An agent-first website keeps its content and value built for humans, then adds a layer on top that AI agents can read, understand and use without guessing. The human still makes the decision; the agent just does more of the reading, comparing and pricing on that human's behalf.
Here's how I've been saying it: the content and the value are human-first, but the infrastructure, or the layer on top of that, is agent-first.
On October 1 and 2, 2026, we rebuilt the layer on top of sidekickstrategies.com around that idea. This is part 1. I'll walk you through why we did it, what an agent-first layer actually is, and a little of how we built ours, with the real commands and real output so you can check my work.
Part 2 comes after our October 16 readout, with real traffic and agent-usage numbers. Measurement is underway right now, so you won't find results here. I'm not going to guess at them.
Why Would a Website Need to Be Agent-First?
Back in July 2025, on HubHeroes, I said this: "I think AI answer engines and zero click have absolutely shrunk the top of the funnel." I still believe it. What's changed since then is who's doing the reading.
More of your buyers ask an AI assistant before they ever visit a website, and we want our answers to be the clear ones when that happens.
That's the whole why. Humans still decide. Their AI assistants increasingly do the reading.
So what's an agent? It's software that reads and acts on websites for a human who asked it to. Cloudflare describes agents as "software acting on behalf of people. They sit somewhere between humans and traditional bots."
Picture what that agent runs into on a typical website: navigation, scripts, pop-ups, and a price hidden behind a "contact us" button. A human can work around that. An agent can't quote what it can't read.
Let Me Be Real Honest About the Stakes
Profound looked at 27 million citations across ChatGPT, Gemini and AI Overviews. On unbranded category questions (questions about a type of service, not a specific company), about 4% of citations pointed to a brand's own website. On questions that mention a brand, brand websites as a group (the brand's own, its competitors' and others) made up about 58% of citations.
So two things are true at the same time. When someone asks about you by name, your own site matters, and it should be easy for an assistant to read. On the broad questions, most of the conversation happens on other sites. That means making your own site legible is one piece of the growth story, not the whole thing.
One more honest line: nothing here is proven to raise rankings or AI citations. Google says plainly that you "don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search." We didn't build this for a ranking bump. We built it so that when a human's assistant reads us, it gets the facts right.
What Does "Human-First Content" Mean Here?
The content stays for humans. That part doesn't change.
Google's own guidance on agent-friendly sites says agents access sites "in 3 primary ways: screenshots, raw HTML, and the accessibility tree," and its advice is semantic HTML and accessibility, not new protocols. That's good news. A page built well for humans is already a big part of the work.
It also fits how we've always worked. Human-powered, AI-assisted. Value-first content. And ungated by default: "Share the secret sauce, because that's what builds trust and starts conversations." None of the agent-facing layer on our site needs a sign-in.
The agent-first part isn't new content. It's the same content, with the same facts, served in a shape an agent can use. Think of it as a translation, not a second website.
What Are the Three Layers of an Agent-First Website?
I think about the infrastructure in three layers. Number one, readable. Number two, legible. Number three, usable.
Layer 1: Readable
Can an agent read your page cleanly and cheaply? Four pieces make that happen.
- Markdown views. The same page as plain, tidy text with no design code, so an agent reads it fast. Vercel's example page went from about 500KB of HTML to 3KB of markdown.
- Content negotiation. One URL. The visitor says "I'd prefer text" or "I'd prefer a web page," and the server hands over the matching version. Browsers still get the normal page.
- llms.txt. A table of contents at the root of your site, written for AI. It's cheap to publish, but keep your expectations low. Ahrefs looked at sites that published one and found "97% of those files received zero traffic in May 2026," and SEO tools sent more of the requests than AI assistants did. Google says llms.txt "will neither harm nor help" your visibility in Google Search.
- Robots and content signals. A public note telling bots where they may go and whether they may use your content for search, for answers, or for training.
Layer 2: Legible
Can an agent understand what you offer, what it costs, and what to do next? This is where our own site fell down. Our tier prices rendered as "$" and "6000" in separate pieces, and two of our FAQs still quoted an old retainer floor that didn't match the pricing page.
Prices and facts go in visible text. In a small searchVIU test from late 2025 (one page, so treat it as directional), none of the five AI systems tested pulled a price that lived only in hidden structured data. If a fact only lives in markup, don't count on an agent finding it.
Pages go answer-first. An answer-first page states the price, the offer and the next step in its opening lines. We rewrote the opening of our top ten pages this way.
Layer 3: Usable
Can an agent ask your business a direct question and get an exact answer back? That's what MCP is for.
MCP (Model Context Protocol) is a menu of exact questions an AI can ask your business, like "what's the price for 20 hours?", with exact answers back instead of a guess pulled from a web page.
Our MCP server can quote a retainer at the hours you need, check cohort and workshop dates, list upcoming events, and find open discovery-call slots. A human finishes every purchase.
We also built a tool that books a discovery call. It's live, with rate limits and guardrails, but as I write this it hasn't completed a verified booking in production. So I'm not going to tell you it books calls. I'll tell you what happened in part 2.
What Did We Actually Build on sidekickstrategies.com?
On October 1, Cloudflare's isitagentready.com checker scored our site Level 1, Basic Web Presence. After the build, it scored Level 4, Agent-Integrated.
Please don't read that as a quality grade. It's one vendor's checklist, and it grades some specs that are still unmerged drafts. It's a useful to-do list. It's not a trophy.
