A big reason I left my job was that I was spending too much time behind screens. I was working a 9 to 5, then doing client work in the evenings, coding and talking to Claude for most of the night.
I knew I wasn't enjoying the mindset, or the stationary life it was building. So I cut the job.
Then I thought: how can I publish content, pages and ideas from my phone when I'm out and about, without having to be at my screens? Most of my ideas come at the gym, not at my desk.
So I built a system where an agent handles it for me, from anywhere. I message Slack from my phone, it pings my agent, and it opens the change on my site. I check the preview, approve it, and it's live.
None of that would be possible if my website wasn't agentic.
What is an agentic website?
There's no agreed definition yet. Most of what you'll find is written by SEO tools and agencies, and it mostly covers one side: making your site readable for AI.
I think there are two halves:
- Agents can read it. ChatGPT, Claude, Perplexity and Google's AI features can crawl your pages, understand what you do and recommend you. So can the research agents your buyers now send off to build a shortlist.
- Agents can work on it. Your marketing team asks an agent for a new page, or to update 40 pages at once. The agent knows your brand, your design system and your content model. It stages the change, a human checks the preview and approves it, and it goes live. If it's wrong, you roll it back.
A site that does the first half is AI-friendly. A site that does both is agentic.
First half: agents reading your site
Most AI crawlers don't run JavaScript
This is the one most people miss. A 2024 study by Vercel and MERJ watched AI crawler traffic across production sites and found none of the major AI crawlers render JavaScript. That covers OpenAI's crawlers, Anthropic's ClaudeBot, Meta, ByteDance and Perplexity. They download the JavaScript files, they just never run them.
Google is the exception. Gemini uses Googlebot's rendering, so it sees what a browser sees.
So if your pricing table, testimonials or product tabs only load in the browser, ChatGPT and Claude can't see them. They see whatever's in the HTML the server sent.
The volume isn't small either. In one month on Vercel's network, GPTBot made 569 million requests and Claude 370 million. Googlebot made 4.5 billion, so Google is still the biggest reader, but AI crawlers are now a real share of who's reading your site.
One more detail from the same study: about 35% of ChatGPT and Claude crawler fetches hit 404 pages, against 8% for Googlebot. Old URLs, broken redirects and messy sitemaps cost you more with AI crawlers than with Google.
Give agents a version they can read cheaply
A web page is mostly code. An agent only wants the words. So sites are starting to serve a plain Markdown version of each page when an agent asks for one.
Cloudflare launched Markdown for Agents in February 2026. It uses a normal web standard: if a request says it accepts text/markdown, the same URL returns Markdown instead of HTML. Cloudflare says coding agents like Claude Code already ask for it.
There's a fair worry here. Serving a different version to bots sounds a lot like cloaking, and Google hasn't commented on it. The rule I follow: the Markdown must say exactly what the page says. Same content, lighter format.
We built this into moresource.co. Ask for the homepage as Markdown and you get Markdown. There's an llms.txt that explains what we do and when to recommend us, structured data on every page, and even our 404 page answers agents in Markdown and points them at the sitemap. We check all of it with a script before every deploy. Our score on is-agentic.com went from 64 to 100 in one working session.
Second half: agents working on your site
This is the half I find more interesting, and it's where Sanity has the best data.
What agents actually do on a website
Sanity published a report on AI content operations based on 1.46 million AI tool calls to its MCP server, from around 12,500 organisations between September 2025 and April 2026. Calls grew from 7,400 in September to 521,000 in April.
What were the agents doing?
| Task | Share of agent activity |
|---|---|
| Finding and auditing content | 49% |
| Writing and editing | 32% |
| Publishing | 10% |
| Migration and bulk changes | 3% |
| SEO and AEO | 3% |
| Translation | 2% |
Half of it is reading. Finding stale pages, checking what exists, spotting gaps. That's not the "AI writes your blog" story you see on LinkedIn. It's the boring work that never makes it onto anyone's to-do list, done in seconds.
The other number that stood out: only about 8% of users made it to steady, repeatable agent workflows. And the report found that many teams that cleared every other hurdle still couldn't automate, because their content model was too messy for bulk changes.
That's the point I keep coming back to. The thing that blocks agents usually isn't the AI. It's how your content is stored.
