AI Agents Struggle To Read B2B Pricing, Report Finds via @sejournal, @MattGSouthern

A Siteline report tested a Claude agent on 100 top B2B software products. It found access errors and hidden pricing caused the agent to visit third-party sites for info it couldn’t find on official pages.

The report, by Siteline founder David Kaufman, involved running a simulated Claude agent through 534 attempts to discover monthly prices of all plans and highlight main features.

The data reaches the funnel stage most agent-visibility coverage misses, after a buyer knows your product and sends an agent to check pricing and features.

Siteline sells agent analytics and an AI agent readiness tool, so it has a commercial interest in the finding that many sites aren’t prepared for agents.

What Siteline’s Agent Was Asked To Do

Siteline tested 20 products in five categories: productivity, developer tools, marketing and sales, customer support, and analytics. It checked if the agent could access the site, retrieve plans and prices, and determine costs in tokens and tool calls.

At the median, a run on Sonnet 4.6 took about 32 seconds and $0.24, with three search-or-fetch tool calls. Siteline shows a 2.2x time and 4.2x cost difference between the fastest tenth of runs and the slowest, mainly due to web-search calls. Linear was more efficient, parsing four plans in one fetch for about $0.11 on the company’s site.

Access Errors Pushed The Agent Toward Third Parties

About 30% of runs face at least one error fetching or searching a site, with roughly a quarter of those errors being access denials from bot blocking or unreadable pages. Most retries recover, but 5% abandon the brand site for third-party sources, which the report notes can be stale or incorrect.

Errors caused a steep content gap, with access-error runs pulling 58% of content from third-party sources, compared to 12% in error-free runs. Siteline notes Anthropic and OpenAI don’t run JavaScript, unlike Google. SEJ covered Vercel data showing AI crawlers make up 28% of Googlebot volume. Siteline’s data shows 13% of runs carried internal mentions of JavaScript or rendering trouble, not counted as errors. SEJ also highlighted that a third of top fintech homepages returned little content, revealing a JavaScript blind spot.

The examples show this: Zendesk’s pricing page loaded, but the plan table was JavaScript-rendered, according to the report, making it unreadable for the agent, who then relied on third-party blogs at five times the Linear cost. Coda’s pricing fetches failed, leading the agent to use third-party pages. Braze’s agent couldn’t access the pricing page and obtained numbers from G2 and Vendr.

One case proposes a straightforward option. Siteline states that the agent used a non-existent pricing URL, which was not in search results, then depended on third parties. The report recommends keeping a pricing page active, even if it doesn’t show prices.

Where Pricing Pages Came Up Short

Across the runs, 65% of plans showed readable prices, while 14% posted no prices and routed to a sales contact. About 30% of marketing, sales, and customer support products had no prices, unlike zero in productivity and developer tools

A ‘Contact Sales’ button is a dead end for buyer’s agents comparing prices. Siteline suggests it may let agents recommend competitors with public rates. Tests reveal issues: FullStory’s page lacked prices, and Databricks was the most expensive at $0.95 per run, with pay-as-you-go rates hidden behind an inaccessible calculator, redirecting to third parties.

Why This Matters

A client-side loaded plan table may seem empty to agents, even if it looks complete in the browser. Siteline’s fixes, covered by SEJ, include rendering pricing and features server-side and highlighting key details early, as agents typically pull only the first 15,000 to 20,000 tokens. Siteline also recommends llms.txt, but Google’s guidance varies, and its value remains uncertain with independent data.

Siteline is among vendors offering agent-readiness scoring, a category SEJ examined when Cloudflare released its own scanner. The benchmark assesses one model on a single task, measuring agent access to specific sites, not how all agents handle every product.

Looking Ahead

As more buyers have agents compare plans before engaging sales, sites with clear, readable plan details on first view help agents represent confidently. The question is whether agent-readiness measures will rely on shared metrics or diverge across vendor scorecards, each focusing on different signals.

Featured Image: Vasilyev Alexandr/Shutterstock