AX is the New AEO

• The web’s prior arrangement served human readers and search crawlers: crawlers found and roughly characterized pages, while people completed the reading in a browser.Source: §1 · p. 1
• The paper describes agents as a new audience that searches, fetches, and reads to answer a person’s current question, unlike crawlers that build an index for later use.Source: §1 · p. 1
• Agent readiness concerns whether a site serves an arriving agent that fetches and reads it; the study treats a site ranked agent-ready as having good AX.Source: §1 · p. 2
• Agents search and fetch through several rounds for one buyer question, building answers from what they read; if the business site is inaccessible, they may use third-party pages, competitors, listicles, or aggregators instead.Source: §1 · p. 2
• The study focuses on AX’s read layer, holding discovery fixed and leaving the act layer to later work; it frames SEO as surfacing and AX as the drill-down step.Source: §1 · p. 2
• The study reports a controlled in-the-wild experiment on agent readiness, covering 37,927 agent journeys across 1,056 live businesses and four harnesses.Source: §1 · p. 2

• Prior AEO/GEO work found that edits can raise a source’s share of answer conditional on retrieval, while later work reports weak or harmful rewrite effects and no stable cross-platform causal discoverability effect.Source: §2 · pp. 2–3
• Earlier agent research established browsing, action interleaving, and tool selection, then evaluated multi-step interactions on real and reproducible sites; Gou et al. (2025) study deep-research correctness and attribution over live browsing.This paper holds agents fixed and compares sites, reversing Gou et al.’s varying-web comparison.Source: §2 · p. 3
• Grounding and attribution research evaluates whether answers draw on retrieved or identified sources; this paper’s first-party grounding and answer-composition measures apply those ideas to the site the agent read.Source: §2 · p. 3
• Prior audits report unsupported or incorrect generative-search answers, while this study holds the engine fixed and varies the source it tries to read.Source: §2 · p. 3
• Third-party pages can expose agents to planted instructions and recommendation steering; Zhang et al. (2026) report that edits of 13 words can place text in 38–51% of deep-research reports.The paper does not measure poisoning and notes that seeded third-party pages are also an attack channel.Source: §2 · p. 3
• Agent-facing web interfaces include MCP, llms.txt, NLWeb, and WebMCP, but the paper leaves structured interfaces aside and measures agents fetching and reading existing site pages.Source: §2 · p. 3

• 7–10% of answers in the four study harnesses came from training memory, compared with 14% for gpt-5.6; the study used a different instrument for its harness estimate.Earlier, gpt-4.1 still built about half its answer from memory; Figure 1 plots training-knowledge share across releases, with the dashed final point measured by this study.Source: §3 · p. 4, Figure 1 · p. 4
• Across three ChatGPT generations, current-year-anchored search rose from 6% to 87%; recency-dependent questions invoke frontier-model search about nine times in ten.Source: §3 · p. 4, Figure 2 · p. 4
• Reddit’s share of ChatGPT citations fell 86–95% within a week, while searches against official domains rose from near zero to roughly a quarter of background searches.Figure 2 presents current-year search, recency-dependent search invocation, and Reddit citation share in three panels.Source: §3 · p. 4, Figure 2 · p. 4
• The paper does not claim memory is irrelevant: stable facts may still be answered in plain chat, but pricing, setup, and comparison questions need current information.Source: §3 · p. 4
• For the buyer-question setting studied, fetching is the default and what the agent reads now decides nearly all of the answer.Source: §3 · p. 4

4.1 Design

• 1,056 businesses—528 per group—were matched on controls within industry and fame bands, with accessibility held out of the matching distance.The controls include third-party citations, fame, and prior model knowledge; agent readiness is scored from site accessibility.Source: §4.1 · p. 5

4.2 Collection

• Each business received pricing, features, and setup prompts, each run three times on four independent harnesses.The harnesses span two model vendors and two search providers, and each run starts in a fresh isolated environment.Source: §4.2 · p. 6
• 131 businesses with extractable facts for at least one intent were graded against full headless-browser captures of their sites.The captures provide ground truth even for sites the agent could not read.Source: §4.2 · p. 6

4.3 Measures

• Grounding measures track first-party evidence share and journeys answered using the site without outside sources; answer sections are separately labeled by source.Source labels distinguish business pages, third-party pages, search snippets, and training knowledge.Source: §4.3 · p. 6
• Two judges rate recommendations from 0 to 4, while factual accuracy scores correct and partial facts against the facts asked, counting omissions as failures.A clear recommendation requires both judges to give the top score; accuracy is graded against captured site facts.Source: §4.3 · p. 6
• Group differences use business-level averages, permutation tests, and bootstrap confidence intervals; site-built and web-built answers are compared within business, harness, and question-type cells.The analysis treats each business as the unit because its nine journeys are not independent.Source: §4.3 · p. 7

5.1 Where the agent actually goes

• In two traces of the same buyer question and harness, twilio.com answered from its documentation in seven steps, while hashicorp.com was blocked twice and searched the open web seven times.The hashicorp.com trace ended on two competitors’ blogs.Source: §5.1 · p. 7, Figure 3 · p. 7

