---
title: "When an AI Agent Shops for Your Customer, What Makes It Pick You"
seoTitle: "When an AI Agent Shops, What Makes It Pick You"
description: "AI agents are starting to book, buy, and shortlist for customers. What makes an agent choose your business over the one down the road, in concrete terms."
datePublished: "2026-07-18T06:08:00Z"
dateModified: "2026-07-18T06:08:00Z"
category: ai
imageAlt: "Iron Goo blog featured image on what makes an AI agent acting for a customer pick one small business over a competitor."
tags: [ai-agents, agentic-commerce, smb-ai, ai-search, ux]
faq: true
---
A man on a job site, hands covered in grout, says to his phone: "find me a same-day locksmith near the warehouse, somebody who does commercial deadbolts, and get one booked for this afternoon." He goes back to work. He is not searching anything. He has handed the whole errand to an AI agent, and a few minutes later it tells him a name and a time. Whether that "agent picks you" or the locksmith two streets over is not one decision the software made; it is a short chain of them, and most owners only ever picture the last link. The agent first had to find you and have enough consistent, machine-readable information about you to even put you on the list. Then it had to weigh you against the others it found and judge you the better match for "commercial deadbolts, this afternoon." Then, the part nobody plans for, it had to be able to actually book you on your own site without a person there to nudge it. Miss any link and the man on the job site never hears your name.
This is still early. Agents are not doing most of the shopping yet, and anyone who tells you they are is selling something. But the behavior is real and it is spreading, and it is not one product. The agent might be powered by Claude, it might be an assistant built into a major platform, it might be one of several others, and from your side of the glass the distinction barely matters: a person delegates a concrete errand to an AI agent, the agent goes and acts on it, and somewhere in that run a business gets chosen and the rest do not. The useful question for an owner is not whether this is the whole future. It is narrower and more answerable. When an agent is sent to handle a job a customer would have handled themselves, what makes it land on your business, and what could make you lose a sale you had already won.
## What makes an AI agent pick your business?
The agent has to find and consider you through consistent, machine-readable information, judge you a better match for the customer's actual need than the alternatives, and then complete the task on your surface. An agent that cannot finish the job abandons you for one that can.
That is the chain in one breath. The rest of this is each link made concrete, because owners tend to collapse all three into a single popularity contest and then optimize the wrong one.
## The pick is a chain, not a popularity contest
Picture how an owner imagines getting chosen by AI. They imagine something like a vote: the AI "knows" the good businesses, and being good enough gets you picked. It feels like one event. Be reputable, be liked, get selected.
That is not what happens when an agent is acting for a real person on a real errand. What happens is closer to how a careful assistant would handle being told "book me a locksmith." First they have to come up with candidates at all, which means you have to exist in a form they can find and read. Then they have to choose among the candidates, which means you have to look like the right one for this specific job. Then they have to actually carry out the booking, which means your site has to let them finish it. Three different competencies. You can be brilliant at one and fail the customer at another, and the customer never sees which link broke. They just get a different business.
Naming the links is the whole point, so here they are plainly:
- **Found and considered.** The agent assembles a shortlist from what it can read about businesses like yours. If your information is missing, thin, or contradicts itself across the places it looks, you are not on the list. You cannot be weighed if you were never a candidate.
- **Weighed and chosen.** Among the candidates, the agent judges which best fits the customer's stated need. Specificity wins here. A business clearly described as doing the exact thing asked beats a vaguer one that might.
- **Operated to completion.** The agent goes to your surface and tries to do the task. If it can finish, you got the job. If it hits a wall it cannot get past, it leaves and picks someone it can finish with, and the pick you won evaporates.
Owners spend almost all their worry on the first link and almost none on the third. The third is where won sales quietly die.
## Getting found and getting chosen
The first two links live on the search side of the world, and there is real depth to them that other writing here owns. I will stay at framing depth and point you to it, because turning this into a full guide to AI visibility would bury the part owners actually miss.
To be **found and considered**, an agent has to be able to read who you are and what you do, and it has to find the same story wherever it looks. Agents do not see your business the way a customer scanning your homepage does. They assemble a picture from machine-readable information scattered across your site, your listings, your profiles, the places that mention you. When those sources agree, you resolve into a confident, considerable business. When they disagree, when one place has an old service you dropped and another has the new one, when your hours or your location or even your name render three slightly different ways, the agent gets a blurry, low-confidence read and reaches for a candidate it is surer about. Corroboration across sources is what makes you a candidate at all. The mechanism behind it, how AI assembles and trusts a picture of a business from agreeing sources, is laid out in [how AI platforms decide which business to recommend](/blog/ai-recommends); the practical first move, seeing what AI currently believes about you, is [a short audit of what AI already knows about your business](/blog/does-ai-know-you).
To be **weighed and chosen** once you are a candidate, you have to match the customer's actual need more precisely than the alternatives. The man asked for commercial deadbolts, same day. An agent comparing a locksmith described plainly as a commercial-and-emergency service with same-day availability against one described only as "locksmith services" has an easy call. The specific one is the safer match for the specific request. This is not about more words. It is about the right ones: the exact service, the real coverage area, the actual availability, stated where an agent can read them and consistent with everywhere else you appear. Being a confirmable, consistent entity rather than a vague one is what wins the weigh step, and that confirmability is itself a trust signal an agent leans on, covered in [the signals that make a business confirmable to AI](/blog/trust-signals).
