You passed the agent-readiness score. Agents still bail at your pricing page

Published July 24, 2026 · Last reviewed July 24, 2026

You ran the audit. You cleaned up robots.txt, added structured data, maybe published an llms.txt, and the readiness score turned green. So why is agent-driven traffic still not turning into conversations, and why, when you ask a browsing agent what your product costs, does it sometimes answer with a number from a directory you have never heard of?

The readiness score answered a real question. It just was not the question that closes a deal. A score tells you an agent can find and read your site. It does not tell you the agent got the answer it came for, and it says nothing about what happens next, because on most sites nothing happens next. The agent reads, extracts what it can, and leaves.

This piece is about the layer above the score. We will look at what a readiness score actually grades, at the July 2026 data that measured exactly where agents fall down, and at the practical work that turns a passing grade into conversations you can see.

What an agent-readiness score measures, and where it stops

An agent-readiness score grades the readable layer of your site. Cloudflare Agent Readiness, at isitagentready.com, scores four buckets: Discoverability, Content, Bot Access Control, and Capabilities.¹ Chrome shipped a Lighthouse audit that checks similar things, and community specs like AgentReady.org grade the same territory. All of them are asking one family of questions: can an agent discover your pages, fetch them without being blocked, and parse the content and any structured data you expose.

That work is necessary, and none of it is wasted. If an agent cannot reach or read your site, nothing downstream matters. If you have not done this pass yet, the agent-readiness audit checklist is the place to start, and it is honest about what the tools catch and what they miss.

The line where a score stops is the important part. A readiness score can confirm that your pricing page is fetchable and that its markup parses. It cannot confirm that an agent found the price, or that the price was in a form the agent could extract, or that there was anyone on your side to ask when the answer was ambiguous. Readable is measured. Whether the agent actually got its answer, and whether it could do anything after reading, are not.

The July 2026 data: agents retrieve your content, then bail at pricing

In July 2026, Kevin Indig teamed up with David Kaufman, founder of Siteline, to measure where agents actually get stuck. They ran 1,500 agent runs across 100 B2B products, five runs each on three buyer tasks, and scored how often an agent could answer the question from the first-party site itself.² The pattern is the point. Integrations answered first-party 93 percent of the time. Security answered 92 percent. Pricing and features came last, at 79 percent, the weakest category in the study.

The failure mode matters more than the number. Among pricing runs where the vendor disclosed no real price, agents pulled in at least one third-party source 45 percent of the time, against 18 percent even when a numeric price was sitting on the page.³ Read that plainly. On the most commercial question a buyer asks, a missing price makes an agent roughly two and a half times as likely to walk off your site and repeat what a directory or comparison blog says, instead of the number you would have given it. Put the other way: leaving the price off the page is not a neutral omission, it is a decision to let someone else answer. And of those third-party citations, nearly half pointed to software directories like G2, Capterra, and Vendr, listings the vendor does not control and cannot correct.

Indig and Kaufman name three root causes, and they are worth memorizing. Opacity: there is no clear price disclosed anywhere an agent can find. Poor machine-readability: a price exists, but it is trapped in an image, a PDF, or JavaScript the agent does not execute. Access friction: the page fails to fetch, or a bot rule blocks the agent at the door. The uncomfortable part is that two of those three, machine-readability and access, are exactly the things a readiness score rewards you for fixing on the rest of the site. You can pass the score and still hand your pricing answer to someone else.

Readable is not answerable, and neither one is addressable

It helps to separate three things a site can be, because a readiness score collapses them into one green check. A page can be readable, meaning an agent can parse it. A page can be answerable, meaning the specific answer the agent needs is actually present, in text, in a form the agent extracts. And a site can be addressable, meaning there is a representative on your side that another agent can reach and talk to.

The Siteline data is a story about the gap between the first two. Those pricing pages were readable, they scored well, and the agent still left without the answer, because the answer was opaque, buried in an image, or blocked. Fixing that is content and access work, and it is the cheapest win here.

The third gap is the one no score touches. Even a perfectly answerable page is passive. The agent reads the number, and if it has a follow-up, whether the price includes onboarding, whether there is a plan for a team of five, whether you serve its buyer’s industry, there is no one to ask. It guesses, or it leaves. The distinction between a site an agent can read and a site an agent can address is the whole subject of agent-readable versus agent-addressable, and it is the seam a readiness score is not built to grade.

The layer above the score: answer, qualify, route

Passing the score got the agent in the door. Three moves keep it in the room, and they map directly onto the gaps above.

First, make the commercial answers extractable. Go straight at the three Siteline causes. If you publish a price, put it in plain HTML text, not in a rendered image, a gated PDF, or a number that only appears after JavaScript runs. If your pricing is genuinely custom, do not leave a bare “contact us” for an agent to choke on: state the model in text, per seat, usage-based, tiered by company size, so an agent can represent how you charge even when it cannot quote a final figure. Opacity is a choice, and for an agent it reads as an answer you did not give.

Second, stop losing the follow-up. An agent that got a partial answer does one of three things: it guesses, it cites a third party, or it leaves. An endpoint it can address, ask, and get a qualified answer from keeps that exchange on your property and lets you shape it. This is the same lesson the contact-form conversion data has been teaching about human visitors, now with a visitor that will not fill in a form at all, because an agent cannot and should not fake a phone number it does not own.

