how you report on AI. If your monthly AI report only shows mentions and citations, it is a surfacing report. Add a second line that asks whether a customer's agent could finish your main task this month, and what changed. The two lines will not always move together, and that is the point of keeping them apart.

6. Four stories of a business that was surfaced but not converted

The four stories below are illustrations, not case studies. They are made-up examples of common patterns, written to show where a visit can end. They describe no real company and report no real results.

Example one: the hotel whose booking form needs a person. A small hotel is named in an AI answer about quiet places to stay near a city venue. The customer's agent visits, finds the rooms and reads the rates. The booking form, though, asks the guest to pick dates from a custom calendar that only works with a mouse, then shows a challenge to prove the visitor is human. The agent cannot pass either step. It reports that the room looked right but could not be booked, and offers a hotel down the road whose booking works end to end. The first hotel did well on being surfaced. It lost on act.

Example two: the SaaS company whose pricing loads only after scripts run. A software company appears in an AI answer comparing tools for small teams. The customer's agent visits the pricing page to check plans and limits. The page loads its plans with scripts after the first response, so what the agent reads is a heading, a few buttons and no prices. It cannot compare the plans, so it tells the customer the price was not available and moves on to tools with readable pricing. The company was surfaced. It lost on read.

Example three: the store whose checkout needs a human step. An online store is cited in an AI answer about a particular product. The agent finds the product, reads the price and adds it to the basket. At checkout, a final step asks the shopper to confirm through a pop-up that only a person can complete. The agent stops and hands the task back to the customer, who is busy and does not come back to it. The store was surfaced and nearly converted. It lost on pay.

Example four: the brand whose bot rules turn agents away. A well-known brand appears in many AI answers. Its bot protection, switched on years ago, treats every automated visitor as a threat and blocks them before they see a page. Customers' agents arrive and are turned away. Nobody in marketing knows, because the rule sits with the security team and nobody connected it to AI agents. The brand was surfaced again and again. It lost on find and trust, at the door.

In each example, the business could have looked at its AEO results and felt fine. In each, the gap was on its own site, under its own control.

Notice also that in none of the four stories did the fix require a new marketing campaign. The hotel needed a booking form an agent could complete. The software company needed prices in the first response. The store needed a checkout step that did not depend on a person. The brand needed bot rules that matched a decision someone had actually made. These are changes to how the business runs online, decided by business owners and carried out by their web teams.

7. Why AEO results are volatile and agent experience is checkable

AI answers are generated each time someone asks. The same question can come back with different wording, a different order and different sources. When the models behind an assistant are updated, the answers can shift again. That is not a flaw in AEO tools. It is the nature of what they measure.

To cope with it, AEO tools re-run the same prompts on a schedule and report trends rather than single results. Scrunch's Core plan, for example, lists 125 prompts and 5,000 responses a month (Scrunch pricing, read 2026-10-18). The result is useful for spotting direction. It is hard to use as a pass or fail.

Agent experience is different. The same check on the same page gives a comparable result until the page changes. Either your robots.txt allows a named agent or it does not. Either the price is in the first response or it is not. Either a guest can complete the booking without a challenge or they cannot. You can check, change something, and check again, and the difference is down to what you changed.

That makes agent experience something you can manage like any other part of your site. You can give it an owner, set a baseline, fix the gaps in order of importance and confirm each fix. The readiness calculator shows the fewest answer changes that move you up a level, and our two-week trial briefing sets out a short, repeatable routine.

It also makes progress easy to explain to a board or a leadership team. Instead of a share of AI answers that rises and falls from week to week, you can report a short list: the bots we allow, the pages that are readable, the tasks a guest can finish, our position on agent payments, and how we verify agents. Each item is either done or not done, and each one has an owner.

Repeatable does not mean permanent. A site redesign, a new checkout provider or a change to your bot protection can undo good work overnight. That is why the commitments at the end of this manifesto include a named owner and a schedule. The check is easy to repeat, so repeat it whenever something important changes.