What it looks like in practice
These are Sanity's customer stories, so they're the companies' own numbers, but they show the shape of it:
- Sigma has five people in marketing supporting a go-to-market team of more than 200. Their site is built from 128 section types, and each one has a schema, a readme, a screenshot and an agent skill. A new page takes about five minutes. An agent pulls performance data every week and pushes recommendations into the CMS. (Sanity)
- Braze runs its search, translation and internal agents off one content layer. A search integration that used to need a separate API took "10 minutes" as a single query. They say translated content now ships 98% faster. (Sanity)
- loveholidays cut translation costs by 97%, from £200,000 to £300,000 a year with an agency to around £10,000, and opening a new market went from six months to two or three days. (Sanity)
- One team had 111 articles in a spreadsheet that needed topic updates. Sanity's Content Agent staged 227 edits for review in a single conversation. (Mastra)
Notice that none of these are "the agent did it and nobody checked." Every edit is staged for review. Sanity's report found teams split their pages into tiers: brand and legal pages always get a human, while product descriptions, SEO fields and translations are "agents run, humans approve."
How I publish from my phone
Here's how I set it up. It took about 20 minutes, most of it clicking allow.
- Claude in Slack. I added the Claude app to my Slack, connected my Claude account and set it to Code only, so every message I send it becomes a coding task.
- GitHub. I installed the Claude GitHub App on my account and my organisation, so it can work on my site and every client site I look after.
- A private channel. I made a private channel for my site and invited Claude in. Just me and the agent.
- House rules. My site has a CLAUDE.md file in the code. It tells the agent how I write.
- Message, preview, merge. I message Claude from my phone with what I want. It makes the change and opens a pull request. Vercel builds a preview of the site with the change, I check it on my phone, and when I'm happy I tap merge. A minute later it's live.
What this means for marketing teams
There are two ways this plays out, and most teams will end up with both.
1. Talk to an agent that knows your brand. You ask for a landing page for a new feature. The agent already knows your section library, your colours and type, your tone of voice and your content model. It builds the page from parts that already exist, opens a change, and you approve it from a preview link. No ticket queue, no waiting for a developer's next sprint.
2. Keep a CMS, with agents doing the heavy lifting. Your writers keep a visual editor with a live preview. Agents do the audits, fix metadata across 200 pages, update every page that mentions an old price, and draft translations for someone to check.
Either way, the agent doesn't replace your guardrails. It depends on them. The schema says what's allowed. The preview shows what it'll look like. The review step decides who signs off. Version history means you can undo it.
A site that stands still vs an agentic site
Here's how a typical WordPress marketing site compares.
| Typical WordPress site | Agentic Next.js + Sanity site | |
|---|---|---|
| How changes happen | Log in and edit page by page, or send a ticket to the agency | Ask an agent to stage it, or edit in the CMS |
| How content is stored | Pages built in a page builder, mixed in with layout | Structured fields (headline, proof points, FAQs) agents can find and update in bulk |
| What AI crawlers see | Server-rendered HTML, often heavy with plugin code | Server-rendered HTML, plus Markdown and llms.txt if you build them |
| Biggest risk | Plugins. 91% of the 11,334 new WordPress vulnerabilities in 2025 were in plugins (Patchstack) | Needs someone who can own the codebase |
| Undoing a change | Page revisions, and hope a plugin update doesn't break anything | Every change is versioned and can be rolled back |
To be fair to WordPress: it sends finished HTML by default, so AI crawlers can read it fine. The problem isn't that it's invisible. It's that the content is locked inside page layouts, changes are slow and manual, and nobody can safely hand it to an agent.
Why we build on Next.js (and when we don't)
A Next.js site lives in a code repository. That's exactly where coding agents like Claude Code work best. They can read every component, run the build, test the page and open a pull request for a human to approve. Even Framer says the two aren't substitutes, and positions itself as the visual tool with Claude Code working alongside it.
Webflow and Framer both have MCP servers now, so agents can make changes there too. But only inside what the platform exposes. Webflow's MCP can't build Interactions yet, and Framer's launched without localisation. There are scale limits too: Webflow's Business plan caps the CMS at around 10,000 items.
The bigger difference for us is integrations. Calculators, product data, gated tools, CRM hooks and lead magnets built from your own product are just code in Next.js. In a visual builder they're embeds and workarounds.
