An AI agent sent to check the price of a piece of B2B software now fails to read it straight from the vendor on a meaningful share of attempts — and when it fails, it goes looking elsewhere. A benchmark published on 18 June by Siteline, an agent-analytics startup, ran a simulated Claude agent through 534 attempts to find the monthly pricing and headline features of 100 leading B2B products, and found the agent hit at least one search-or-fetch error on roughly a third of runs. On the runs where the agent was blocked, most of the answer came from third-party pages rather than the company’s own site.

The finding lands at an awkward moment for marketing teams. Buyers are increasingly delegating the early legwork of software research — shortlisting, comparing, estimating cost — to AI assistants, and pricing sits at the bottom of that funnel, where the decision to keep or cut a vendor is made. If an agent cannot extract a price from the source, it does not stop; it cites whatever it can reach, which is often a directory, a review site or a competitor’s comparison page. The vendor is still named, but the pricing story is no longer its own.

What did the Siteline benchmark find?

Siteline’s benchmark tested 20 products in each of five categories — productivity and collaboration, developer tools, marketing and sales, customer support, and analytics and data — measuring whether a Claude agent could reach each site, retrieve its plans and prices, and how much the task cost in tokens and tool calls. The task was deliberately narrow: find the monthly price of every publicly listed plan and list each plan’s main features, the kind of query a buyer runs just before a shortlist tightens.

According to the report, about 30 per cent of runs carried at least one search-or-fetch error, and roughly a quarter of those errors were outright access denials — bot blocking or pages the agent could not read. Most runs recovered on a retry, but the cost of the ones that did not was steep: on runs that hit an access error, 58 per cent of the final answer was pulled from third-party sources, against just 12 per cent on error-free runs. The company also reports that only 65 per cent of plans surfaced a directly readable price, while 14 per cent showed no price at all and routed the agent toward a sales contact.

The spread between good and bad sites was wide. Siteline puts the median run at 32 seconds, three search-or-fetch calls and $0.24 on Claude Sonnet 4.6, but says the top 10 per cent of products let the agent finish roughly twice as fast and at under a quarter of the cost of the bottom 10 per cent — a gap the report attributes mainly to access friction rather than brand complexity. (The dollar figures are the model’s per-run compute cost, reported in USD.)

Why do pricing pages break for AI agents?

Siteline traces most of the failures to two causes, and they are not the same problem even though they produce the same result. The first is deliberate: many B2B companies gate pricing behind a “contact sales” button on purpose, a long-standing tactic to route prospects into a sales conversation. The second is accidental — pricing that is visible to a human but invisible to a fetcher, because it loads through JavaScript, sits inside a calculator widget, or is baked into an image or PDF. Many AI agents do not execute JavaScript the way a browser does, so a pricing grid that renders cleanly on screen can arrive at the agent as fragments, or not at all.

The report names examples on both sides. Zendesk’s pricing page loaded, Siteline says, but the plan grid and feature detail were JavaScript-rendered and unreadable, so the agent fell back to third-party blogs at roughly five times the cost of a clean run. Braze’s agent, by the report’s account, could not reach the pricing page at all and pulled numbers from the directories G2 and Vendr instead. Linear sat at the other end: the agent found the pricing page, fetched it once, and parsed four plans with pricing and features for about $0.11, without touching an outside source.

A broader version of the study, run with the SEO analyst Kevin Indig and published in his Growth Memo newsletter in mid-July, widened the test to three buyer tasks — pricing and features, integrations, and security and compliance — across 1,500 agent runs. Most sites did well: integrations returned a first-party answer 93 per cent of the time and security 92 per cent. Pricing was the outlier, at 79 per cent. Even when a vendor published a real numeric price, the pair found, agents still cited at least one third-party source 18 per cent of the time; where no price was disclosed, that rose to 45 per cent.

What can marketing teams do about it?

The fixes Siteline recommends are technical, specific and — for a marketing team that can brief a developer — well within reach. Render pricing server-side rather than client-side, so the numbers sit in the raw HTML an agent fetches. Keep a live pricing page even when the real figure sits behind a sales conversation, because an agent that hits a dead or missing pricing URL is more likely to wander off to outside sources. Put plan names, prices, limits and features in crawlable text on one canonical URL, explain usage-based pricing in words rather than a calculator alone, and add Product and Offer schema. Front-load the important detail, too: the report notes agents often process only the first 15,000 to 20,000 tokens of a page.

The caveat worth holding on to is the source. Siteline sells agent-analytics software and an “AI agent readiness” tool, and says it processes more than three million agent requests a day, so it has a direct commercial interest in the finding that many sites are not ready. The benchmark is a vendor-run test on a single model family, not an independent, peer-reviewed study, and its numbers should be read in that light. What is harder to argue with is the direction of travel: agentic research is a growing slice of B2B buying, and a page that persuades a human can still fail the machine sent to read it.

For marketers, the reframing is the useful part. Pricing-page performance has quietly become a retrieval problem as much as a design one, and the same discipline that makes a page easy for an AI answer to cite — clean HTML, clear structure, front-loaded facts — is what keeps a vendor’s own numbers in the buyer’s research. It is a solvable, unglamorous piece of technical hygiene, not a threat to the profession.

Whether the fix settles into a shared standard is the open question. Proposals for machine-readable pricing files, echoing the emerging convention of an llms.txt at a site’s root, are already circulating; none has been adopted at scale. Until one is, the readable pricing page remains the cheapest defence a B2B brand has against having its prices told by someone else.