8. The business decisions behind readiness

Most readiness gaps are business decisions as much as code. The code matters, but someone has to decide what the code should do, and in most businesses that someone sits in marketing, ecommerce, digital, legal or finance rather than in engineering.

A useful way to start is to list the three tasks that matter most to your business: the booking, the trial sign-up, the order, the quote request. For each one, ask who decides whether an agent should be able to do it, who owns the pages involved, and who owns the rules that might stop it. Often the answers are three different people who have never discussed it together.

Which bots you let in. AI visitors are not all the same. Some collect pages to train models, some build search indexes, some fetch a page to answer a live question, and some act for a customer. You can allow some and refuse others. The decision belongs to whoever owns your content and your customer relationships, not only to whoever set up the firewall.

Which tasks you open to agents. You may be happy for an agent to check prices and availability but not to cancel an order, or to book a table but not to change a reservation. Writing that list down is the first step. Deciding which tasks to offer through a direct route, such as an MCP server or WebMCP tools, comes after.

How you handle payment. Some businesses will want the customer to confirm every payment. Others will accept agent payments within limits. Your fraud rules need a position on purchases that an agent started, and someone has to sign off on it.

How you handle trust. Decide whether you will tell verified agents apart from unknown bots, and treat them differently. Make sure your policies are public text. Decide who owns your policy on AI agents and where it is written down.

None of these decisions needs a developer to make. All of them need a developer to carry out. That is why this site offers questions to ask your web team rather than implementation guides.

9. Who does what

Tools in this area do different jobs, and their names do not always tell you which. Here is how the tools named on this site group, in plain words. Each entry, with its sources, is on tools that check.

• Agent experience: ora.ai checks discovery, access, usability and payments, and its checks fit SaaS and brand sites. Its Journey tool records an agent attempting a task on a public page.
• The closest comparison: Scrunch tracks how a brand appears in AI answers and, on its Enterprise plan, serves AI-optimized pages to agents.
• Bot and agent security (DataDome): decides which agents get in and blocks fraud.
• Publisher tools: TollBit is built mainly for publishers to control and charge for AI access, and its site also lists brand and commerce offerings. Axioma works on the publisher side with access rules and monetization.
• Developer tools: 514.ax runs coding agents to see whether they find, choose and use a developer product.
• AEO platforms: for example Profound, which reports how a brand shows up in AI answers and which AI crawlers visit.

Two more names sit apart. Vercel is a partner of ora.ai, and is-agentic.com was built by Vercel with ora.ai, with every scan run by ora.ai. They are listed under Partners and built with Ora on tools that check.

ora.ai is an agent experience platform, not an AEO tool.

10. Ten commitments for an agent-ready business

If you take one thing from this manifesto into a planning meeting, take this list.

1. We will decide, in writing, which AI bots and agents we let in and why, and make sure our robots.txt and bot protection match that decision.
2. We will make sure the pages that matter most, the products, plans, rooms, prices and policies, are in our sitemap and easy for an agent to reach.
3. We will show prices, including tax and delivery, as text in the first response, not only after scripts run.
4. We will publish returns, cancellation, delivery, privacy and terms as public text that needs no login and no pop-up.
5. We will check that a guest can complete our main task end to end, and remove challenges that are not needed.
6. We will choose which tasks we open to agents and which we keep for people, and say so.
7. We will agree a position on purchases started by agents, and make sure our fraud and payment rules reflect it.
8. We will learn to tell verified agents from bots that only claim a name, and treat them accordingly.
9. We will give agent experience one named owner and review it on a set schedule.
10. We will keep measuring surfaced and converted separately, and not mistake one for the other.

11. Sources

Sources: RFC 9309: Robots Exclusion Protocol · llms.txt proposal · Model Context Protocol specification · WebMCP proposal · x402 · Cloudflare Web Bot Auth · Cloudflare, "The agentic internet" · Scrunch Agent Pages (read 2026-10-18) · Scrunch pricing (read 2026-10-18) · DataDome Agent Trust Management (read 2026-10-19) · ora.ai · TollBit (read 2026-10-18) · Axioma (read 2026-10-18) · 514.ax (read 2026-10-18) · Profound · Undated sources read September 2026