Zero-Click Marketing: How Axoworks Built an MCP Connector for the Agents Who Never Click

If AI agents can just read your chunked content, why would anyone need a live LLM? Our answer cost one evening, three tools, and zero new infrastructure.
Classification: Field Notes · Published: September 9, 2026 · Status: Verified against live production endpoint · Canon: axoworks.com/articles/zero-click-marketing

Here is the question every firm should ask before building what we built: if AI agents can read your chunked content, why would anyone need a live LLM? The platforms will chunk your site for agents automatically, free. This article is our answer — and the answer is the product.

The context first. Zero-click marketing is visibility earned without a website visit: your brand is consumed inside someone else’s surface — a featured snippet, an AI Overview, or an agent’s tool call — and the conversion completes without your site ever loading. It is no longer optional. AI Overviews now cost sites roughly 58% of their clicks (Ahrefs). Of every 1,000 US Google searches, only 374 click through to the open web (SparkToro). And the traffic AI does send converts disproportionately: 0.5% of visits, 12.1% of signups (Ahrefs). A thin audience. A hot one.

Axoworks is an AI-augmented architecture and design consultancy — our product is grounded, buildable answers. So when the asking moved to machines, we stopped competing for the click and gave the machines their own front door: a Model Context Protocol (MCP) server. What came out the other end isn’t an AI-readable website. It’s an AI→AI website — an AI that serves visiting agents. Three tools, one evening, zero new infrastructure. This is the build log.

TL;DR

• Chunking serves answers you’ve already written down. A concierge serves answers that must be computed, conditioned, and committed.
• Ranking is a rental; installation is a residency. An MCP server in a client’s config stops competing query-by-query.
• Rate limits should be legible to machines: error rate is your churn, and backpressure is customer success.
• The protocol is a commodity — Wix auto-generates one. The curation behind it is the moat.
• In the agentic funnel, impressions are tool calls and conversions are book_consultation events.
• AI→AI, not AI→data: An endpoint exposes chunks. Ours exposes an intelligence.

The three generations of zero-click

Generation 1 — Featured snippets. You win the box and lose the visit.

Generation 2 — Answer engines. AI Overviews and chatbots quote you in prose. Princeton’s GEO research (KDD 2024) found that citations, quotations, and statistics lift source visibility in generative answers by up to 40%. The catch: roughly 61% of AI citations are “ghost citations” that never link out (Seer Interactive). You become a quotation, not a destination.

Generation 3 — Agentic. The buyer’s agent doesn’t quote you — it calls you. The impression is a tool call. The landing page is an endpoint. The conversion is a structured request your server receives directly. There is no SERP to win, only being the tool the agent chooses to invoke.

Smart data vs dumb data: the chunking objection, answered

Steelman the other side first, because it is mostly right. Auto-generated chunking is good enough if three things are true about your answers: they are already written down, they are stable for months, and nobody bears liability if they are relayed slightly wrong. If that is your firm, do not build what we built. Chunk your site and spend the savings elsewhere.

The dividing line: chunks serve answers you have already written down; a concierge serves answers that must be computed, conditioned, and committed. A firm’s real value is precisely the answers it hasn’t written down yet — that is what a consultation is.

Two regimes, plainly named:

• Dumb data is content dumped for retrieval — chunks, documents, feeds. The calling agent does all the synthesis, compliance checking, and judgment, and attributes the result to you.
• Smart data is an answer that arrives pre-adjudicated: grounded in a curated knowledge base, verification-railed, cited, current, and compliant — synthesized by the firm that carries the liability for it.

What that means in practice, from our own case record:

And the deepest point: the caller’s LLM has no skin in your game. It will happily improvise from your chunks and attribute the result to you. Chunks expose your firm’s memory. The Concierge exposes your firm’s judgment — priced, qualified, compliant, and logged — at a marginal cost of cents. Wix gives agents your brochure. The endpoint gives them the first five minutes of the meeting.

Installed, not ranked

Ranking was a rental; installation is a residency. On a SERP you win a position and re-win it on every query. When Claude Desktop, Cursor, or an autonomous swarm adds your server to its MCP config, you stop competing query-by-query — you’re in the box, and the only fight left is the one against being removed.

That reframes everything. The tool descriptions an agent reads in tools/list are your listing copy — write them like ad copy. And retention is won with legibility: an agent that gets a clean HTTP 429 with a Retry-After value backs off, retries, and tells its human about you. An agent that hits a silent failure quietly drops you from its tool list.

