Where AI agents get stuck on B2B sites

The next major wave of artificial intelligence is agentic: Google is introducing agentic tasks in Search, the web now receives more visits from bots than humans according to Cloudflare, and Salesforce reports that 20 percent of revenue already flows through agents. Sixty percent of companies run agents in production, and three out of four are actively investing in AI agents. For B2B brands, this means websites are no longer read only by humans, but by systems that solve tasks, extract facts, and cite sources.

To assess how ready B2B websites are for agentic visitors, a joint analysis with David Kaufman, founder of Siteline, examined how agents scan websites and where they get stuck. The result: most sites are fundamentally agent-ready, but there is one critical breaking point. AI agents do not read websites like humans. They receive a task, search the web, fetch pages, extract facts, and cite the sources they used. A page can persuade a human and still fail an agent if facts are hard to find (opacity), hard to fetch (machine-readability), or hard to cite (access friction). Agents turn websites from showrooms into barcodes.

How agent behavior was measured

The study simulated realistic buyer scenarios for 100 B2B products. The agent had to find the official vendor website on its own, without provided starting links. Three buyer-related tasks were repeated five times each to account for the probabilistic nature of LLMs: find pricing and features, find integrations, and clarify security and compliance. What mattered was not whether information existed somewhere on the web, but whether the agent could reliably answer from the vendor's own site.

Pricing pages break first-party sources

Once prospects look for pricing, they compare and enter with high purchase intent at the bottom of the funnel. That makes pricing the hardest and most important test of whether a vendor site can serve agents directly. Good pricing pages must satisfy three requirements at once: companies want to control price disclosure, buyers want fast comparison, and agents need clear, fetchable, citable facts.

When retrieving pricing information, agents get stuck far more often than with security or integrations. The first-party answer rate for pricing and features was 79 percent, with an 84 percent first-party citation share. Integrations reached 93 and 99 percent, security 92 and 99 percent. Pricing and feature queries caused 77 percent of all third-party citations. The most common breaking point is therefore not the missing website, but the pricing page.

Hidden pricing is only half the story

Hiding prices forces agents to external sources. Published prices do not fully solve the problem. Among runs without real price disclosure, 45 percent cited at least one third-party source, while 55 percent stayed with the vendor, usually noting contact sales. Even with numerically visible prices, agents still cited third parties in 18 percent of runs. The price can be on the page and still be hard to extract, low-trust, or unclear to cite.

Pricing pages often look complete to humans but are not reliable enough for agents. If you use complex pricing models, explain them clearly and make them machine-readable instead of relying on external directories you cannot control.

Three reasons agents fail

Agents fail to retrieve pricing for three reasons: opacity, machine-readability, and access friction. Opacity means no concrete price is published or it is only vaguely packaged. Machine-readability fails when prices exist but cannot be confidently extracted because of page structure, JavaScript, calculators, toggles, screenshots, PDFs, or ambiguous tables. Access friction comes from fetch errors, rate limits, blocking, or unreachable pages.

Access errors appeared in only 7 percent of all runs but had a massive impact: for pricing queries, third-party fallback rose from 17 to 77 percent. The harder a site is to fetch, the more work an agent invests before citing or falling back.

The fallback web is messy

Fallback means agents turn to third-party sources because first-party information is missing or unusable. This is the biggest risk because external data is patchy and not under your control. Of 580 third-party pricing citations, 52 percent came from editorial sources such as blogs, media, and comparison articles, 46 percent from directories like G2, Capterra, or Vendr, and 2 percent from app stores, marketplaces, or partner pages. Agents reconstruct pricing from a mixed web instead of a clean vendor source.

How to make a website agent-proof

An agent-proof pricing page keeps citations with your own vendor instead of directories. The measures follow the three failure modes.

• Fix opacity: Publish real prices as text for every self-serve tier. For custom models, explain what drives the price instead of only writing contact sales. Bundle plan names, prices, limits, and features on one canonical pricing URL and point all references there.
• Ensure machine-readability: Deliver prices in crawlable server HTML because many agents do not execute JavaScript. Add schema.org Product and Offer markup with price and priceCurrency. Explain usage-based models in text, not only in calculator widgets.
• Reduce access friction: Allow AI crawlers in robots.txt, for example GPTBot, ClaudeBot, PerplexityBot, and Google-Extended. Do not block server-side fetches on pricing pages. Place price information early in the DOM and keep pages lean.

Address opacity and machine-readability first because they drive most fallback. Then test the central buyer question yourself: find all pricing and features for product X and measure whether the agent reliably cites your own domain.