Found but Unread: Why AI Agents Recommend the Websites They Can Access

TL;DR for operators

A business can appear in AI search results without supplying the information an agent needs to recommend it. Across 1,056 matched businesses, websites classified as agent-ready received clear recommendations in 20.4% of agent journeys, versus 10.6% for less accessible websites. First-party-grounded answers were also more frequent: 77.8% versus 55.5%.

The operational issue arises after discovery. Agents unable to retrieve a business’s own pages usually continue answering from other sources. That fallback increases search effort and can leave requested information out of the final answer. Within matched answer-source comparisons, first-party-built answers achieved 48.3% graded factual accuracy, compared with 34.3% for answers built from other web evidence.

For businesses investing in search-engine optimization (SEO) or answer-engine optimization (AEO), the findings support adding a separate agent-readability audit. Test whether agents can retrieve pricing, features, and documentation, then measure grounding, omissions, recommendation strength, and retrieval cost. The observed advantages are associations, not guaranteed returns from modifying a website.

A website can be discovered without being read

A buyer asks an AI assistant to compare software providers. The questions are straightforward: What does each product cost? Which features are included? What is required for setup?

An agent finds a vendor’s website, but its pricing information depends on browser-side JavaScript or its documentation blocks automated retrieval. The agent can still search comparison websites, cached snippets, and other external sources. It may produce a reasonable-looking response without recovering the details the vendor intended customers to see.

For the business, search visibility has worked. Information delivery has not.

In AX is the New AEO, Ido Finder and colleagues investigate this problem through 37,927 completed agent journeys involving 1,056 real businesses.¹ They distinguish the surfacing stage, where a business enters an agent’s consideration set, from the drill-down stage, where the agent retrieves evidence and constructs an answer.

The authors call the latter concern agent experience (AX). Their empirical analysis concentrates on website accessibility: whether business-controlled content can be fetched and read by agents.

That distinction introduces a separate operational question. SEO and AEO may help an agent find a business, but what determines whether its own information survives into the final response?

Comparable visibility, different recommendations

The researchers divided 1,056 businesses into two equal groups using the accessibility component of ora’s proprietary agent-readiness score. Agent-ready businesses scored at least 0.65; not-agent-ready businesses scored at most 0.50.

They matched businesses within industry and fame categories on measured website fame, prior model knowledge, third-party citation breadth, and discovery scores. The matching controls were closely balanced, leaving accessibility as the principal measured difference.

Four browsing-capable agent configurations then answered standardized questions about pricing, features, and setup.

The main comparisons are substantial:

Source: Sections 5.2–5.7, especially Table 7. Comparisons are aggregated at the business level.

A clear recommendation required both independent model-based judges to assign their highest score. Under that relatively strict criterion, agent-ready businesses had a 1.93-fold recommendation rate (95% bootstrap confidence interval: 1.74–2.15; permutation p < 0.0001).

The direction of the recommendation advantage was consistent across all four agent configurations and twelve business categories. That consistency is useful evidence, although the configurations differ in their models and retrieval tools, and category-level effect sizes vary.

First-party grounding shows a related pattern. The grounded-answer measure identifies journeys that use the business’s own website without outside sources. Its rate was 77.8% among agent-ready businesses, compared with 55.5% among their matched counterparts.

These results suggest that discoverability alone does not explain how agents construct recommendations. Website accessibility is associated with a separate downstream outcome: the extent to which the agent can base its response on business-controlled evidence.

When retrieval fails, the agent usually keeps answering

An inaccessible website does not usually terminate an agent’s task.

The researchers report that approximately 99% of journeys encountering a site dead end still produced an answer. The agent typically compensated by searching elsewhere.

That compensation is measurable. Not-agent-ready businesses generated 57% more web searches, 23% more agent turns, and 15% longer execution times. Their average journey cost was about 16% higher, while the reported cost premium per first-party-grounded answer averaged 64% across the four harnesses.

The source composition of finished responses also changed. First-party pages accounted for approximately 78% of answer text for agent-ready businesses, versus 58% for not-agent-ready businesses. Content derived from web search increased from approximately 12% to 25%.

This provides a plausible mechanism linking accessibility to recommendations.

If directly retrieved pricing or product documentation is unavailable, an agent must reconstruct the answer from material that may not contain the requested details. It can still mention the business, but may provide fewer specifics, qualify its recommendation, or refer the buyer elsewhere.

The study’s recommendation judgments align with this interpretation: not-agent-ready businesses generated more access disclaimers, secondhand-source qualifications, and vague or noncommittal responses.

For operators running browsing agents, the increased searches and turns represent additional execution overhead. For the business being researched, the consequence is reduced control over which published facts inform the answer.

Neither consequence requires the agent to abandon the task.

Missing facts are a larger problem than fabricated ones

The study’s accuracy analysis examines what happens when agents construct answers from different sources.

Researchers captured business facts using a headless browser and graded 2,499 answers against 31,127 requested atomic facts from 131 businesses. Correct facts received full credit, partially correct facts half credit, and incorrect or omitted facts zero.

The central source-comparison analysis retained 196 matched business–harness–question cells from 90 businesses, each containing both site-built and web-built responses.