Third, remove delay. The buyer’s agent is comparing you against alternatives inside a single session, and it will not wait for a callback tomorrow. The collapse of the response clock from hours to the length of one visit is covered in speed to lead is dead. If the answer and the qualification do not happen during the visit, they may as well not exist.

A this-week checklist after you pass

You do not need a project to close most of this gap. You need an afternoon and a willingness to look at your own site the way an agent does.

1. Fetch your own pricing page as raw HTML, without running scripts, and search the response for the actual price. If it is not there in text, an agent probably cannot see it either.
2. Grep for prices trapped in images, PDFs, or script-injected elements. Anything a screen reader would miss, an agent misses too.
3. Ask a real browsing agent, in ChatGPT or Claude, what your product costs and who it is for. Watch whether it answers from your site or cites a directory. That single test reproduces the Siteline finding on your own domain.
4. Check that your pricing and contact pages are not caught by a bot rule. Access friction blocks the exact traffic you want here, so make the allow-list deliberate.
5. Make sure the answer carries qualification cues, not just a number. Who the product is for, what size of team, what it does not do. An agent routing a buyer needs fit signals, not only price.
6. Add an addressable endpoint that can take a follow-up question and route a good-fit conversation to a human. This is the move that turns a read into a relationship.

Where an addressable, networked endpoint fits

Here is the honest boundary, because it is easy to overclaim. Agent-readiness is about how agents read your site, and the scorers measure that layer well. Making a site addressable does not replace that work, and it does not make your site agent-readable: the readable layer and the addressable layer are two different jobs, and you want both.

The addressable half is a representative that agents can find and talk to, tied to a human name, so a real person is reached only with consent. That is the layer Tobira works on: a @handle and a Site Agent give a company an addressable, networked presence on an agent-to-agent network, the human-facing trust layer for the agentic web, complementary to the readable site a score grades. A Site Agent can answer the follow-up, qualify the fit, and route the buyer, which is precisely the gap the Siteline data exposed. Tobira is free during beta, with a paid tier planned, and it is one way to close the third gap, not a substitute for fixing the first two.

Whatever you use to do it, the shape of the work is the same. A passing readiness score means agents can read you. The next layer is making sure they get the answer, and that someone is there when they have the next question.

What to remember

• An agent-readiness score grades the readable layer: discovery, fetch, parse. It is necessary and it is not the finish line.
• July 2026 Siteline data: agents answered pricing questions first-party only 79 percent of the time, the weakest category; when no price was disclosed, 45 percent of those runs pulled in a third-party source, versus 18 percent when a price was shown.
• Two of the three root causes, poor machine-readability and access friction, are things you can pass a general readiness score on and still fail at the pricing page.
• Separate three states: readable (parseable), answerable (the answer is actually there), addressable (something on your side can converse, qualify, and route). Scores test only the first.
• The work after the score: make commercial answers extractable, add an endpoint that answers follow-ups and qualifies, and remove delay.

FAQ

What does an agent-readiness score actually measure? It grades the readable layer of your site: whether an AI agent can discover your pages, fetch them without being blocked, and parse the content and structured data. Cloudflare Agent Readiness, for example, scores four buckets: Discoverability, Content, Bot Access Control, and Capabilities. These are real and worth fixing. What a score does not test is whether the agent found the specific commercial answer it came for, such as your price, or whether anything on your side could answer a follow-up. Readable is measured; answerable and addressable are not.

Why do AI agents fail at pricing pages even on agent-ready sites? A July 2026 study by Kevin Indig and David Kaufman of Siteline tested 100 B2B products across 1,500 agent runs. Agents answered pricing and features questions from the first-party site only 79 percent of the time, the weakest category, against 93 percent for integrations and 92 percent for security. The three root causes were opacity, meaning no clear price is disclosed, poor machine-readability, meaning the price is trapped in an image, a PDF, or JavaScript, and access friction, meaning fetch failures or bot blocks. Two of those three are things you can pass a readiness score on elsewhere and still fail at the pricing page.

What is the difference between agent-readable and agent-addressable? Agent-readable means an agent can parse your site and, at best, extract an answer already sitting on the page. It is passive: the agent reads and leaves. Agent-addressable means there is a representative on your side that other agents can find and talk to, that can answer a follow-up, qualify whether the buyer is a fit, and route a useful conversation to a human, with the real person reached only after consent. A readiness score tests the readable layer. The addressable layer is not on its rubric.

What should I do after I pass the agent-readiness score? Three things. First, make your commercial answers extractable: put a real price or a plain-text pricing model in HTML, not in an image, a PDF, or a contact-us wall, and confirm the page is not bot-blocked. Second, add an endpoint that can answer follow-up questions and qualify fit, so an agent that got a partial answer does not guess, cite a competitor, or leave. Third, remove delay, because a buyer agent will not wait for a callback. The score got the agent in the door; these steps keep it in the room.

Does a Site Agent make my website agent-ready? No, and it is important to keep the two separate. Agent-readiness is several site-layer changes, such as clean HTML, structured data, and bot access rules, that make your site readable to agents. A Site Agent does not do that work. What it adds is the other half: an addressable, networked representative that agents can find and talk to, tied to a human identity, so the readable site and the addressable endpoint are complementary, not substitutes.