An AI agent is software that reads the web and acts on it. It shortlists suppliers, drafts the comparison, books the call. More of your customers use one every month, often without saying so. Getting ready means making your business readable and trustworthy to software, not only to people. Four things carry most of the weight, and a small firm can do all four in a week. Most do none of them, which is why the firms that bother are already pulling ahead in AI answers.
Here is what an agent actually does with your business, where the gaps usually are, and how to close them.
Why this is worth your time now
ChatGPT reached about 900 million weekly users by April 2026, up from 400 million a year earlier. Google’s AI Overviews serve answers to around 2 billion people a month. Perplexity handles tens of millions of queries a day. When a buyer asks an assistant ‘who does fractional CTO work in Wales’, the assistant reads the web, picks a few names and hands them over. If it cannot read you, or does not trust what it reads, you are not on that list.
This is not a far-off shift. It is happening in the searches your buyers run today, and the firms that appear in those answers are not always the ones with the biggest marketing budgets. They are the ones whose sites a machine can read and quote without guessing.
What an agent actually does with your site
It is worth being concrete, because the word ‘agent’ hides a very ordinary process.
A buyer types a question. The assistant decides it needs current information, so it sends a crawler to fetch pages. It reads the raw HTML of each page, pulls out the structured data if any exists, looks for a direct answer to the question, and judges whether the source is credible enough to repeat. Then it writes an answer, names two or three sources, and sometimes offers to take the next step, such as drafting an enquiry.
Every one of those stages can fail on your site. The crawler can be blocked. The page can be empty until JavaScript runs, which the crawler does not do. The answer can be buried under marketing copy. The structured data can be missing or broken. And the next step can have no obvious path. Each failure is cheap to fix and expensive to leave.
The four things to get right
Let the right crawlers in. Check your robots.txt allows the AI crawlers you want, such as GPTBot, ClaudeBot and PerplexityBot. Many sites block them by accident, usually because a security plugin or a well-meaning developer added a blanket rule years ago, and then wonder why they never appear in AI answers. Allowing them is a one-line decision. If you are nervous about it, remember that blocking them does not protect your content, it only removes you from the answer.
Publish clean, structured facts. Agents prefer sources that state things plainly and mark them up. A direct answer in the first hundred words of a page, clear headings, a short FAQ, and schema.org data (Organization, Service, FAQ, Article) all help software understand what you do and quote it correctly. Make sure your pages render their content without needing JavaScript, because many agents read the raw HTML. A page that looks fine in a browser and is blank to a crawler is a common and invisible problem.
Keep a current llms.txt. This is a plain markdown file at the root of your site that gives an AI a short description of your business and links to your key pages. It was proposed by Jeremy Howard at Answer.AI in September 2024 and has since been adopted by Anthropic, Cursor and others. Think of it as a menu you hand the agent so it does not have to guess. A stale one is worse than none, because it points the agent at the wrong map. If you have launched a service, renamed one or added a section since you wrote it, it is already out of date.
Make your key actions followable. If an agent, or the person it is helping, wants to contact you or book a call, the path should be obvious and machine-readable. A clear contact form, a visible booking link, plain phone and email. No dead ends, no ‘click here’ with nowhere to go, no contact details hidden in an image.
What this looks like in practice
Take a 30-person professional services firm with a decent website. The crawlers were blocked by a plugin nobody remembered installing. The service pages opened with three paragraphs of brand story before saying what the service was. There was no schema beyond the default. There was no llms.txt. The contact page had a form that only worked with JavaScript.
An afternoon of work fixed all of it. The plugin rule was removed. Each service page got a two-sentence direct answer at the top and a short FAQ at the bottom. Service and FAQ schema were added. An llms.txt was written listing the six pages that mattered. The contact form got a plain email and phone number alongside it.
Nothing about the business changed. Within a few weeks it started appearing in assistant answers for the questions it actually wanted to be found for, because for the first time a machine could read it, understand it and trust it.
The common mistakes
Treating it as a marketing project. It is mostly a technical hygiene project, and the technical fixes are quicker and more durable than any copywriting.
Writing an llms.txt once and forgetting it. The file has to track the site. Put it in the checklist for every new page and every renamed service.
Blocking crawlers out of caution. The content is already public. Blocking the crawler removes you from the answer without protecting anything.
Over-optimising for one assistant. The basics, open access, structured facts, direct answers, work across ChatGPT, Perplexity, Google and Copilot. Chasing any single one is wasted effort.
The honest caveat
None of this is a magic switch. Google has not committed to reading llms.txt for ranking, and the standard is young. Do the durable basics first, allow the crawlers, structure your pages, tell the truth clearly, and treat llms.txt as a cheap, sensible addition rather than the whole answer. The firms that win here are the ones with genuinely useful, well-structured content, which is the same thing that has always earned trust online. Agents have simply made the gap between well-structured and badly-structured sites far more visible.
Where to start
Run one honest audit of your own site. Can an AI reach it, read it without running scripts, and understand who you are and what you sell. Fix what it cannot, in the order above. An AI opportunity review covers this alongside where AI can help inside the business, our piece on how agentic search changes discovery covers the buyer side, and our Insights library shows the structure in practice.
Frequently asked questions
What is an AI agent, in plain terms?
Software that reads the web and takes steps for a user, such as researching suppliers, summarising options or filling a form. It works from what your site publishes, so what you publish decides what it can do with you.
Do I need an llms.txt file?
It helps and it is cheap to add. It gives an AI a clean summary and a map of your key pages. Keep it current, because a file that describes an old version of your site does more harm than good.
Will blocking AI crawlers protect my content?
It also removes you from AI answers, where a growing share of buyers now start. For most businesses that want to be found, allowing the main AI crawlers is the right call.
Does my site need to work without JavaScript?
The content should be present in the HTML the server sends, because many agents do not run scripts. A page that only fills in after JavaScript runs can be invisible to them.
Which schema should a small business add?
Organization on the home page, Service on each service page, FAQPage wherever you answer questions, and Article on blog posts. Those four cover most of what an agent looks for.
How do I know if my site is agent-ready?
Check three things. The crawlers are allowed, the content is in the HTML rather than loaded by script alone, and your key facts carry schema markup. If all three hold, most agents can read and quote you.