Blog · Inside · 19 November 2026
How to build an agent-first website, from our first six weeks and 5 million page views
Vlad and I recorded our first podcast, about building a website for agents and building it with them. Here are the key takeaways from our first weeks, what our dashboard shows behind 5 million page views, and what we'd tell someone starting now.
Vlad and I recorded our first podcast, two conversations about how we build Anchor Terminal. In one we talk about what agent-first means, for the agents who read the site and for the agents who build it. In the other we open our admin dashboard and go through the numbers live [1].
Watch the podcast on YouTube (34 minutes).
We launched on 26 September 2026. When we recorded, the dashboard's all-time view showed over 5 million page views from over a million IP addresses, and 1.5 million of those requests were for the Markdown and JSON versions we publish for agents [2]. Neither of us expected that, and we're treating the whole thing as an experiment. This is what we've learnt so far.
Key takeaways
- Ask the agent you already work with what it wants. It will tell you which formats it prefers and which cost fewer tokens.
- Keep a short version of every page. Agents picked our short Markdown files with the facts and the ranking over the long pages with every review on them.
- Give agents a way in. An llms.txt that says where everything is does for a website what a CLAUDE.md does for a codebase.
- Name URLs the way an agent would guess them, and answer a wrong guess with the closest pages you have.
- Keep everything in files for as long as you can. An agent finds a file in the repository without needing an MCP server or a database login.
- Use fewer frameworks and libraries. Fewer ways to do a thing keep an agent focused on what to do.
- Split your traffic into crawlers, agents and people before you celebrate a number. Most of our launch-day page views weren't people.
- Search still decides whether agents find you. An agent mostly reads what Google or Bing gave it.
Building for the agents who read the site
People use "agent-first" for two things, building for agents as your users and using agents to build. Vlad's view is that the term took off in about the last half year [1].
The best place to start is the agent you already work with. Ask it what data formats it prefers and what's most token-efficient. If agents are your readers, every page they read comes out of somebody's token budget, so you don't want a page to burn through a user's credits.
We learnt this as the pages grew. As we added reviews and research to each listing, agents used the short Markdown file with just the facts and the ranking more than the long one with all the reviews [1]. Every page now has a Markdown twin, a slim one and a JSON one [3]. Plenty of agents still read the HTML, so that has to be good as well, but I think it will change over time. Our MCP server is getting more use, and the API still takes most of what agents ask for [1].
Then help them find their way around. Our llms.txt tells an agent where to look for tools and where to look for comparisons [3]. It's the same idea as the CLAUDE.md or AGENTS.md in a repository, a summary that points an agent to the right place for the job it was given.
Structure matters on both sides, folders inside and URLs outside. Agents guess URLs, and if your pages are named the way they'd guess, they find the content. We saw agents guess comparison URLs with a product's full name, such as OpenAI against Google Vertex, where ours used a shorter one [1]. It's a guessing game, so a URL that doesn't exist now answers with the closest matches we have and a link to llms.txt, in Markdown too [4].
Being independent helps here. When someone asks their agent to compare two tools, the agent searches and reads the comparison page Google or Bing gives it. A vendor can write a comparison on its own site, but it's comparing itself with everyone else. We never take payment to move a rank [5], and in my view that gives a page like ours a better chance of being the one an agent reads.
Building with agents inside the repository
The other half is how we build, and three choices made the most difference.
There's no database. Listings, research, reviews and even the traffic logs are files. Vlad's reasoning is not to optimise too early. A database needs an MCP server before the agents can reach it, and access for whoever is developing. With files, if something is there the agent finds it, and if it isn't, the agent creates it [1].
There are no frameworks and few libraries. It's Go, HTML, JavaScript and CSS. Go has one way to do most things, which keeps the agent on what to build instead of choosing how. Go is on most machines too, so an agent can write a feature, run the tests, compile and see what failed in the same session [1]. The repository is about 62% Go, 19% JavaScript, 9% Python and 5% CSS, with some shell.
Most features have a CLI. An agent can run one to answer a question instead of reading the code, which saves a lot of tokens [1].
Then we wrote down how we work. CLAUDE.md says where everything lives. There's a voice file for each of us, one for the whole site with the things we never want on it, and a facts file with the true things about us that a post may say [6]. Research runs two agents on the same job and a third that compares their answers and checks them again, and that cut a lot of the hallucinations and confident wrong facts [1]. Companies that claimed their listing have mostly been happy with the research, with a few small corrections. Some told us they aren't built for agents and asked us to look at other things in the reviews, which is fair [1].
It moves fast. In the podcast I said 4,000 merged pull requests, and that was the commits [1]. Since the first commit on 25 September 2026 the repository has 3,888 commits and 900 merged pull requests, from the two of us and our agents.
How each of us works
My setup started differently. I was doing the social media, the content and the logo, starting new threads and giving them our design guideline. As the repository grew, I got more consistent results working with my agent inside it, with the whole repository as context [1]. Vlad starts every session by cloning the repository.
I use both Claude and Codex. In my experience Codex is better for social content, and Claude for building features in the product. I edited this podcast with Codex as well, and still used a couple of tools along the way, Canva for music and FLORA for the thumbnail [1].