5.2 Recommendation: agent-ready businesses get recommended

• 20% of answers clearly recommended agent-ready businesses against 11% for businesses that were not agent-ready, a 1.9× difference.The lift appeared in every harness and all twelve categories; weak-to-recommend answers were 2.45× more common for not-agent-ready sites, alongside more frequent hedging.Source: §5.2 · pp. 7–8, Figure 4 · p. 8

5.3 Grounding: when agents can’t read you, the answer is built without you

• 78% of agent-ready answers were built from the business’s own pages against 58% for not-agent-ready businesses, while training knowledge remained at 7–10% either way.As accessibility fell, mean searches rose from 1.8 to 4.5; losing readability raised searches across every stack and category.Source: §5.3 · pp. 9–10, Figure 7 · p. 9, Table 4 · p. 10

5.4 Cost: blocked sites cost the agent more per answer

• 64% more was spent per grounded answer for not-agent-ready sites on average across the four stacks.Those runs also took 23% more turns and 15% longer to reach an answer; per-run costs varied because stacks differed in how much they read and searched.Source: §5.4 · pp. 10–11, Figure 9 · p. 10

5.5 Accuracy: the common failure is a missing fact

• 48.3% of asked facts were correct in site-built answers against 34.3% in web-built answers, while web-built answers were 3.7× more likely to contain none of the requested facts.The fact-level difference was mainly omission: facts were never mentioned in 45% of web-built answers versus 29% of site-built answers; stated-wrong facts rose from 4% to 6%.Source: Table 5 · p. 11, §5.5 · p. 11, Figure 10 · p. 11
• Pricing showed the largest accuracy gain from site-built answers (+64%), features showed +18%, and setup showed no confirmed channel difference (+2%, p = 0.87).Across harnesses, the gain was largest for claude-code (+382%) and absent for openclaw (+10%, p = 0.05) and eve (−5%, p = 0.39). Accuracy also rose from 41% to 56% across bins of first-party evidence share.Source: §5.5 · pp. 11–12, Figure 11 · p. 12
• The assigned-group accuracy contrast was not confirmed: the graded subset scored 51.7% against 49.8% (p = 0.52), and establishing the group effect directly was left to future work.On pricing, the group comparison was 61.7% against 55.6%, but none of the group comparisons reached significance on 131 businesses.Source: §5.5 · p. 12

5.6 The AEO control: no effect of its own

• After controlling for accessibility and other factors, neither AEO proxy had a confirmed link to grounding, recommendation, or cost; accessibility had a strong link to all three.The one hint was discovery and recommendation (r = 0.10, p = 0.002), below the study’s confirmation bar.Source: §5.6 · p. 12, Table 6 · p. 12

6 Conclusions

• Recommendation, grounding, and search contrasts hold across every harness and business category tested, while accuracy and cost effects vary by harness.The accuracy and cost findings are averages; the recommendation, grounding, and search findings are consistent across the tested groups.Source: §6 · p. 13
• Only 7–10% of answers comes from training knowledge; agent-ready answers draw 78% from businesses’ own pages, against 58% for not-agent-ready sites.The training-knowledge share is similar whether or not the site is agent-ready.Source: §6 · p. 13
• 20% of agent-ready businesses receive a clear recommendation against 11% of not-agent-ready businesses, a 1.9× difference.The recommendation contrast holds across every harness and business category tested.Source: §6 · p. 13
• Within the same business, harness, and question, site-built answers get 41% more asked facts right; web-built answers are 3.7× more likely to contain none.Wrong facts change only from 4% to 6%, while unmentioned facts rise from 29% to 45%.Source: §6 · p. 13
• A grounded answer about a not-agent-ready business costs 64% more on average, with the premium varying by harness.The reported increase ranges from 11% on eve to 93% on two other stacks.Source: §6 · p. 14
• With accessibility held fixed, neither citation breadth nor discovery predicts grounding, recommendation, or cost; accessibility predicts all three.The study reports this pattern across all 1,056 businesses.Source: §6 · p. 14
• The authors conclude that answer-engine optimization decomposes into SEO for surfacing and AX for the drill-down step and everything after it.They characterize being readable as more consequential at the drill-down step than being talked about.Source: §6 · p. 14

7 Limitations and Future Work

• Content depth is unmatched, so thinner content among not-agent-ready businesses could widen the grounding and recommendation gaps.The accuracy comparison is within the same business, but the between-business contrasts cannot rule out this content-depth difference.Source: §7 · p. 14
• Ground truth covers 131 businesses, and the 41% accuracy contrast is not significant at this sample size.The contrast points in the same direction across question types and harnesses; a larger graded set would test it directly.Source: §7 · p. 14
• The cohort is SaaS- and commerce-heavy and English-first, and the questions concern business-fact lookups.The authors identify more verticals, languages, and transactional tasks as areas for extension.Source: §7 · p. 14