Siteline Finds AI Agents Misread B2B Pricing

Siteline, a startup that analyzes AI agent web traffic, ran a simulated Claude (Sonnet 4.6) agent through 534 attempts across 100 top B2B software products, asking it to find monthly pricing and top features for every publicly listed plan. According to Siteline's report, only 65% of plans surfaced directly readable prices; the rest pushed the agent toward a "contact sales" flow with no listed price. Nearly one in three runs hit at least one access error (bot blocking, broken pages, or unreadable JavaScript-rendered tables), and when that happened the agent pulled 58% of its final answer from third-party sources such as G2, Capterra, or vendor-comparison blogs, compared with just 12% on runs with no errors. Siteline names specific examples: Zendesk's pricing table loaded but was JavaScript-rendered and unreadable, so the agent turned to third-party blogs at roughly 5x the cost of a clean run; Braze's agent could not reach the pricing page at all and pulled numbers from G2 and Vendr instead. Search Engine Journal's coverage of the report adds that the median run took about 32 seconds and $0.24 in model cost, with a roughly 4x cost spread between the fastest and slowest 10% of sites.

Siteline sells agent-analytics and an agent-readiness tool, so it has a commercial interest in positioning "agent readiness" as a problem worth paying to fix -- its findings should be read as a vendor-run benchmark on one model family rather than an independent, peer-reviewed study.

If you operate a B2B site, this points to concrete fixes: render pricing tables server-side rather than client-side, keep a live pricing page even if exact numbers sit behind a sales contact (an agent that finds a dead or non-existent pricing URL is more likely to fall back to outside sources), and front-load key plan details since agents typically only process the first 15,000-20,000 tokens of a page. If you build or deploy agents that shop or compare vendors on a user's behalf, treat any price pulled from a third-party site as lower-confidence and worth a freshness check, since Siteline's data shows that fallback path is disproportionately triggered by the same failures (JS rendering, gating) that also make the data most likely to be stale.

Whether vendors respond by adding server-rendered pricing or machine-readable formats like llms.txt; whether agent providers add JavaScript-rendering capability to reduce this failure mode; and whether other agent-analytics vendors publish comparable benchmarks that corroborate or complicate Siteline's numbers.