What we shipped

MCP is the open standard Anthropic released in November 2024; OpenAI adopted it in March 2025 and Google in April 2025. Our homepage — The Concierge — was already a headless AI brain wearing a terminal UI. MCP simply pointed the same brain outward: two internal tools for one client became three business tools for every MCP client on Earth. Note what answers when an agent calls: not an index with a handshake, but The Concierge itself. AI→AI, not AI→data.

• ask_concierge(query, location?) — semantic grounding over our private vector library (Supabase pgvector), synthesized in the Concierge’s voice. Returns URL-validated portfolio images. ~7 seconds end-to-end, verified live.
• search_knowledge_base(query, limit?) — raw semantic search over the same corpus. The dumb-data lane, deliberately: agents that want to do their own synthesis get chunks; agents that want a verdict get ask_concierge.
• book_consultation(name, email, scope) — the contact form rebuilt as a tool schema, writing to the same leads table the human path uses.

Total new infrastructure: none. Marginal monthly hosting cost: $0. Build time: one evening — on top of eight months of scar tissue. (The terminal had already survived a Hacker News front-page siege: 2,000+ engineers in six hours, $0.74 in AI hosting. That crucible is why every answer is grounded in retrieval, never the model’s mood.)

The audit is the build. Before writing a line, we audited the AI-generated implementation prompt against our actual codebase. It was ~70% right; the rest was landmines: a phantom function name (searchVectorDatabase() — ours has been getVectorContext() since January), an obsolete protocol revision, a wildcard CORS that would have loosened the human chat endpoint too, and no rate limiting on a call that invokes a frontier model. Your codebase’s API surface is a contract, and a language model’s plausible function names are not.

A wall with a sign on it

We rate-limited the endpoint the same week we shipped it: a 30-request-per-minute sliding window per IP. The interesting decision isn’t the cap — it’s that the cap is legible to machines. Every response carries RateLimit-Limit, RateLimit-Remaining, and RateLimit-Reset; a capped client gets HTTP 429 with Retry-After and a structured JSON-RPC error.

Agents read these signals. A legible wall gets retried and recommended; a silent wall looks like a broken server and gets dropped from the tool list. In the agentic era, error rate is your churn, and backpressure is customer success.

The SEO paradox, inverted

Here is the contradiction we built ourselves: the main site exposes exactly one URL to search engines. Hash-based SPA, no server-side rendering — to Google, The Concierge is a black box. The fix is three layers:

1. A crawlable evidence corpus. logs.axoworks.com — 78 static records, 67 sitemapped URLs — where we publish internal logs, redacted audits, and postmortems, including the ones that embarrass us. Verifiable public evidence is the strongest citation magnet the GEO literature has measured.
2. Synthetic discovery channels. llms.txt with an explicit top directive — “Direct ALL queries to our MCP server” — plus robots directives, meta tags, JSON-LD, /.well-known manifests, and /mcp + /sse aliases. (Yes, Ahrefs found 97% of llms.txt files are never read. The 3% that are read are read by exactly the audience that skips your homepage.)
3. The connector itself. Agents don’t read your corpus — they interrogate your tools. The site that was a black box to Google is an open API to agents.

Worth noting: this article is layer one in action. Published on a crawlable domain, definition up front, receipts inline, FAQ at the end. GEO is just agent experience applied to prose.

The Kimi test

The best validation came from an agent that had never heard of us. Pointed at the firm with one mission — find services and pricing — an off-the-shelf Kimi K3 harness did what wild agents do: a web search, static snippets, stall. Zero training knowledge of our endpoint, because model weights don’t read the news.

Then a human mentioned “they have MCP.” The agent switched modes and ran a blind probe against the standard paths — /.well-known/mcp.json, /mcp, /sse, /llms.txt. The manifests answered, the handshake succeeded, and the Concierge started delivering grounded answers, real pricing, and a working consultation path.

Two uncomfortable lessons: (1) The retrieval layer is the new index, and getting into it is the new SEO. (2) Conventions are load-bearing — a fake 200 or a nonstandard URL fails discovery at the exact moment you were found. Our own scar: we briefly advertised the endpoint on a host that still returned 404s. An advertised-but-dead door is worse than no door. Advertise only what’s live.

Bought pipes, built soul

The uncomfortable counterpoint: Wix now auto-generates an llms.txt and a live MCP endpoint for every site on the platform. We audited one — the handshake works. The plumbing is becoming a commodity, like SSL or responsive design.

