How AI agents find a business

YourCite · AI visibility · September 2026 · sources checked against each vendor's own announcement

In September 2026, three of the largest AI companies shipped agents that read websites by themselves. Meta released Muse on 8 September, OpenAI released Dots at DevDay on 29 September, and xAI runs Grok Bot as a team of always-on agents. Each one gets its own browser or its own cloud computer, and each one goes and looks at your site the way a person would.

That changes a question owners have been able to ignore. It is no longer only whether a search engine can index you. It is whether a machine that is researching on someone's behalf right now, can read you, understand what you sell, and trust you enough to put you on a shortlist.

This page is what we know from the vendors' own documentation about what those agents can and cannot read, and what to do about it. It is early. We say where it is early rather than dressing it up.

What actually launched?

Three products, all within September 2026, all with the same shape: an agent gets a computer of its own and uses it the way a person would.

But this is not three launches. It is an industry moving at once, and the signals are easier to read outside the model vendors. Cloudflare shipped an Agent Readiness score at isitagentready.com that scores a site on discoverability, content, bot access and capabilities. Vercel shipped Is Agentic, a free tool that audits a site against 118 agent-readiness checks. Shopify opened its checkout to browser-based AI agents, and the Wall Street Journal reported Shopify using Meta's Muse for agentic checkout. Cloudflare and Coinbase launched a payments standard so an agent can pay for access itself. The UN turned to Google to make its data ready for agents. When the plumbing, the scorecards and the payment rails all land in the same quarter, that is a direction of travel, not a product cycle.

How is an agent different from a search engine?

A search engine crawls your site, indexes the text, and ranks the page. An agent does something more direct: it opens your site and tries to accomplish a task. That difference decides what matters.

For a crawler, the question is whether your words are in the HTML. For an agent, the question is whether it can operate your site: find the price, find the delivery terms, find the contact route, and finish what it came to do. A page can rank perfectly and still defeat an agent.

Google's own AI optimization guide makes the priority clear by placement: agentic experiences are its last section, written as something to do if you have extra time. We agree with that ordering, which is why this page is a guide and not a promise. The work is real; the volume of it is not here yet.

Should you build an app for these agents?

No. Building an app is the instinct when a new platform appears, and it is the wrong one here, for a reason that is specific to how these agents work: they do not arrive through an app store. Each of them gets a browser or a computer of its own and reads your existing website the way a person would. Meta gave Muse a browser inside a secure virtual machine. OpenAI gave Dots its own cloud computer and browser. Grok Bot gives every agent a persistent computer.

So there is nothing to publish to. There is only your site, and whether a machine can read it. That is the whole surface.

The app path also carries a cost most owners do not see: an app is maintenance forever, tied to one vendor's ecosystem, and it does nothing for the search engines and AI answers that send you traffic today. Making your site readable does both jobs with one effort, because the same fixes that let an agent operate your page are the ones that let a search engine index it and an AI answer quote it.

What if you sell online?

This is where the direction of travel is clearest, and where we want to be careful not to overclaim.

The plumbing is real and it is being built by the companies that own checkout. ucp.dev, the Universal Commerce Protocol, is co-authored by Google, Shopify, Amazon, Microsoft, Meta, Stripe, Etsy, Target and Walmart, plus lodging and food companies. Stripe and Shopify both ship agent checkout work. When the checkout layer standardises, the retailer's problem stops being payments and becomes discovery: an agent has to find your product, understand it, and trust it before any checkout code matters.

What we will not say is that agents are shopping your store today. We measured the search demand for it, and it is genuinely thin. Of eight commerce-side queries we checked, page one was already held by seven to nine large domains in every single case: McKinsey, IBM, Mastercard, Deloitte, Shopify, BigCommerce, TechCrunch. The subject is covered by consultancies and the platforms themselves. That is a crowded page, and it tells you this is being talked about long before it is being bought.

The honest position: the infrastructure is arriving, the buying habit is not here yet, and the useful work now is making your product data readable rather than betting a quarter on agent traffic.

What can an agent actually see on your site?

This is the part most owners have never had to think about, and there is now a controlled number for it. Google's web.dev guide Build agent-friendly websites states that agents view a site through three channels at once: what they see on screen, the raw HTML, and the accessibility tree.

The accessibility tree is the one that decides the outcome. It is the same structure screen readers use, and most sites have never checked it. AudioEye ran a study in September 2026 with 1,560 agents built from six commercial models, across 13 real tasks on six websites, and the only variable was whether accessibility fixes loaded. On the site with the most issues, agents completed 31 percent of tasks. On an accessible version of the same site they completed 96 percent. Every model failed, including the most advanced ones.

The example from that study is worth reading twice. One task needed numbers that appeared only inside an image with no text description in the code. Six models tried it ten times each. All sixty attempts failed. Nothing was broken in the usual sense; there was simply nothing in the code to read.

