When the Buyer Isn't Human: Getting Chosen by AI Agents That Shortlist Vendors

Imagine a customer who never visits your website, never reads your reviews and never sees the homepage you paid for.

They tell an assistant: "find me a licensed electrician who can do a panel upgrade, has good reviews, and can come out this week." The assistant searches, compares, checks what it can verify, and hands back three names.

You are either one of those three names or you are invisible. There is no page two to be on.

What actually changed

The shift is not that people ask AI questions. That has been true for a while. The shift is that AI has started acting — running the comparison, checking the constraints, and returning a decision rather than a list of links.

The numbers behind that shift are worth stating precisely, because vague claims about "the AI revolution" are what make business owners tune out.

• An IBM study in 2026 found roughly 45% of consumers already use AI for at least part of the buying journey. Not all of it. Part of it — which is exactly the research and shortlisting stage.
• Adobe Analytics measured a 4,700% year-over-year increase in generative-AI traffic to US retail sites between July 2024 and July 2025.
• McKinsey estimates the agentic commerce opportunity at $3–5 trillion globally by 2030. Treat long-range forecasts with appropriate suspicion, but the direction is not in dispute.
• eMarketer projected AI platforms would account for roughly $20.9 billion in US retail spending in 2026, close to quadrupling the previous year.

Meanwhile the payment infrastructure quietly finished being built. Mastercard went live with agent transactions. Visa launched its Trusted Agent Protocol commercially after piloting with more than a hundred partners. American Express added purchase protection specifically for registered AI-agent purchases. All three major US card networks now support agent-initiated payments.

Two different games

Here is where most coverage of this topic goes wrong for the average business owner. Nearly all of it is written about e-commerce — product feeds, catalogues, checkout integration.

If you sell physical products online, that coverage is relevant and you should read it. But there is a second version of this, and almost nobody is writing about it.

A plumber will never transact through a checkout protocol. A law firm will not publish a product feed. But both get shortlisted — or skipped — dozens of times a week by systems evaluating them on data they have never audited.

Game two has no integration requirement and no vendor to hire. It rewards businesses whose information is simply correct and consistent everywhere. Which is why it is winnable right now, by almost anyone, with a week of unglamorous work.

The protocol layer, briefly

You do not need to implement any of this to benefit from understanding it. But knowing the plumbing explains why the advice further down works.

Three open standards now carry most agent-driven commerce.

Again — if you run a service business, none of that is your homework. It matters because it tells you how seriously the platforms are taking this, and because of what happened next.

The reversal nobody expected

For about eighteen months the dominant fear went like this: AI platforms will insert themselves between you and your customer, own the relationship, and reduce you to a supplier with no direct contact and no loyalty data.

It is a reasonable fear. It is roughly what happened with food delivery apps and hotel booking sites.

Then in March 2026, OpenAI deprecated Instant Checkout — the feature that let people complete purchases inside ChatGPT. The model shifted from "buy in the chat" to discover in the chat, transact on the merchant's own site.

So the practical question stops being "how do I avoid being disintermediated" and becomes something much more actionable: when an agent assembles a shortlist in my category, am I on it?

How an agent judges you differently than a person does

This is the part worth sitting with, because the instincts you have built over years of marketing to humans actively mislead you here.

Read the right-hand column again, because every item is a strategy implication.

"Cannot see your design at all." The thing you spent the most money on is invisible to this evaluator. Not de-prioritised — invisible.

"Discounts promotional language." Superlatives are the single least useful thing you can write for an agent. "The best plumber in Denver" carries no information because every competitor says it. "Licensed, bonded, insured, serving Arvada and Wheat Ridge, emergency calls answered within two hours" is all signal.

"Skips you rather than asking." This is the harsh one. A human with an unanswered question calls you. An agent with an unanswered question moves to the next candidate, and you never learn it happened. There is no bounce rate for being left off a shortlist.

Why this quietly favours smaller businesses

Here is the genuinely encouraging part, and it is not a motivational flourish — it follows directly from the mechanics above.

Consider what large incumbents actually bought with their advantage: television awareness, decades of name recognition, sponsorships, a brand people trust on sight. Every one of those is worth precisely nothing to an agent.

Now consider what agents reward: accurate data, consistency across sources, substantive reviews, clear availability, specific service descriptions. A ten-person company can achieve all of that completely. In fact it is easier at ten people than at ten thousand, because the multi-location enterprise has franchise listings, legacy directory entries and regional variations quietly disagreeing with each other.

We put a version of this argument on our generative engine optimisation page, and it is the reason we push clients toward data hygiene before campaigns. Fixing what a machine reads about you is cheaper than buying attention, and right now it is also rarer.

The six signals that decide it

Across the categories we work in, the same six things determine whether a business gets named. None of them require a developer, and four of them are free.

If you want the deeper mechanics on the second and third of those, we go much further in the guide to entity signals — that is the foundation everything here sits on.

What disqualifies you

Being skipped is usually not a close call. It is normally one of these, and the first one accounts for more exclusions than the rest combined.

• Conflicting information. Three phone numbers, two sets of hours, an address that changed in 2023 but lives on in four directories. The agent cannot tell which source is authoritative, so it favours a competitor whose data agrees with itself.
• Blocked crawlers. A robots.txt rule, often inherited from a template or added by a plugin, that excludes GPTBot or Google-Extended. Total exclusion, invisible cause.
• No verifiable specifics. A site that describes a "commitment to excellence" and "customer-focused solutions" without ever stating what you do, where, for whom, or at what price. There is nothing there to quote.
• Stale or thin reviews. Four ratings, none recent, no responses. An agent reads that as a business with no current evidence of operating well.
• Unanswerable availability. No hours, no service area, no indication of whether you take new clients. An agent handling "who can come out this week" cannot include a candidate it cannot verify is available.
• A slow or broken site. Timeouts and errors during retrieval get treated as unreliability. The agent does not retry patiently on your behalf.

Notice how mundane that list is. There is no clever technique missing. It is six varieties of "your information is not in order," and every one is fixable without a marketing budget.

Test it this week

You do not need a tool or an agency for this. Forty-five minutes and a notepad will tell you most of what you need to know.

We turned a longer version of this into a standalone walkthrough: the 15-minute AI visibility self-audit, with the exact prompts and a scoring sheet.

What to measure

There is no dashboard for this yet. Anyone selling you one is selling an estimate. Here is what can honestly be tracked today.

That last row is the most underused diagnostic in this entire article. Your server logs record every AI crawler that visited. If the count is zero, no amount of content strategy will help until access is fixed.

The uncomfortable summary

A category of buyer now exists that cannot see your branding, does not know your name, ignores your headlines, and will not call to clarify. It evaluates you on whether your facts are correct, consistent, specific and machine-readable — and it moves on silently when they are not.

That is genuinely bad news for businesses whose advantage was built on recognition. It is unusually good news for businesses willing to spend a week making their information accurate everywhere.

Most competitors have not noticed yet. That is the entire opportunity, and it has an expiry date.

If you are unsure whether this is even your bottleneck — plenty of businesses have a pricing or capacity problem wearing a visibility costume — our free answer desk walks through the diagnostic without asking for anything first. And if the honest answer turns out to be that agentic visibility is not your problem right now, it will tell you that too.

Questions we get about this