The Price of a Page Just Became a Number: Cloudflare Now Lets Any Site Charge AI Agents Per Request

Cloudflare Now Lets Any Site Charge AI Agents Per Request

Cloudflare launched the Monetization Gateway on July 1, 2026, letting any site behind it attach a price to a web page, a dataset, an API or an MCP tool, then collect that price from the agent that requests it. Prices are written as expressions in a dedicated product API. Payment is discovered through an HTTP 402 response then settles in stablecoins in under a second.

The published examples matter more than the product, because they are the first time anyone with distribution has put numbers against machine consumption. Cloudflare's own illustrations run from a $0.001 base fee on an upload endpoint to $0.99 for a resolved support escalation, paid only when the work succeeds. That is a span of nearly three orders of magnitude across resources that all sit on the same domain.

Digital Strategy Force reads that spread as the actual news. The question the field has been asking is how to get cited. The question this launch forces is different, then harder: what is a given surface worth when a machine consumes it, then who decides.

Read the right-hand column carefully. Two of those five prices are not charges for content at all. They are charges for work performed, then one of them pays out only if the work succeeded. That distinction is where the money is, then it is the distinction most site owners are about to get wrong.

The Old Price Was One Number for Every Crawler

Charging machines is not new. What changed on July 1 is granularity, so the baseline is worth stating precisely. Under Pay Per Crawl, Cloudflare's documentation sets the minimum at $0.01 USD per crawl, charged on each successful content retrieval, meaning an HTTP 200 response.

The binding limitation was never the amount. It was that a single default applied to the whole property. A homepage, a decade of archives, a pricing calculator, then a live inventory feed all carried the same number, which meant the number had to be low enough for the cheapest of them. Pricing to your least valuable resource is how a site ends up concluding that machine revenue is not worth pursuing.

There is a second limitation worth naming, because it explains why so few sites ever switched the old system on. A flat rate has to be justified to every buyer at once, so the owner ends up defending the price of an archive page nobody wants against the price of a live feed everyone does. That argument has no good answer, then most teams resolve it by leaving the whole thing at zero. Granularity removes the argument entirely, since each resource now answers for itself.

The last row is the one to sit with. Moving from one number to any number does not produce revenue by itself. It produces a pricing exercise, then the exercise is the work. A capability nobody configures earns exactly what the old flat rate earned, which for most properties was nothing.

The Price Travels in Headers, then the Agent Arrives With a Ceiling

This is not a posted rate that buyers accept. It is a two-sided negotiation conducted entirely in HTTP, which is the single most important thing to understand before setting any number.

On the Cloudflare wire, the origin returns a crawler-price header carrying the cost to access the content, alongside a crawler-charged header stating the exact amount billed. The requesting agent sends its own headers, then one of them is decisive: crawler-max-price, the maximum that agent is configured to pay for any content at all.

A price above that ceiling is not a negotiation the agent loses. It is a request the agent never completes, in the same way that a control the browser cannot name is one the agent never clicks. The underlying mechanism is the x402 Payment Protocol, which uses the long-dormant HTTP 402 Payment Required status code, then carries its terms in base64 headers so that no account, subscription, or API key is involved anywhere.

Step three is the whole commercial reality of this launch. Somebody else has already decided what your page is worth to them, then encoded it as a number their software carries into every request. A price above that number does not earn less revenue. It earns none, then loses the visibility too.

A Price Only Clears If It Beats What the Request Costs the Agent

Most owners will set a price by asking what the content cost to produce, or what the lost advertising was worth. Both anchors produce prices no agent pays, because neither describes anything the buyer feels. The only anchor that clears is the agent's own cost structure.

That structure is unusually favourable to sellers. Research measuring token consumption across eight frontier models on a standard software benchmark found that agentic tasks consume roughly 1000 times more tokens than single-turn chat, with up to 30 times variation between runs on the very same task. Agents are expensive to operate, then their costs are dominated by input rather than output.

Separate work on token economics puts the input dominance at a ratio exceeding 150 to 1, then sizes the demand curve directly: weekly token processing on one platform rose from 0.4 trillion in December 2024 to 27.0 trillion by March 2026, a nearly 68-fold increase in fifteen months.

Fetching is also slow, which is its own kind of cost. Measurements of agent workloads put the mean web fetch at just under thirty seconds across more than eighteen hundred observed calls, inside trajectories whose accumulated context runs to 146,000 tokens at the upper end.

Put those together, then the pricing rule falls out on its own. A resource that saves an agent tokens, turns, or seconds has a defensible price, because the agent is already paying for those things. A resource that simply exists has none, however expensive it was to write.

The Reserve Price Model Prices Four Kinds of Work, Not Four Kinds of Bot

The DSF Agent Reserve Price Model is a four-tier method for setting the minimum an owner should accept for machine access, tiered by the kind of work the resource performs rather than by the identity of the caller. Tier one is the Reference Read, tier two the Bulk Ingest, tier three the Computed Endpoint, then tier four the Completed Outcome.

Classifying by caller is the intuitive move, then it is the wrong one, because the same agent wants different things at different moments. Classifying by work survives that, since a computed answer is worth the same whoever asks for it.

A page is worth what it saves the machine reading it, never what it cost to write.— DSF Commerce Engineering Division

The rule governing the model is the Deliberate Zero Principle: for most sites the correct reserve price on tier one is exactly zero, chosen on purpose because the citation is worth more than the toll, rather than defaulted into because nobody ever priced it. A zero you selected is a strategy. A zero you inherited is an oversight.