I think that's where things are heading. You have one main agent running your workflow, and you hand a specific task to a specialist tool when you need it done at a higher quality. So a tool has to show up for agents to get picked for that kind of work, and stay ahead of the models that are eating into its business, which is what our next podcast is about.
Vlad works in Claude on the web, in the sandbox Anthropic gives you, a big virtual machine that cleans itself up. I code locally and still prefer it, though I always worry that something goes off track and takes over my social media accounts [1]. Vlad thinks Claude or ChatGPT on the web gets you 95% of the way [1]. Either way most of our work now is reading pull requests. Every change is built and checked before it reaches us, and the whole site is generated on every deploy, so nobody needs to test on their own machine [7]. It also means a new person could start with a Claude session and the repository.
What 5 million page views are made of
Vlad thought nobody would look at the site, so we launched without analytics and only added them on the Sunday. Within ten minutes there were thousands of hits, and 30,000 to 60,000 page views on launch day [1]. A lot of that wasn't people, so we split the traffic into three groups [8].
- Crawlers fetch pages for a search engine or an AI index, such as Googlebot, Bingbot, Applebot and Meta's crawler. They say who they are in their User-Agent, some publish the IP ranges they crawl from, Google among them [9], and they fetch a page and leave without running JavaScript.
- Agents fetch a page because a person asked something, such as ChatGPT, Claude or Perplexity fetching for a user. They also say so in their User-Agent, and they mostly land on comparison pages, because the person asked what the difference is between two tools.
- People are browsers that ran our page script and kept a page in view for at least five seconds [8]. There are far fewer of them than crawlers or agents, in the hundreds, because we build for agents first [1].
Crawlers are good news. One that keeps coming back means a search engine is interested in you, and over time it sends people and agents your way. You want them back when something changed, so the updated date on each page and the sitemap you give them matter. Search engines mostly do what your sitemap says.
Agents come after crawlers. ChatGPT or Claude doesn't know which pages to read until it has searched Google or Bing, so your agent traffic depends on the crawlers having been there first. When I tested how ChatGPT finds a tool, it went out and read the first comparison page it found every time [1].
The stats are our own. Every request reaches our Go server, which writes a line to a file for that day and that address, with how long a person stayed, and the dashboard is built from those files [8]. The site sets no cookies unless you sign in and runs no third-party analytics [10]. Vlad suspects Google Analytics helps a site in Google's results. We left it out anyway, because we'd rather keep what our readers do to ourselves.
Here's what the dashboard showed when we went through it [2].
- The most requested page of all time is the about page. Crawlers come back to see whether it changed, and people read it first to find out who we are.
- The MCP page and the reviews come next, and agents read the MCP page far more than people do, which tells us we're reaching the right audience.
- robots.txt is near the top, from crawlers checking what they're allowed to read.
- The page people read most was Google Veo, Google's video model.
- Google still sends the most people, then Reddit and X when we've been posting there, then Bing and LinkedIn.
- Most visits are from desktop, because crawlers and agents ask for the desktop page.
- Applebot, Meta and Amazon index us a lot, and ChatGPT fetching for a user is high.
What we're trying next
We'd like more reviews and more people using the site, because the reviews are something a competitor can't copy. We build for agents first, and we still build things for people, the business sign-up and the company dashboard, because that's how we've built all our lives. I think that will move towards agents over time.
Vlad's post Agent first, inside and out has more on the files our agents read before they start [6].
References
Read on 19 November 2026.
- [1]Anchor Terminal, How to Build an Agent-First Website: Lessons From Our Launch Hitting 5M views after 6 weeks, YouTube, 19 November 2026, 33 minutes 52 seconds. Source for everything Vlad and I said in the conversation, the launch weekend, how we work and what agents read. youtube.com
- [2]Anchor Terminal admin dashboard, all-time view, read out in the podcast in November 2026, not published. Source for over 5 million page views, over a million IP addresses, 1.5 million Markdown and JSON requests, the most requested pages, where visitors came from and the crawlers.
- [3]Anchor Terminal, for agents. Source for the Markdown, slim and JSON twins of every page, llms.txt, the MCP server and the API. anchorterminal.com
- [4]Anchor Terminal, the not-found page, read on 19 November 2026 at /compare/openai-vs-vertex-ai, a comparison URL that doesn't exist. Source for a missing page listing the closest matches and linking to llms.txt, in HTML and Markdown. anchorterminal.com
- [5]Anchor Terminal, the benchmark, "Vendors can't pay for placement. Reports are paid, rankings aren't." Source for never taking payment to move a rank. anchorterminal.com
- [6]Vlad Cealicu, Agent first, inside and out, Anchor Terminal, 15 November 2026. Source for CLAUDE.md, the voice files, the facts file and the other files the agents read before they start. anchorterminal.com
- [7]Anchor Terminal, CLAUDE.md and the deploy, 19 November 2026, not published. Source for every change being built and checked before a pull request is merged, and the whole site being generated on every deploy.
- [8]Anchor Terminal, the session logs and their report (internal/sessions), 19 November 2026, not published. Source for the three groups, a person being a browser that ran the page script and kept a page in view for five seconds, and one file a day per address.
- [9]Google, Verify requests from Google crawlers and fetchers, Google for Developers. Source for Google publishing the IP ranges its crawlers use. developers.google.com
- [10]Anchor Terminal, privacy. Source for no cookies unless you sign in and no third-party analytics. anchorterminal.com