Every Public Page Has a Markdown View
All 1,391 public pages have one. Ask for markdown in the header and the same URL returns it:
$ curl -sS -H "Accept: text/markdown" https://www.sidekickstrategies.com/pricing | head -25
---
title: "Pricing"
url: https://www.sidekickstrategies.com/pricing
description: "Transparent HubSpot partnership pricing. 10, 20, or 30 hours per month. ..."
type: page
canonical: https://www.sidekickstrategies.com/pricing
---
# We Don't Do Retainers. We Do Partnerships.
Sidekick partnerships are billed monthly for a set block of HubSpot hours. Starter is $3,500 a month
for 10 hours, Growth is $6,000 a month for 20 hours, and Scale is $7,500 a month for 30 hours.
The response comes back as Content-Type: text/markdown; charset=utf-8 with Vary: Accept, so caches keep the markdown and HTML versions apart. You can also add .md to any URL: /pricing becomes /pricing.md.
Notice the first paragraph. Every price, in plain text, right up top. That's Layer 2 showing up inside Layer 1.
An llms.txt That Points to the Markdown
# Sidekick Strategies
> HubSpot consulting, content operations and enablement for SMB leaders and HubSpot Super Admins. ...
Every page on this site is available as markdown: append `.md` to any URL, or request the page with `Accept: text/markdown`.
## Services
- [All services](https://www.sidekickstrategies.com/services.md): Every service across strategy, implementation, AI, design and content.
A Robots File With an Opinion
Content-signal: search=yes, ai-input=yes, ai-train=no
That line says yes to search, yes to AI assistants using our content in answers, and no to training. Our robots.txt also names training crawlers like GPTBot and ClaudeBot separately as part of that opt-out.
Here's the difference that matters. A training crawler gathers content for future models. A human-triggered fetcher grabs a page because a human asked a question right now. Blocking the first doesn't have to block the second.
And to be clear, Cloudflare itself says content signals "express preferences; they are not technical countermeasures against scraping." It's a public statement of what we want, not a lock on the door.
An MCP Server Anyone Can Connect
If you use Claude Code, one line connects it:
claude mcp add --transport http sidekick https://www.sidekickstrategies.com/api/mcp
Ask for a quote at 20 hours a month and here's what comes back (trimmed):
{"requested_hours_per_month":20,"match":"exact",
"tier":{"name":"Growth","hours_per_month":20,"price_usd":6000,"price_label":"$6,000/month",
"effective_rate_usd_per_hour":300,"billing":"monthly"},
"terms":"Month-to-month. The only thing we ask is 30 days notice ...",
"next_step_url":"https://www.sidekickstrategies.com/strategy-call",
"note":"20 hours a month is the Growth tier."}
No guessing. The exact tier, the exact price, the terms, and the next step. And the next step is a conversation with a human.
The Bug We Found While Testing
Here's the honest part, because failure is feedback.
While testing, we checked what header Claude Code actually sends. It's this:
Accept: text/markdown, text/html, */*
Markdown and HTML show up with equal weight. That's a tie, and our site broke the tie toward HTML. So the exact tool we were building for was getting the web page instead of the markdown.
We fixed it on October 2. On a tie, the type listed first now wins. Browsers never list markdown, so humans still get the normal page. And the .md links in llms.txt worked the whole time, so agents that read it still got markdown.
The lesson? Test with the real tool's real request, not the one you assume it sends.
What Did We Deliberately Skip, and Why?
Here's what we passed on, on purpose.
- llms-full.txt. It's not in the llms.txt spec, version 1 or 2. It's a vendor convention that doubles the maintenance.
- FAQ schema theater. Google removed FAQ rich results from Search on May 7, 2026. We kept the FAQ content as visible text for humans and removed the FAQPage markup.
- Proxying the site through Cloudflare. Vercel's docs say its Bot Protection doesn't work behind a reverse proxy, and Vercel already verifies signed bots natively. Cloudflare stays DNS-only for us.
- Agent-run checkout. Services don't fit the shopping feeds. Google Merchant Center says "Labor, time, effort, expertise, or actions, which do not result in ownership of a tangible product are not allowed." And for retainers, a call stays the next step. HubSpot is a practice, not a project. I want to be the sidekick on speed dial, not the vendor you call once, and that starts with a conversation.
A fair critique worth naming: Google's John Mueller called converting pages to markdown "such a stupid idea." His target was sites that sniff for bots and serve them different content. Our markdown view lives at the same URL and carries the same facts as the page a human sees.
Where Should You Start?
You don't need all of this on day one. If I were starting over, here's the order I'd follow.
- Ask an AI assistant about your own business. Ask it what you offer and what it costs. Write down what it gets wrong.
- Put your prices, offers and next step in visible text. Rewrite the opening of your top pages so they answer first.
- Read your robots.txt. Decide on purpose whether you welcome assistants, training crawlers, both, or neither.
- Publish markdown views and test them for real. Use
curlwith the exact header the tools send, including the tie. - Run a baseline on isitagentready.com. Treat the result as a checklist, not a grade. Consider an MCP server once you know the exact questions your buyers ask.
What's Next?
Measurement runs through October 15, and the readout is October 16. Part 2 shares what actually happened: markdown fetches, MCP calls by tool, and whether a real booking ever landed.
Want to see it for yourself? Head to /agents and point your own AI at our site. Want help doing this for your team? We teach HubSpot teams how to make their sites readable to AI agents, from markdown pages to their own MCP server. It's a project, and we price it on a strategy call. Start at /services/agent-readiness.
So here's my question for you: when a buyer's assistant reads your website tomorrow, will it get your answers right?