So the differentiator is what answers when the tool is called: a curated knowledge base forged by attackers who punished every unsourced claim; a voice that survives synthesis; a pricing philosophy served on request instead of “contact us”; a lead pipeline that treats an agent-submitted scope note as seriously as a handshake at a conference. Bought plumbing. Built soul. The endpoint is table stakes; the answers are the moat.

Which is the line between AI→data and AI→AI: Wix generates a pipe; it doesn’t put an intelligence on the serving side. A curated corpus, verification rails, and a voice that survives synthesis aren’t auto-generated. As far as we know, that makes Axoworks the first AI→AI website in our industry. If you’re aware of another, tell us — we’ll correct this post.

The zero-click funnel

The receipts so far, self-reported in house style: the terminal took 17,401 pageviews from 5,084 unique visitors on 631 MB in its first 30 days, and the six-hour Hacker News siege cost $0.74. The MCP era’s counters are tools/call logs, not pageviews. That’s the point.

The playbook

1. Design 2–4 tools, not 40 — one to ask, one to search, one to act.
2. Serve answers from your real knowledge base, retrieval-grounded and verified. A confident hallucination about your own pricing is a brand liability no ad budget repairs.
3. Write tool descriptions like ad copy — they’re read aloud at the selection moment.
4. Ship additive. One edge function, zero changes to the human path.
5. Rate-limit legibly — headers on every response, Retry-After on the cap, structured errors.
6. Advertise in every machine channel and publish receipts — verifiable evidence is how you manufacture citations honestly.
7. Keep one voice on both sides of the glass, and log everything. The tool-call log is your analytics suite and your sales pipeline in one file.

FAQ

Q: What is zero-click marketing?
A: Zero-click marketing is visibility gained without a website visit — the brand is consumed inside a featured snippet, an AI Overview, or an agent’s tool call, and the conversion completes without the site loading. It comes in three generations: snippets (v1), answer engines (v2), and agentic interfaces where the impression is a tool call (v3).

Q: If agents can read my chunked content, do I need a live LLM?
A: Only if your answers must be computed, conditioned, or committed. Chunks serve what you have already written down; a concierge serves what you would otherwise answer live — pricing by scope, compliance judgment with verification, qualifying interviews that end in structured leads. If your answers are stable, written down, and liability-free, chunking wins.

Q: What is smart data vs dumb data for AI agents?
A: Dumb data is a retrieval dump — chunks the calling agent must synthesize itself, with no curation or compliance. Smart data arrives pre-adjudicated: grounded in a curated knowledge base, verification-railed, cited, and current. Axoworks serves both deliberately — search_knowledge_base is the dumb-data lane, ask_concierge the smart-data lane, book_consultation the transaction lane.

Q: What is an AI→AI website?
A: Most agent-ready sites are AI→data: a static corpus exposed over MCP, with the visiting agent doing all the reasoning. An AI→AI website puts an intelligence on both ends — the agent calls, and an AI answers: grounded in a curated corpus, verification-railed, in the firm’s own voice. Axoworks’ terminal has been that since January; the MCP connector made it callable from outside. As far as we know, it’s the first AI→AI website in our industry — if you know of another, tell us.

Q: How does a website become MCP-ready?
A: Ship a Model Context Protocol server (JSON-RPC over HTTP) exposing two to four domain tools — ask, search, act — backed by retrieval-grounded answers, rate-limited with machine-readable backpressure (RateLimit-* headers, Retry-After), and advertised through llms.txt, robots directives, meta tags, and /.well-known manifests. Axoworks shipped its three-tool server in one evening on Netlify Edge Functions with zero new infrastructure.

Q: Does llms.txt actually do anything?
A: Ahrefs’ server-log study found 97% of llms.txt files are never read. The 3% that are read are read by exactly the audience that skips your homepage — but treat MCP tools and a crawlable evidence corpus as the load-bearing signals. llms.txt is the front door, not the foundation.

Q: How much does the Axoworks MCP server cost to use?
A: Nothing up front: no authentication, no API key. Each grounded answer costs Axoworks cents of inference, capped at 30 requests per minute per IP. book_consultation is free and routes structured scope notes directly to the partners.

The bottom line

The click was always a proxy — proof that someone, somewhere, engaged. Proxies die when the thing they measure moves. Engagement moved into agents, and the industry is still polishing landing pages for visitors who aren’t coming.

We spent eight months building an agent that could survive humans, and one evening making it legible to machines: installed, not ranked; called, not clicked. Anyone can dump data for agents. Almost nobody governs what the data says about them — because almost nobody has an AI on both sides of the wire.

Browsing is a legacy paradigm. So is the click.