Learn
Agent readiness score vs purchase attempt
An agent readiness score grades what a storefront publishes about itself, and a purchase attempt grades what happens when something tries to buy. The first reads robots rules, structured data, feeds and policies without pressing anything. The second opens a product, chooses an option, adds it to the cart and drives the checkout to the payment step, and it reports the step where the agent stopped.
What a readiness score reads
A readiness score is a signal check. It fetches the things a storefront publishes about itself and grades how complete they are. That is genuinely useful work: it is cheap, it runs in seconds, and it covers the part of the journey where a store can be invisible without knowing it.
Cloudflare’s free Agent Readiness score, at isitagentready.com, scores four dimensions: Discoverability, Content, Bot Access Control and Capabilities. Every one of them is something the site states, which is why they can all be graded from outside.
- Whether robots.txt allows an agent to fetch the catalogue at all.
- Whether there is a product feed and a sitemap to read it from.
- Whether product pages carry structured data with a price, an availability state and identifiers.
- Whether shipping and returns terms exist as text rather than as a graphic.
- Whether a bot wall returns a challenge instead of a storefront.
The honest limit of a signal check is that it never presses anything. It grades the store’s description of itself, which is a claim, and it has no way to test whether the claim survives contact with a checkout.
What a purchase attempt drives
An attempt is the other half. An agent takes a real product on a real storefront and works through six steps in order, and any one of them can end the journey on its own:
- Open the product
- Choose an option
- Add to cart
- Open the cart
- Start checkout
- Reach the payment step
The attempt stops at the payment step and never submits payment. Reaching that step is the measurement; going further would mean putting a real order into a real merchant’s system. What comes back is not a grade but a location: the step the agent reached, and the stall that stopped it going further.
An attempt has an honest limit too. It is one journey against one product at one moment, so it can miss a fault that only shows on a different variant or under a different load. What it cannot do is be wrong about the failure it did observe.
What lives in the gap
The gap between the two is not academic. It has a membership, and the membership is roughly the buy stage of the journey: failures that leave no trace in what a store publishes, because they only happen while something is interacting with the page.
- A control an agent needs is not exposed as one
- Agent could not reach the payment step
- Checkout requires creating an account
- The cart can only be paid with an express wallet
- CAPTCHA stops the agent at checkout
- Checkout returned an error
- Size or colour can only be chosen by hovering
- Choosing a variant needs JavaScript the agent cannot run
- A popup blocks the path to the cart
- Add to cart needs JavaScript the agent cannot run
- The store refused the agent’s add to cart
- An age check asks for a date of birth
- A simpler agent stopped at a step an AI agent got past
- An agent can buy only the option already selected
- A popup covers a button on the way to checkout
- A cookie banner covers the page until it is answered
- The first press of checkout does nothing
- One item cannot be checked out below your minimum order
- Nothing we picked could be bought
A store can publish complete, accurate, well marked up product data and still be unable to take an agent’s order, and every one of the failures above is invisible to a check that only reads pages.
The gap has a published version
You do not have to take our word for the boundary, because the scoring tools state it themselves. Writing about its own Agent Readiness score, Cloudflare says:
Additionally, we check if the site supports agentic commerce standards including x402, Universal Commerce Protocol, and Agentic Commerce Protocol, but these do not currently count towards the score.
That is a careful and correct thing to say. Commerce capability is detected, and it is not scored, because a signal check can see that a standard is advertised but cannot see whether a purchase completes. The transaction is the part that has to be attempted.
What to do with a good score
Keep it. It is worth having, and losing it is easy: a theme update drops the structured data, a security setting starts challenging agents, a feed stops regenerating. Then treat the score as the start of the list rather than the end of it.
- Confirm the discover and evaluate stages, which is what the score already told you.
- Have something attempt a purchase and record where it stopped. That is the buy stage, and it is the heaviest weighted part of the journey for a reason.
- Fix by stall rather than by score. A named failure with a page and a change to make is a ticket a developer can close.
- Re-run after the fix. A stall that is gone should be gone in a measurement, not in a belief.
Questions
Is an agent readiness score worth running?
Yes, and it is free, so there is no reason not to. It covers the discover stage properly, which is the stage where a store can be invisible without anyone noticing. Treat the result as a passed first round rather than as a clean bill of health.
Why can a signal check not tell whether checkout works?
Because the evidence only exists while something is trying. A variant picker that cannot be activated without a mouse looks identical in the page source to one that can. A checkout that errors on the third step publishes nothing about that error. There is no signal to read, so nothing is graded.
Does a good score mean agents can buy from me?
It means an agent can find you and read you. Those are real and necessary, and they are the first two of four stages. Whether an agent can complete a purchase is a separate question, and the only way to answer it is to have one attempt the purchase.
Does Ottom submit payment during an attempt?
No. The attempt stops at the payment step, which is the thing being measured. Going further would put a real order into a real system. The agents identify themselves with a published user agent from a published set of addresses, so the traffic is recognisable in your logs.
Should I run both?
They answer different questions and cost different amounts, so yes. A signal check is cheap enough to run continuously and catches the whole class of problems that are visible from outside. An attempt is what tells you whether the money can actually arrive.