The Invisible Gatekeepers: How B2B Software Sites are Failing the AI Agent Test
In the rapidly evolving landscape of B2B procurement, a new type of buyer has emerged: the AI agent. These autonomous programs, powered by Large Language Models (LLMs), are increasingly being tasked with the heavy lifting of software evaluation—parsing through pricing tiers, feature lists, and technical documentation to provide recommendations to human decision-makers. However, a groundbreaking new study suggests that the majority of the world’s top software companies are inadvertently slamming the door in these agents’ faces.
A comprehensive report released by Siteline, an AI agent analytics firm, has revealed a startling "readiness gap" in the B2B software sector. After testing a Claude-based agent on 100 of the top B2B software products, the study found that technical barriers, hidden pricing, and architectural choices are forcing AI agents to abandon official brand websites in favor of third-party sources. This shift not only diminishes a brand’s control over its own narrative but also risks the dissemination of stale or inaccurate information to potential buyers.
Main Facts: The High Cost of Digital Inaccessibility
The Siteline report, authored by founder David Kaufman, highlights a critical friction point in the modern sales funnel. While much of the current discourse around "AI SEO" focuses on how agents discover products, Siteline’s data focuses on the "consideration" stage—the moment after a buyer has identified a product and sends an agent to verify specific details like monthly costs and core functionality.
The findings paint a picture of a digital ecosystem ill-equipped for the automated future. Key takeaways from the 534 simulated agent attempts include:
- The Error Threshold: Approximately 30% of all agent runs encountered at least one error when trying to fetch or search a site.
- The Third-Party Pivot: When faced with access denials—often caused by aggressive bot-blocking or unreadable JavaScript—5% of agents abandoned the official brand site entirely, seeking information on third-party platforms like G2, Vendr, or community blogs.
- The Content Gap: In runs where access errors occurred, agents pulled a staggering 58% of their information from third-party sources. In contrast, error-free runs relied on third parties for only 12% of their data.
- The "Contact Sales" Dead End: 14% of the analyzed products provided no public pricing, instead routing users to a sales contact. For an autonomous agent, this is effectively a hard stop that may lead to the agent recommending a competitor with transparent rates.
Chronology of the Study: 534 Attempts to Bridge the Gap
To gather this data, Siteline conducted a rigorous simulation using a Claude 3.5 Sonnet agent. The study was structured to mimic the behavior of a sophisticated procurement agent tasked with a specific mission: discover the monthly pricing for all available plans and highlight the primary features of the software.
Phase 1: Categorization
The researchers selected 100 top-tier B2B products, divided equally into five distinct categories:
- Productivity Tools
- Developer Tools
- Marketing and Sales Platforms
- Customer Support Software
- Analytics Engines
Phase 2: Execution and Tool Calls
The agent was granted the ability to use "search-or-fetch" tool calls. Over 534 attempts, the study tracked the efficiency of these interactions. At the median, a successful run took roughly 32 seconds and cost $0.24 in API tokens. However, the study noted a massive variance in efficiency. The slowest 10% of runs were 2.2 times slower and 4.2 times more expensive than the fastest 10%, primarily due to the agent having to perform multiple web searches to compensate for unreadable primary pages.
Phase 3: Analysis of "Agentic Friction"
The researchers then analyzed why certain sites, such as Linear, were processed with high efficiency (parsing four plans in a single fetch for $0.11), while others, like Databricks, cost nearly a dollar per run ($0.95) because the agent had to navigate complex, inaccessible calculators and third-party references.
Supporting Data: The JavaScript Blind Spot and Token Limits
The technical root of many "agent failures" lies in how modern websites are built. Many B2B sites rely heavily on client-side rendering (JavaScript) to display dynamic content like pricing tables.
The Rendering Problem
Siteline’s data confirms a long-suspected "JavaScript blind spot" for AI agents. Unlike Google’s crawlers, which have spent decades perfecting the ability to render and index JavaScript, many current LLM-based agents (including those from Anthropic and OpenAI) often do not execute JavaScript when fetching a URL.