Both of these are necessary. Neither is sufficient. You can be perfectly found and perfectly matched and still lose the job at the next link, which is the one this post exists to put in front of you.
## The link owners ignore: can the agent actually finish?
Here is the uncomfortable part. Say you win the first two links. The agent found you, considered you, weighed you against the locksmith two streets over, and chose you. It is now on your booking page, on the customer's behalf, trying to complete the task. The customer is back at work, expecting an appointment.
And the agent cannot finish.
Not because anything is broken. Your booking page works fine; people use it every day. It works for people because people have eyes and improvise. They see a control that looks like a date field and click it even though nothing names it a date field. They recognize a button by how it looks even though the markup never says it is a button. They watch a confirmation flash green and trust the job is done. An agent has none of that. It is driving without eyes, reading your surface through its structure, its text, the names and states the page actually exposes. Where your page communicates only by appearance, the agent has nothing to act on. It reaches the spot where a person would have squinted and guessed, and it stops, because it cannot squint.
When that happens, the agent does not wait. It does not call the customer. It abandons the task and the business attached to it, goes back to its shortlist, and tries the next candidate: the locksmith two streets over, whose booking page it can read end to end. That competitor did not necessarily get found better or matched better. They just built a surface an agent could operate to the finish. The sale moved on the strength of the one link you never thought to check.
::::comparison{title="Two businesses an agent chose"}
:::side{label="Chosen, then abandoned"}
The agent picked this locksmith as the best match and went to book. It found the service, started the form, and reached a time-slot control that the page renders as styled boxes with no readable names and no machine-readable mark on which slots are open. There is nothing under the appearance for the agent to act on. It cannot tell which slot is "this afternoon, available," so it cannot select one. The booking dies one step from done. The customer hears about a different locksmith. This business never knows it was first choice.
:::
:::side{label="Chosen, and completable"}
The agent picked this locksmith too, on a different run, and went to book. Each step of the flow exists as something the agent can read and act on: the service is named, the time slots carry real labels and an observable open or taken state, the confirm control is an actual button, and a readable message confirms the appointment. The agent works through it and finishes. The customer gets a time. Nothing here was fancier. It was simply built so the task, not just the look of the page, could be completed by an actor without eyes.
:::
::::
The difference between those two businesses is not reputation, not marketing, not how well either gets described to AI. Both got chosen. One had a surface an agent could finish on and one did not, and that alone decided who got paid.
:::callout{type="key" title="Being chosen and being operable are two different wins"}
An owner can do everything right on visibility, win the agent's pick outright, and still lose the job at a checkout or booking flow the agent cannot operate. Getting picked gets the agent to your door. Whether it can finish the task once inside is a separate question, and it is the one most owners have never asked.
:::
## Why this is the hinge, and where it hands off
The reason the completion link matters so much is that it inverts a comfortable assumption. Owners assume the contest ends at the choice, that once you are good enough to be picked, the work is done. With a human customer that is roughly true; a person who chose you will usually push through a clumsy checkout to finish, because they want the result and they can cope. An agent will not cope. For an agent, being chosen and being able to finish are the same decision in two parts, and failing the second silently undoes the first. A perfectly discoverable, perfectly matched business loses the job at a step a person would have muscled through.
So preparing to be picked by agents is two jobs, not one. The first is everything that makes you findable, confirmable, and the clear match, which is the search-and-entity work the neighbors above go deep on. The second is making sure that once an agent arrives with a task in hand, your surface lets it finish. That second job, [whether an agent can operate your booking or checkout page once it gets there](/guides/ux/designing-for-ai-agents-as-users), is its own discipline: how to build a surface whose tasks an agent can actually carry through to completion, rather than one that only looks done to a person watching it animate.
I am deliberately not teaching that here. How to make a page an agent can finish on, what real structure versus visual-only looks like, how a task becomes a path an agent can follow, the one test that tells you whether an agent could complete a job on your site, all of that belongs to the guide and is covered there in full. This post has a narrower job: to put the whole chain in front of you and make the last link visible, because it is the one that decides the pick when nobody is watching. Naming it is the point. Fixing it is the guide.
There is also the honest question of who does this work. Some owners will read their own surface against this and find it fine. Many will find that the pages a customer completes by reflex are exactly the ones an agent stalls on, and that getting them operable is real work. [The work of preparing a business's surface for agents acting on a customer's behalf](/services/aio) is a service like any other; the point of naming the chain is that you can now tell the difference between someone selling you visibility and someone who can also make sure the job actually completes once the agent shows up.
## Where to start
Do not start by trying to fix everything. Start by walking your own most important task, the booking, the quote, the reorder, the way an agent would have to: through the structure and the readable labels and the states the page exposes, not through the appearance you already know how to read. The first place you find yourself relying on "well, a person would just know to click that" is the first place an agent picks you and then loses you. Walk that one flow against the agent-legibility guide below before the agents start arriving in earnest. That single pass tells you whether the businesses you are about to lose to are doing anything you cannot.
