When an AI Agent Can’t Read Your Price, It Quotes G2

A buyer already knows your product. They send an AI agent to check the pricing and features before booking a call. The agent goes to your site, tries to read your pricing page, and can’t. So it does the next logical thing. It goes and finds your prices somewhere else, on a review site or a competitor’s comparison blog, and answers the buyer using numbers you didn’t write and might not agree with.

That scenario is more common than most B2B teams realize, and a new study puts numbers on it.

What the study found

The data comes from Siteline, a company that tracks AI agent traffic, and it was written up this month by Kevin Indig in Search Engine Land. Siteline ran a simulated Claude agent through 534 attempts to find pricing across 100 top B2B software products. The pattern was consistent. Only about 65% of plans had prices the agent could read directly. Another 14% listed no price at all and pushed the agent toward a contact-sales form.

The interesting number is what happened when the agent got stuck. Roughly a third of runs hit some kind of access error, a blocked bot, a broken page, or a pricing table the agent couldn’t parse. On those runs, the agent pulled 58% of its final answer from third-party sources, against 12% on the runs that went smoothly. So a page it couldn’t read didn’t stop the agent. It sent the agent to G2, Capterra, and vendor comparison blogs instead, and those became the source of truth.

Two named examples made it concrete. One well-known vendor had a pricing table that loaded fine for a human but was built in JavaScript the agent couldn’t read, so it turned to third-party blogs at roughly five times the cost of a clean run. Another vendor’s pricing page couldn’t be reached at all, and the agent pulled its numbers from a review site and a procurement platform.

Why the page keeps breaking

Indig’s write-up traced the failures to three root causes, and none of them are exotic. The first is opacity, simply not showing a price and routing everyone to sales. The second is poor machine-readability, where the price technically exists on the page but is locked inside JavaScript, an image, or a PDF that the agent doesn’t execute or open. One thing to remember, the agents behind Claude and ChatGPT don’t run JavaScript the way Google does, so a table that renders in a browser can come back empty to them. The third is access friction, where bot blocking or a slow, oversized page makes the fetch fail before it starts.

The technical fixes are well understood. Serve prices in plain server-rendered HTML rather than client-side JavaScript. Use real table markup so plan names map cleanly to prices. Add Product and Offer structured data with the actual price and currency. Keep a pricing page reachable even if the real numbers wait behind a conversation, so the agent has something to read instead of guessing. None of that is groundbreaking, and that’s rather the point. All of it is old-fashioned technical hygiene, applied to a new kind of visitor.

The part that’s really about authority

Here the problem stops being purely technical. Fixing your pricing page controls what the agent reads on your site. It does nothing about what the agent reads everywhere else, and everywhere else is where it goes the moment your page falls short.

When an agent gives up on your site and turns to G2 or a comparison blog, the answer your buyer gets is only as good as those third-party sources. If the review sites list stale pricing, if a competitor’s comparison frames you unfavorably, if the top-ranking roundup for your category barely mentions you, that becomes the version of your product the agent presents. You can build a flawless pricing page and still lose the narrative on someone else’s.

So the full defense has two halves. Make your own pricing readable to a machine, and shape what the credible third-party sources say about you, because the agent treats those as a fallback the instant it needs one. That second half is familiar work. Earning accurate, favorable coverage on the review platforms and comparison sites that AI agents trust is what digital PR has always done, and it decides which outside sources speak for you when the agent can’t read your own.

A couple of caveats belong on this. Siteline sells agent-readiness tooling, so it has a stake in the finding that sites aren’t ready, which is reason to read the numbers as directional rather than gospel. And an unreadable pricing page doesn’t directly lower your Google rankings. What it changes is quieter and possibly more expensive, whether your own numbers or a stranger’s represent you at the exact moment a buyer is deciding.