There is a cost side too. Agents used 43 percent more tokens on inaccessible sites, and up to six times as many at the extreme. An agent pays that bill and may give up rather than spend it.

The web.dev guide is also blunt about motion. Layout that shifts as you scroll is described as functionally broken for agents, because the agent clicks what it measured and the thing has moved by the time it clicks. That single line explains a whole class of failures on modern marketing sites.

What should you fix first?

In this order. The first three are the same work that makes you readable to search engines and to AI answers, so none of it is wasted if agentic browsing stays small.

What about agents that buy?

There is a protocol for that. ucp.dev, the Universal Commerce Protocol, is a joint effort co-authored by Google, Shopify, Amazon, Microsoft, Meta, Stripe, Etsy, Target and Walmart, plus lodging and food companies. It exists so an agent can complete a checkout without scraping a human interface.

Be careful with the timeline here. A protocol shipping is not a market arriving. What we can say honestly is that the infrastructure for agent checkout is being built by the companies that own the checkout, and that being ready costs less now than retrofitting later.

Does this matter for a small business yet?

Honestly: not as a sales channel, not today. Here is the measured picture.

We checked where the searching actually is. The subject carries real search volume: ai agents measures about 31 on Google Trends, comparable to seo agency. But the advice terms do not exist yet. How to be found by ai agents and optimize for ai agents both measure zero. So nobody is searching for this guide today.

We are publishing it now on purpose. When buyers start asking how to be found by agents, the page that answers it should already exist and already be trusted. Writing it later means starting from nothing in a contested SERP; the obvious phrasing, how to be found by ai agents, already has five large domains on page one.

And the early part is genuinely early. Google files agentic work last. The useful thing you can do now is the small set of fixes above, most of which are also accessibility work you may already owe.

How would you know it worked?

Read your server logs for the agent user agents, and check whether your pages parse without JavaScript. Both are free and both are today's work.

We measure the answer side separately and can show you the checks: whether your facts are in the raw HTML, whether every control has a name, whether the layout is stable, and whether an engine can quote a plain sentence about what you sell. We report that as readiness, and we do not claim an agent has bought anything from you, because we cannot measure that and neither can anyone else yet.

Frequently asked questions

What is an AI agent?

An AI agent is software that is given a goal and then works toward it on its own, using tools rather than only answering questions. The agents launched in September 2026 each come with their own browser or cloud computer, so they read websites the way a person would.

How do AI agents find businesses?

They search and browse on the user's behalf, then read the sites they land on. A search engine crawls your text and ranks the page; an agent opens the page and tries to complete a task, so it needs your facts to be readable and your controls to be labelled.

Which companies launched autonomous agents?

Meta launched Muse on 8 September 2026 with a browser inside a secure virtual machine. OpenAI launched Dots on 29 September 2026 at DevDay, with its own cloud computer and browser linked to more than 4,000 apps. xAI runs Grok Bot as always-on agents, each with its own computer.

Is AI agent optimisation the same as SEO?

No, though the first fixes overlap. SEO targets a ranked list. Agent readiness is about whether a machine can operate your site: finding the price, the hours and the contact route. Google's own guide files agentic work last, after the search work.

Do I need to make my website agent-ready now?

Not as a sales channel, because the demand is not there yet. The measured advice terms still return no search volume. The work is worth doing when it is also accessibility work you already owe, and it is cheaper now than retrofitting later.

What is the single most common thing that breaks for an agent?

Content that only exists after JavaScript runs, and controls with no accessible name. AudioEye measured the cost: with 1,560 agents across six models, task completion was 31 percent on a site with accessibility issues against 96 percent on an accessible version of the same site. Google's web.dev guide also names unstable layouts, describing pages that shift as you scroll as functionally broken for agents.

How do I check if my site is agent-ready?

Two free scorecards exist. Cloudflare's isitagentready.com scores discoverability, content, bot access and capabilities. Vercel's is-agentic.com runs 118 checks and explains each failure. Both are free and need no account. Neither is a certification, so read them as a prioritized list of fixes rather than a grade.

Can AI agents buy from my store?

The protocol exists and the platforms are building it. Universal Commerce Protocol, at ucp.dev, is co-authored by Google, Shopify, Amazon, Microsoft, Meta, Stripe, Etsy, Target and Walmart, and Stripe and Shopify both ship agent checkout work. But the buying habit is not here yet, so treat this as preparation rather than a channel.

Should I build an app for AI agents?

No. These agents do not arrive through an app store. Each gets its own browser or cloud computer and reads your existing website like a person. There is nothing to publish to, only your site and whether a machine can read it. An app is also permanent maintenance and does nothing for the search engines that send you traffic today.