This leads to a "ghost page" effect. To a human in a browser, the page looks complete; to an AI agent, the page appears empty or missing its most vital data. The report found that 13% of all runs contained internal mentions of JavaScript or rendering troubles that weren’t even officially logged as "errors," but still hindered the agent’s ability to gather data. This corroborates recent industry data from Vercel, which shows that while AI crawlers account for 28% of Googlebot-equivalent volume, many sites remain "invisible" to them.
The Token Window
Another critical data point involves the "first-view" priority. Agents typically only pull the first 15,000 to 20,000 tokens of a page. If a site is bloated with header code, marketing jargon, or heavy scripts before reaching the actual pricing and feature data, the agent may "timeout" or truncate the content before it reaches the information the buyer actually needs.
The Category Divide
The report also revealed a cultural divide in how different software sectors handle transparency:
- Productivity and Dev Tools: 0% of these products hid their pricing behind a "Contact Sales" wall in the study.
- Marketing, Sales, and Customer Support: 30% of these products lacked public pricing, creating a significant barrier for automated procurement.
Official Responses and the Emerging "Agent Readiness" Industry
While the software companies mentioned in the report—such as Zendesk, Coda, and Braze—have not issued formal rebuttals, the findings have ignited a debate within the SEO and web development communities.
Siteline, which has a commercial interest in selling agent analytics and readiness tools, argues that the current state of the web is "agent-hostile." They are not alone in this assessment. Earlier this year, Cloudflare released its own "AI Audit" tool and an "Agent Readiness Score," signaling that the infrastructure layer of the internet is preparing for a world where bot traffic is no longer just "spam" to be blocked, but "customers" to be served.
The report highlights specific failures as cautionary tales:
- Zendesk: While the pricing page loaded, the plan table was rendered via JavaScript. The agent found it unreadable and turned to third-party blogs, resulting in a five-fold increase in cost and a loss of brand control.
- Braze: The agent was blocked by security protocols on the pricing page, forcing it to obtain data from G2 and Vendr—platforms where the brand cannot guarantee the accuracy of listed tiers.
- Databricks: Hidden "pay-as-you-go" calculators proved to be a "black hole" for the agent, leading to the most expensive and least efficient runs in the study.
Implications: Strategic Shifts for the AI Era
The Siteline report serves as a foundational document for what many are calling "Agentic SEO" or "AI Optimization" (AIO). The implications for B2B companies are profound and require a shift in both technical architecture and sales philosophy.
1. The Death of the "Contact Sales" Friction
In a world where an AI agent is making the first cut of a vendor list, friction is fatal. If an agent can’t find a price for Product A but finds a clear, readable price for Product B, Product B is statistically more likely to make the shortlist. Companies may need to provide "starting at" prices or agent-readable ranges to ensure they aren’t filtered out by autonomous buyers.
2. Server-Side Rendering (SSR) is No Longer Optional
To be visible to AI agents, the most critical data—pricing, features, and FAQs—must be rendered server-side. If the information requires a browser to execute JavaScript to become visible, it is effectively invisible to the current generation of AI agents.
3. The Rise of llms.txt
The report touches on the recommendation of using an llms.txt file—a proposed standard for providing a markdown-based, agent-friendly version of a website’s key information. While Google’s guidance on this remains inconsistent, the Siteline data suggests that providing a lightweight, text-only "cheat sheet" for agents could drastically reduce the cost and error rates of site visits.
4. Brand Control vs. Third-Party Decay
When an agent is forced to use third-party sites like G2 or Reddit, the brand loses the ability to update its pricing in real-time. This creates a "data decay" problem where agents may recommend products based on discontinued tiers or outdated feature sets.
Looking Ahead: The New Benchmark for Success
As we move toward a future where "buyers have agents," the definition of a "good" website is changing. It is no longer enough for a site to be aesthetically pleasing to humans; it must be computationally efficient for machines.
The Siteline benchmark is just the beginning. As David Kaufman’s report suggests, the question for B2B leaders is no longer if agents are visiting their sites, but how much it costs those agents to find the truth. Companies that prioritize "agent-readiness" today by simplifying their architectures and embracing transparency will likely find themselves at the top of the AI-generated shortlists of tomorrow. Those who continue to hide behind JavaScript walls and "Contact Sales" buttons may find themselves increasingly invisible to the modern buyer’s most trusted assistant.
