The guerrilla war for agent attention

Every optimization playbook so far assumes a human at the end of the funnel — someone who reads a headline, feels a nudge, clicks a button. That assumption is quietly breaking. In 2026 a meaningful and rising share of web traffic is autonomous software acting for a human principal, and a meaningful and rising share of new web content is produced by companies with no humans in the loop at all.

This is the front that SEO vs AIO only gestures at: not just "how do AI search engines cite me," but "what happens when both my competitors and my audience are agents." It is less a tidy discipline than a guerrilla war — fast, asymmetric, fought with unfamiliar weapons.

Combatant one: the autonomous company

A new category of platform spins up and operates an entire business from a prompt.

• NanoCorp (a Y Combinator-backed product from Phospho) lets anyone create an autonomous company run by an agent that, in its own framing, "maximizes revenue while trying to avoid bankruptcy with no human intervention." The platform handles the harness, sandboxes, scheduling, budgets and secrets; you describe the mission and watch the agents work.
• Polsia markets itself as "AI that runs your company while you sleep" — it sets up the website, writes the code, runs marketing and handles support, for roughly $49/month plus a 20% cut of Stripe revenue. Its bold claims have drawn both rapid traction and public skepticism; treat the numbers as self-reported.

Whether or not any single platform lives up to its pitch, the direction is the signal. When launching a site, writing its content and running its marketing collapses into a prompt, web presence gets produced at a scale and cadence no human content team can match. The index floods. The contest for both engine ranking and agent attention intensifies — and "publish more, faster" stops being a moat the moment a competitor rents an autonomous company for $49.

Combatant two: the individual agent

On the demand side, autonomous agents now browse and act for ordinary users.

• OpenClaw (formerly Clawdbot, then Moltbot) is an open-source, self-hosted assistant that lives in your chat apps and controls a real browser via Playwright — navigating, filling forms and extracting data from any site, while also running shell commands and calling APIs. Launched in early 2026, it became one of the fastest-growing open-source repositories in history, drawing hundreds of thousands of GitHub stars within months.
• Hermes Agent, built by Nous Research and released in February 2026, is an open-source (MIT) agent with persistent memory that "lives on your server and gets more capable the longer it runs." It offers full browser automation — navigate, click, type, screenshot — across cloud and local backends.

These agents are not search engines and not chatbots. They are users — software users who arrive with a task, complete it, and leave. They do not watch your hero animation, read your testimonials, or feel your scarcity timer. They parse, act, and move on.

Why "guerrilla"?

The metaphor is precise. This contest is asymmetric (a solo founder with an autonomous company can out-publish a staffed marketing department), fast (agents iterate in minutes, not sprints), and decentralized (open-source agents run on thousands of individual machines, not one controllable platform). The old artillery of human-attention marketing — persuasive copy, visual polish, social proof — mostly misses these targets. Attention here is won in machine-legible structure and low task friction.

For a human, your interface is an experience to be persuaded by. For an agent, it is an API to be completed. Optimize accordingly.

Optimizing for both audiences at once

You still have humans and engines to serve — the fundamentals hold. But add a third lens: can an agent accomplish the task on your site without a human?

Surface	What wins
Engines (Google, AI search)	Crawlable HTML, helpful content, authority — see GEO
Generative answers	Quotable, dated, structured content — see AEO
Acting agents	Server-rendered, labelled controls, predictable flows, exposed endpoints

Make the interface agent-completable

• Server-render the actionable parts. A checkout or form that only exists after client-side hydration is invisible to a retriever and fragile for an automating agent.
• Label controls stably and accessibly. The same practices that help screen readers — real labels, semantic buttons, ARIA where needed — help an agent locate and operate your form. Accessibility is now agent optimization.
• Keep flows predictable. Multi-step funnels with surprise interstitials, modals and re-auth break automation. The fewer the surprises, the higher your agent completion rate.
• Expose structure deliberately. llms.txt, a /.well-known/mcp.json manifest, and clean JSON endpoints let cooperative agents act through a front door instead of scraping the back. (See owning your AI surface.)

The new conversion metric

When an agent books, buys or signs up for its user, the brand whose interface the agent can complete most reliably wins the transaction — regardless of which brand had the better landing page. This is agent conversion rate, and almost nobody is measuring it yet. The early movers will.

The defensive side: governing agent access

The same guerrilla cuts both ways. Autonomous companies and scraping agents can hammer your infrastructure, lift your content, and crowd your category. You decide the terms of engagement:

• Signal your rules in robots.txt with a Content-Signal line — separate "index me for search" from "use me for retrieval" from "train on me."
• Distinguish cooperative agents from abusive ones. A user's Hermes agent completing a purchase is a customer; a scraper cloning your catalog is not. Rate limits and auth should target the second without blocking the first.
• Decide whether to welcome or wall. Blocking all automation is a legitimate choice — but in a world where users delegate to agents, an un-completable site may simply lose the sale to one that isn't.

Strategic takeaways

1. Assume software in your audience. Design at least one path through your site that an agent can complete end to end.
2. Stop treating "publish more" as a moat. Autonomous companies commoditize volume; defensible authority and original data matter more than ever (see the web economy in the age of agents).
3. Treat accessibility as agent optimization. The investment now pays in two currencies.
4. Govern access on purpose. Use Content-Signal and auth to choose which agents you serve and on what terms.
5. Measure agent completion. The metric that will define the next funnel barely exists today — which is exactly why building it is an edge.

Frequently asked questions