Nearly every commercial site on the web is entirely tier one, which is why the first honest finding of any pricing exercise is that the existing property is worth very little to machines. That is not a reason to abandon the exercise. It is the reason to run it, because tiers three then four are things a business builds rather than things it already has, which is the same argument for exposing a purpose-built interface to agents instead of hoping they parse the pages you already had.

Want to Know Which of Your Surfaces an Agent Would Actually Pay For? Digital Strategy Force tiers your public surfaces against the model, so the pricing decision starts from what the work is worth rather than from what the page cost.

Charging Agents Costs You Something, then That Cost Has to Be Priced Too

The case against pricing is stronger than most coverage of this launch admits, then it deserves stating before any answer. Three costs are real.

First, a price is a wall, then walls forfeit surfaces. A resource behind a price is not freely retrievable, so it cannot be summarised or cited. That is a Crawl Gate in the strict sense, a boundary a machine meets before any content is served. This is not new territory: Microsoft published controls in September 2023 letting publishers keep content out of chat answers, then every publisher who used them traded away the surface to gain the control.

Second, the volumes are large but the unit values are tiny. Cloudflare has reported that 32 percent of traffic across its network is automated, generating over ten billion AI bot requests per week. At a penny per retrieval that sounds enormous, then divided across every site on the network it is rounding error for most of them.

Third, some paths must stay open regardless. Cloudflare's documentation names five permanently free paths, including robots.txt and sitemap.xml, so the policy itself stays discoverable to the agents being charged.

The resolution is not to price everything, then it is certainly not to price nothing. It is to price the tiers where the agent is buying work, then leave the reference layer open so the citations keep arriving. Selective pricing is the only version of this that does not trade visibility for pennies.

Declaring the Wall Is What Keeps a Priced Site Inside the Rules

Pricing creates a compliance problem most owners will not anticipate, because serving one version to a paying agent then another to a crawler is the textbook definition of cloaking. The fix is declaration, then the markup for it already exists.

Google's documentation on paywalled content states that the structured data exists precisely to help it differentiate paywalled content from the practice of cloaking. The mechanism is an isAccessibleForFree flag, then a hasPart block naming the CSS selector of the gated region, so a machine can tell exactly which portion is behind the wall.

There is also a precedent for metering rather than walling. Google's flexible sampling guidance recommends a monthly allowance, suggesting a value between 6 and 10 articles per user per month for most daily news publishers. The same shape works for agents: a free allowance that establishes value, then a price beyond it.

The consequence of skipping declaration is not a warning. It is that the crawler which indexes a page for ordinary search sees the free version, while the agent that pays sees a fuller one, then no machine anywhere can tell those two views apart from a site that is deliberately serving different content to different visitors. Declaring the boundary converts an undeclared discrepancy into a stated commercial policy, which is the entire difference between a paywall and a penalty.

Declaration is cheap then it is not optional. A site that prices machine access without declaring it has not built a business model. It has built an undeclared difference between what people see then what machines get, which is the thing search engines have penalised for two decades.

The Seller Writes the Price, then the Intermediary Owns the Rails

The headline question is who sets the price. The answer is best read from the neighbouring standards rather than asserted, because four different rails are being built at once, then they do not agree on who holds the pen.

On Cloudflare's rail the site owner authors the number, then Cloudflare acts as Merchant of Record while also providing settlement. On Stripe's, a Shared Payment Token bounds the agent by a maximum amount then an expiry, so the buyer's envelope is fixed before the seller quotes anything. On Google's, the AP2 Payment Mandate is created by the shopping agent, then merely displayed to the user for approval.

Read the second column, then the third. In three of four rails the seller writes a number inside an envelope somebody else defined, which is the honest answer to the headline question. Sellers set prices. Intermediaries set the range prices may occupy, then own the settlement that makes any of it enforceable.

That is a smaller victory than the launch language suggests, then it is still the first real one. For thirty years the price of a page to a machine was fixed at zero by nobody's decision. It is now a number, written by the owner, inside limits the owner did not choose. The businesses that will do well here are the ones that stop asking how to be found for free, then start asking which of their surfaces do work a machine would rather buy than reproduce.

For most companies the honest first answer is none of them, which is uncomfortable but useful. It says the property was built for readers who arrive, look, then decide. The next one has to be built for callers who arrive, take exactly what they need, then leave a payment behind. Those are different buildings, and the second one has a price list.

FAQ — Charging AI Agents

Next Steps — Charging AI Agents

• ▶ Inventory every public resource, then sort it into the four tiers: reference reads, bulk ingest paths, computed endpoints, completed outcomes.
• ▶ Leave tier one open at a deliberate zero, so the pages that earn citations stay retrievable while pricing is tested elsewhere.
• ▶ Price one computed endpoint first, on a compute basis with a published ceiling, committed as code rather than set by hand in a dashboard.
• ▶ Declare every priced region in markup before it goes live, so a paid resource is never mistaken for cloaking.
• ▶ Confirm the permanently free paths stay open, so the pricing policy itself remains discoverable to the agents being charged.

Digital Strategy Force tiers your public surfaces against the Reserve Price Model, then builds the ones worth charging for. Talk to the answer engine optimization team.