Site-built answers achieved 48.3% graded accuracy, compared with 34.3% for web-built answers—a 14-percentage-point difference, equivalent to a 41% relative advantage (p < 0.0001).

The more revealing finding concerns the type of error.

Source: Section 5.5, Table 5 and Figure 10.

Web-built answers were 3.7 times as likely to contain none of the requested facts. The increase in omission was considerably larger than the increase in explicitly incorrect statements.

This matters because evaluating only factual claims that an AI agent actually makes can overlook information it failed to provide. An answer might avoid obvious fabrication yet still be inadequate for a buyer deciding between vendors.

The effect also depends on the question. Pricing showed the strongest site-built advantage: 60.4% versus 36.9% graded accuracy. Features showed a smaller difference, 45.6% versus 38.8%, with p = 0.018—not meeting the paper’s stricter p < 0.001 confirmation threshold. Setup showed essentially no difference: 23.9% versus 23.5%.

The results therefore support prioritizing accessible, explicit commercial facts, especially pricing. They do not show that first-party retrieval automatically solves every documentation or setup question. Even site-built answers left substantial requested information unreported.

AEO metrics do not replace retrieval measurements

The authors also examine whether two conventional AEO-related proxies retain independent associations with agent outcomes after accounting for website accessibility and other controls.

The proxies are a discovery score, representing website findability, and citation breadth, representing external mentions across independent sources.

Under the study’s p < 0.001 confirmation threshold, neither proxy had a confirmed independent relationship with grounding, recommendation, or execution cost in the reported partial-correlation analyses.

Accessibility, by comparison, had partial correlations of +0.62 with first-party grounding, +0.41 with clear recommendations, and −0.27 with cost per journey, each at p < 0.0001.

This is not evidence that AEO or SEO has become obsolete. Those activities still address discovery, which the study deliberately distinguishes from subsequent retrieval. The analysis also tests specific proxies, not every possible optimization practice.

One detail illustrates the boundary: discovery score had a positive partial correlation of 0.10 with recommendation at p = 0.002. It did not satisfy the paper’s stricter threshold, but that is different from demonstrating a zero relationship.

For measurement teams, visibility and readability should be treated as separate variables, with completed answer quality assessed downstream of both.

What an agent-readability audit should measure

For a business with established SEO and AEO activity, the immediate decision is whether its existing website architecture adequately serves browsing agents.

A useful audit starts with representative buyer questions, not a universal readiness grade.

First, test whether an agent can retrieve pricing pages, product specifications, and documentation without depending on client-side rendering. Record access failures, missing content, and broken navigation paths. Review bot restrictions with security requirements intact; granting unrestricted automated access is not a necessary response.

Next, inspect completed agent trajectories. Determine whether answers actually use first-party pages, how frequently the agent searches elsewhere, and which requested facts are missing. Repeat the questions across more than one agent configuration because the study identifies meaningful differences between harnesses.

Finally, measure changes after specific website improvements. If pricing information becomes directly retrievable, compare first-party grounding, factual completeness, recommendation strength, search effort, and cost before and after the intervention. Keep the substantive content and test conditions as stable as feasible to distinguish accessibility improvements from content revisions.

This shifts website evaluation from a single visibility metric to an observable sequence: discovery, retrieval, evidence use, and answer quality.

The sequence also helps assign responsibility. Marketing teams can assess discovery; website engineering and security teams can assess fetching; content owners can assess completeness; and AI-channel evaluators can assess what agents actually tell buyers.

What the study does not establish

The largest inference boundary concerns causality.

The readiness groups were matched observationally, not randomly assigned. Businesses with accessible websites may also publish more complete or better-organized information, and content depth was not matched. The accessibility measure itself comes from the authors’ proprietary ranker.

Critically, the direct accuracy comparison between readiness groups was 51.7% versus 49.8%, with p = 0.52. The study does not establish a statistically significant group-level accuracy advantage.

The 41% relative accuracy advantage instead compares site-built and web-built responses within matched cells. That design holds business, harness, and question type fixed, but the agent’s choice of evidence source remains a selected pathway, not a randomized intervention.

Additional boundaries concern the English-first sample, its concentration in SaaS and commerce, model-based grading, and differences between agent stacks. The tasks evaluate information retrieval for buyer questions, not autonomous purchases, transactional APIs, or MCP-based operations.

Consequently, the paper provides meaningful comparative evidence for the read layer of agent experience. It does not establish the financial return from improving any particular website.

From search visibility to evidence availability

The study identifies a website performance dimension that conventional discovery metrics can miss.

A business can be sufficiently visible for an agent to find it while remaining difficult for that agent to research. When first-party retrieval fails, the agent generally substitutes external evidence, incurs additional work, and risks leaving requested facts unanswered.

For businesses already investing in AI search visibility, the next measurable question is whether their own information can be retrieved and used accurately after discovery.

Agent-readable content deserves evaluation alongside SEO and AEO. Its value should be established through observed agent behavior and controlled website improvements, not inferred from a readiness score or an observational recommendation ratio alone.

Cognaptus: Automate the Present, Incubate the Future.