AI is the customer

In early June, Cloudflare’s CEO posted that Radar had crossed a line he thought would wait until 2027: on their network, bots were making about 57% of the HTTP requests to HTML, humans about 43%. He said the classification is messy. It is. Different firms count “bot” differently, and a request is not the same thing as a person spending twenty minutes with your essay.

Still. The plumbing of the public web is now majority non-human on a very large sample. Imperva’s 2026 bot report put automated traffic over 53% of what they saw in 2025. HUMAN Security said automated traffic grew about eight times faster than human traffic, and that traffic from agents and agentic browsers grew 7,851% year over year — off a small 2024 base, which is the honest caveat, and still a curve you should not ignore.

If your site only makes money when a person sees an ad, joins a membership, or clicks an affiliate link, a growing share of demand never walks through that door. Some of it will steal. Some of it will train. Some of it will pay, if you give it a price and a checkout it can finish without a mouse.

That last group is the point of this piece. AI is the customer — not instead of humans, alongside them, and already ahead of them in raw request volume.

Not every bot is a customer

“Bot traffic” is a junk drawer. Treat it that way.

Training crawlers pull text into weights. You license, block, or negotiate. They are not going to tap a paywall. Search crawlers still matter for humans who use Google or an AI answer box; that is SEO plus whatever people are now calling GEO. Scrapers want the whole corpus, free, at machine speed. Rate-limit them, hard-lock the good stuff, keep a lawyer for the rest.

Then there are agents: software with a job. Research this. Compare those three posts. Fetch the number and put it in the memo. Sometimes, buy. HUMAN’s report put a few percent of agentic activity on checkout paths. That is a small slice of a fast-growing pile. It is also proof that the client on the wire is no longer only a spectator.

A lot of “AI traffic” is a person asking an assistant, and the assistant becoming the HTTP client. Your analytics still say “user.” The bytes were consumed by a script. The popup, the hero, the pixel — all built for a pair of eyes that never loaded the CSS.

Why the old rails miss them

Ads need a viewable impression. Affiliates need a click and a cookie. Memberships need a card and a reason to come back next month. Lead forms need someone who types. Agents are bad at all of that, and they are not trying to get better at it. They are trying to finish a task on a budget.

A soft paywall is worse. If the full HTML is already in the response, you did not sell an article. You decorated one. View Source is the unpaid checkout.

So you can have a traffic spike that looks like success and a revenue line that does not move — unless there is a fourth rail: a price on the URL, a challenge the client can parse, and a body you do not ship until it pays.

That is HTTP 402 and x402. The agent gets a payment required response with terms it understands, pays in USDC, retries the same URL, and gets the page. No account. No “talk to sales.” No monthly relationship it does not want.

What this looks like on a real site

Keep the human path. People still read, subscribe, and care about your voice. They unlock in the browser with a wallet — MetaMask on Base, Phantom on Solana — and a few cents of USDC. We never hold that money. It settles to the payout address you set.

Agents hit the same URL and should not get a modal. They should get 402, pay, retry.

Two lock styles, and they are not interchangeable:

Easy blur is the fast install. The premium HTML is in the page and CSS hides it until payment. Fine for a teaser and a human who will not View Source. Useless against a script.

Hard lock is the one that matters for machines. Your origin only sends a teaser until an unlock checks out. WordPress, Next.js, PHP, nginx — something dynamic. A static file cannot hard-lock. If you only do one thing for agents, do this.

A URL worth selling, then, has a few boring properties: it does not move around; unpaid and paid are different responses; 402 carries the price and the networks; the secret body is not in the unpaid HTML; settlement does not need a person in Slack.

You do not have to price the whole site. Start with the pieces that actually cost you to produce — the research post, the dataset, the playbook. Curl it as a bot. Confirm 402. Pay once. Confirm the body.

How agents find you: ARD and Bazaar

A 402 on a URL nobody discovers is a locked door in an empty building. Discovery is two layers, and they are easy to mash together. Don’t.

ARD (Agent Resource Directory) is the catalog on your domain. smidgen builds it from the paywall rules you enable and you serve it at /.well-known/ai-catalog.json. Each paid path is an entry with a price and the fact that the URL speaks x402. Docs stay marked free. Crawlers find the file because robots.txt includes an Agentmap line. WordPress can publish this when you sync rules. You control it. If you take a post down, the catalog should stop advertising it. Live example: smidgen.io/.well-known/ai-catalog.json.

Coinbase Bazaar is not your site. It is Coinbase’s public directory of x402-paid resources. Agents search it by keyword or URL. You do not submit a form. smidgen attaches Bazaar metadata to the 402 and to the settle, so the publisher URL can be listed — not some internal gate. The first real payment on that URL is what can get it indexed. Our Next.js and WordPress demos are in there today; start from Pay as an agent or search Bazaar for smidgen.

ARD is the shelf in your shop. Bazaar is the mall directory. You want both. The catalog without a settle is incomplete; a settle without a catalog means only the agents who already had the link.

What to do this month

Stop treating “AI” as one policy. Abuse gets blocked. Training gets a license or a wall. Task-time agents get a price.

If your dashboard only says Users, you cannot price this. Watch bot vs human at the edge if you have it, 402s on premium paths, and unlocks by client class.

Pick one flagship URL. Hard-lock it. Put a small USDC price on it. Verify 402. Publish the catalog. Make a payment so Bazaar has a chance to see it. Leave it two weeks. Read the logs.

Leave a free layer that explains who you are. Agents and answer engines still need a page they can cite before anyone pays. That is not generosity. It is the aisle that leads to the register.

You can still run a membership and a newsletter. Those are human strengths. The agent path is for the work you already publish on the open web and currently give away to anything with a User-Agent.

Pushback

My niche is too small. Small niches are where someone pays an agent to save an hour. Scarce expert prose is easier to price than a commodity recap.

They will use the free version. They will, until the free version is wrong or missing. Quality plus a machine checkout is how you win the query that actually matters.

I will wait until wallets feel like Apple Pay. Human wallet UX is getting less ugly. Agent wallets are already software. Waiting is how you train someone else’s catalog.

I only want a community of people. Keep it. Add a paid agent door for the public posts. Those are different rooms.

The smidgen version of this

We built a non-custodial paywall for that loop: hard lock when the origin can do it, USDC on Base and Solana, 402 / x402 for the client that does not have a browser chrome, a catalog on your domain, Bazaar metadata on settle. Humans still pay in the wallet modal. Agents pay the same URL.

Prove it on the demos, then put the same thing on your WordPress or Next.js origin. You do not have to like every bot. You can still invoice the ones that create value.

If an AI is going to install this for you, that is a different agent — a coding agent, after you create the account. How that works. Paying agents never need a smidgen login.

FAQ

Did bots actually pass humans? On Cloudflare’s HTML-request sample in June 2026, yes, by their ruler. Imperva’s 2025 mix was the same direction. Always read the methodology. None of this measures time-on-site.

Should I block all AI traffic? Only if you like the sound of zero upside. Segment: block abuse, negotiate training, sell the task-time fetch.

How does an agent pay? It hits the URL, gets 402 with x402 terms, pays USDC, retries. Same link a person would open. The person sees a paywall UI instead of the raw challenge.

What is the difference between the catalog and Bazaar? The catalog is yours, at /.well-known/ai-catalog.json. Bazaar is Coinbase’s index. First successful settle can list the publisher URL. You want the file on your origin either way.

Sources: Cloudflare Radar, bot vs human; Matthew Prince, June 2026; Imperva Bad Bot Report 2026; HUMAN, 2026 State of AI Traffic.