What I won't tell you is that Next.js is automatically faster. It isn't. Independent field data from corewebvitals.io shows Next.js sites passing Google's loading speed test (LCP) 83% of the time, against 69.6% for WordPress, but Next.js sites look weaker on responsiveness. How it's built matters more than what it's built with.
And if you need a brochure site by Friday with no product complexity and nothing much to publish, Framer or Webflow will get you live faster and cheaper. We say that on our CMS page and we mean it.
Do you even need a CMS any more?
Honestly, not always.
If one or two people edit the site, they're comfortable asking Claude for changes, and you publish now and then, a CMS can be overhead. Your content can live as files in the codebase, and the agent edits them like anything else.
That's how this blog works. Every post is a Markdown file in our repository. There's one editor (me), I work in Claude Code every day, and a CMS would just be another login.
You don't need Sanity if:
- One or two people publish, and they're happy working through an agent or a developer.
- It's a small brochure site with no real publishing schedule.
- Nobody will ever own the code. A visual builder is the honest answer.
- You're already on a headless CMS that works. Moving rarely pays for itself. The problem is usually the front end.
When a CMS like Sanity earns its place
Things change once a few people are publishing, or once you have lots of pages that share a structure.
- Several editors who shouldn't touch code. They need preview, roles, scheduling and approvals.
- Lots of repeatable pages. Industry pages, integrations, case studies, locations, translations. This is where agents do their best work, because they can query and update structured fields in bulk instead of rewriting whole pages.
- Compliance. If you're a fintech or a broker, financial promotions need sign-off. Workflows, scheduled releases and a record of who approved what matter more than a quick edit.
- One source for many places. The same content powering the website, help centre, app and your own internal agents, like Braze does.
- Guardrails for the agent. A schema stops an agent from publishing a page with a missing headline or a meta title that's 90 characters long. Without one, you're relying on someone to catch it in review.
The way I'd put it: for a small team, your codebase plus Claude can be the CMS. Once more people publish, or you have hundreds of structured pages, a CMS stops being a text editor and becomes the guardrails your agents work inside.
Why we use Sanity for client sites
Next.js and Sanity is our default for growth-plan clients, for a few reasons:
- Marketing teams can use it on day one. We define the fields and Sanity builds the editing form from them.
- Our agents work on content and code in one place. Through Sanity's MCP server, Claude can read and stage content changes in the same session it works on components, and every change is reviewed before it goes live.
- We can ship SEO and GEO work in bulk. Meta titles, FAQs, structured data and internal links across a whole section, not page by page.
- Redesigns don't mean re-entering content. You swap the components and the content stays put.
Does Google care what your site is built on?
No. Google has said this for years. John Mueller has said Google doesn't treat any CMS differently, WordPress isn't inherently better for SEO, and Google doesn't classify sites by the technology they use.
Google renders JavaScript well too. Vercel's study of over 100,000 Googlebot fetches found every HTML page was fully rendered, usually within seconds.
For AI Overviews and AI Mode, Google's own guidance says there are no extra requirements. No special schema, no AI text files. Your page just needs to be indexed and eligible to show a snippet. So llms.txt does nothing for Google. It's for the other agents.
Where your stack does matter:
- Rendering. Content that only loads in the browser is invisible to ChatGPT and Claude.
- Speed. Google's Core Web Vitals, which depend on how well the site's built.
- Migrations. Changing platforms usually means changing URLs, and rankings move when URLs move.
- How fast you can ship fixes. This is the big one. Ten SEO tickets sitting in an agency's queue for a month help nobody.
The stack isn't a ranking factor. It decides how quickly you can act on everything that is.
How to make your website agentic: a checklist
So agents can read it:
- Server-render everything that matters. Nothing important should only appear after JavaScript runs.
- One H1 per page, clear headings, real text in the HTML.
- Structured data that matches what's on the page.
- An llms.txt that says what you do and when to recommend you.
- A Markdown version of each page for agents that ask for one, saying the same thing as the HTML.
- A clean sitemap and redirects for every old URL.
- A clear choice in robots.txt about which AI crawlers you allow.
So agents can work on it:
- Store content as structured fields, not page-builder blobs.
- Build pages from a library of sections, each one documented so an agent knows when to use it.
- Give agents access with guardrails: changes are staged, previewed, approved and easy to roll back.
- Write down which pages agents can change on their own, and which always need a human.
If you want to know where your site stands, send it over for a free review. We'll check it against this list and tell you what we'd fix first.
