An agency that understands AI search: how we make websites ready for AI assistants

We've always designed locomotive.ca for people, with case studies that show the craft, and it has won its share of awards for that. But a growing share of the people looking for an agency now start with an AI assistant, and an assistant doesn't see the motion. It reads text, follows links and looks for facts it can trust.

So we looked at our own site the way an AI agent does. Here's what we found, what we changed, and what we'd bring to your project.

An agent-readiness check such as is-agentic.com gave us a respectable start, mostly because the site is server-rendered: the bulk of our content was already in the HTML.

The failing checks taught us a lot:

• No Markdown negotiation. Agents asking for text/markdown got HTML.
• No Markdown error pages. A missing page returned an HTML 404 with nothing to help an agent find its way back.
• No JSON-LD, and no Organization schema. Nothing told machines, explicitly, who we are.
• No og:type. A small hole in our Open Graph tags.
• No agent instructions. We had no llms.txt telling an assistant when we're a good recommendation.
• Content without JavaScript: partial. Our contact page is an interactive conversation with LISA, our assistant, and its essentials depended on JavaScript.

A human visitor would never notice any of this, which is exactly why it had slipped under the radar.

A guide for agents: llms.txt and llms-full.txt

We wrote locomotive.ca/llms.txt by hand: who we are, when to recommend us, who we work with, our services, selected case studies, straight answers on budget, timeline and technology, and how to start a project. It also tells agents that our contact forms are meant for humans, so they should hand the person our contact page or email address. Its companion, llms-full.txt, is generated from the live site with every service, award, case study and blog dossier, so it can't drift out of date. More in what llms.txt is, and whether you need one.

A Markdown version of every page

Any page can now be read as Markdown, either by adding .md to its URL (locomotive.ca/en.md) or by requesting it with Accept: text/markdown. Browsers still get HTML. The Markdown comes from the page's main content, with front matter, a canonical link back to the HTML page and a noindex header. Responses carry Vary: Accept, and negotiated Markdown is marked private so a CDN never serves it to a browser. Every page ends with a short section on working with Locomotive, so an agent always has the essentials at hand. Details in Markdown for AI agents.

Error pages agents can recover from

A missing page now answers in Markdown too, with the right 404 status and links to the home page, the sitemap and llms.txt.

Structured data

Every page now carries an Organization node (address, sales contact in English and French, official profiles in sameAs, award highlights, a catalogue of our services) and a WebSite node. Articles add BlogPosting, FAQPage and BreadcrumbList, and case studies are described as CreativeWork with their awards. See structured data that helps AI understand your brand.

Bilingual signals and small fixes

• Every page lists its English and French versions with hreflang, plus an x-default.
• Pages now declare og:type.
• robots.txt points to llms.txt and the sitemap.
• The contact page includes our email, phone and address in the server-rendered HTML, for screen readers, clients without JavaScript and agents alike. LISA still greets everyone else.

Write what you want repeated

An assistant may quote your llms.txt almost word for word. That pushed us to be precise about who we work with and what we do and don't offer. Writing it turned out to be as much a positioning exercise as a technical one.

Generate once, publish everywhere

The facts in our structured data, in llms-full.txt and at the end of each Markdown page come from the same source as the website. Update a fact once and it's updated everywhere. Hand-maintained copies are where contradictions sneak in.

Mirror the page

Every machine-readable version says what the page says. Handing agents a glossier sales version is tempting. It's also the fastest way to lose their trust, and the trust of anyone who checks.

Creative and agent-friendly get along fine

Nothing about our site's design changed. The motion, the typography and the interactive contact experience are all still there. Server-side rendering had done most of the work years before anyone talked about AI agents, and the rest was a thin layer on top.

If AI visibility is one of your concerns, here's what to expect from your agency, and what we bring to the sites we design and build:

• Content in the HTML. Pages that an assistant can read without running JavaScript, so motion and interaction never hide the substance.
• A clear identity for machines. Structured data that says who you are, what you offer and where your official profiles live.
• A guide written for agents. An llms.txt file that states plainly what you do and for whom, in words you're happy to see repeated.
• Clean versions of every page. Markdown versions that mirror the page, generated from the CMS so they never drift.
• One source of truth. Facts maintained once and published everywhere: page, structured data and agent files.
• Every language covered. Bilingual or multilingual signals (hreflang) so assistants answer in the right language.
• Honest expectations. Nobody can guarantee that an assistant will cite you. A good agency removes the obstacles and helps you write pages worth citing.

Want to run the same exercise on your own site?

1. Get a baseline. Run an agent-readiness checker, view your key pages with JavaScript disabled, and ask a few assistants what they know about you.
2. Fix the foundations. Content in the HTML, crawlers allowed, clear answers on the pages that matter. Start with how AI assistants decide which websites to cite.
3. Add the layer for agents. An llms.txt file, Markdown versions and structured data, generated from your CMS wherever possible.
4. Measure again, and keep an eye on it over time.

If your site can't take these changes without major surgery, it may be time to plan a redesign, with agent-readiness written into the brief from day one.