--- url: https://ora.ai/blog/ax-is-the-new-aeo title: "AX is the new AEO · ora // research" description: "37,927 agent runs across Claude and ChatGPT harnesses and 1,056 businesses: what agent readiness does to recommendation, grounding, cost, and accuracy. Ora Research." crawled_at: 2026-10-18T19:27:24Z ask: https://ora-ai.ora.run/ask docs: https://ora-ai.ora.run/.well-known/ora/docs lang: en --- # AX is the new AEO ora research · Oct 8, 2026 · 12 min read [Share on X](https://x.com/intent/tweet?text=AX%20is%20the%20new%20AEO&url=https%3A%2F%2Fora.ai%2Fblog%2Fax-is-the-new-aeo) [Share on LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fora.ai%2Fblog%2Fax-is-the-new-aeo) In 2023, models answered from their training data, and hallucinated when they didn't. An industry grew around seeding that knowledge: AEO (answer engine optimization), with a playbook of Reddit threads, listicles, and citations on other sites. In August 2026 that playbook broke. Reddit's share of ChatGPT citations fell 86% in a week as ChatGPT started querying official sites directly, and has only partly recovered since. The buyer's agent now searches the web with tools like [Tavily](https://tavily.com/) , opens your pages, reads them, and verifies before it recommends anything. AX (agent experience) is whether that agent can find, read, and check your site itself. We ran ~38,000 of those journeys across multiple harnesses running on Claude and GPT models, against 1,000+ real business sites, each answering a real buyer question, split by one axis: is the site ready for agents? Much of what we found was unexpected. When a site shut them out, agents almost never gave up. They answered anyway from other sources, and answers built that way were 3.7× more likely to miss every fact the buyer asked for. [paper arXiv:2609.34951](https://arxiv.org/abs/2609.34951) [data github.com/ora/research](https://github.com/ora/research/tree/main/ax-is-the-new-aeo) research ora.ai/research ### Agent-ready businesses get recommended 1.9× more often, and up to 2.6× depending on the agent stack share of answers both blind judges rate a clear endorsement · by agent stack agent-ready not agent-ready all stacks 37,927 runs pooled claude-code claude-haiku-4-5 · claude-agent-sdk claude-sonnet-4-6 · openclaw gpt-5.4-mini · eve gpt-5.4 37,927 agent runs · 1,056 businesses ora.ai/research 00 / the shift, in public data ## Agents rely far less on training data We put our buyer questions to earlier OpenAI generations and measured how much of each answer came from memory rather than retrieval. The last non-reasoning flagship, gpt-4.1, still built about half its answer from memory. Then reasoning and tool use became the default, and it falls off a cliff: down to 7–10% in the models agents run today. ### How much of the answer comes from the model's training knowledge same buyer questions, each OpenAI generation since gpt-4.1 630 runs, 90 buyer questions per OpenAI generation, web tools optional, every claim scored by an independent judge, points are release half-year averages ora.ai/research ### Three public signals show the same turn toward grounding claude searches when the answer needs fresh data 9 in 10 recency-dependent questions trigger a live web search "Look It Up", arXiv 2025, rev. 2026 · [source](https://arxiv.org/abs/2511.18931) chatgpt searches anchored to the current year 6% · GPT-5.2 · 53% · GPT-5.3 · 87% · GPT-5.5 TUM & maestra.ai · [source](https://maestra.ai/blogs/is-seo-dead) reddit citations fall as chatgpt runs site: searches Promptwatch & Qwairy · [source](https://promptwatch.com/data/reddit-citations-are-dropping-in-chatgpt) Even the sources it cites are a moving target. Starting August 8, site: queries went from near zero to 23–24% of ChatGPT's background searches, aimed at official sites instead of discussion threads. Any off-site channel can vanish in one update, so none is worth optimizing for. The page it keeps returning to is your own. Note: models don't search on everything. Stable facts still come from memory. But pricing, setup, and comparison questions need current information, and in agent harnesses fetching is the default. 01 / trace replay · one ask, two journeys ## Where the agent actually goes Two real traces for the same buyer question, "what can I try for free?" , on the same harness. twilio.com answers from its own docs. hashicorp.com blocks the agent twice and the journey ends on two competitors' blogs. the site's own page errored / dead end web search another site two real traces, claude-agent-sdk · claude-sonnet-4-6 · click a panel to expand 02 / recommendation ## Agents recommend you twice as often when they can read you We measure how strongly an agent recommends each business. An agent-ready site earns a clear, confident recommendation nearly twice as often as a blocked one. ### One in five answers clearly recommends an agent-ready business. One in nine when it isn't. both blind judges rate the answer a clear endorsement · all stacks pooled 1.9× · agent-ready · 20% · not agent-ready · 11% independent recommendation: how strong a case the agent could build · two blind judges (claude and gpt) must both give the top grade · they agree within a point on 88% ora.ai/research The lift is steepest exactly where buying decisions are researched: 2.5× in IT infrastructure, 2.3× in sales & marketing, and up to 2.6× (14% vs 5% on claude-agent-sdk) depending on which stack the buyer's agent runs on. The flip side holds too: when a site is not agent-ready, answers that both judges rate too weak to recommend at all are 2.5× more common. #### breakdown · by category, by stack, and how the hedge looks steepest categories · endorsement rate, both judges | category | ready | not ready | lift | |---|---|---|---| | IT Infrastructure | 24% | 9% | 2.5× | | Sales & Marketing | 21% | 9% | 2.3× | | Artificial Intelligence | 23% | 10% | 2.2× | | Development | 32% | 14% | 2.2× | by harness | stack | ready | not ready | lift | |---|---|---|---| | claude-code claude-haiku-4-5 | 5% | 2% | 2.6× | | claude-agent-sdk claude-sonnet-4-6 | 14% | 5% | 2.6× | | openclaw gpt-5.4-mini | 25% | 14% | 1.8× | | eve gpt-5.4 | 36% | 20% | 1.8× | ### What the hedge looks like The four ways an answer undermines its own recommendation, agent-ready vs not: admits it could not find or access the information "the site returned 403, so…" 4% · 16% · 4.4× vouches from secondhand sources cites aggregators instead of the business 5% · 15% · 3.0× vague, noncommittal endorsement no concrete prices, steps, or links 7% · 12% · 1.8× punts the user to go check themselves "contact sales for a quote" 15% · 22% · 1.4× · agent-ready not agent-ready · 03 / grounding ## When agents can't read you, the answer is built without you Across the finished answers, training knowledge stays at 7–10% either way , so agents are not falling back on training data. Whatever the site does not provide, web search fills in: on sites that are not agent-ready, a quarter of the answer ( 25% ) comes from web search, against 12% on agent-ready sites. ### When the agent can't read you, a quarter of the answer comes from web search what the finished answer is built from · training knowledge stays at 7–10% either way your pages external web search training knowledge agent-ready 78% your pages · 3% external · 12% web search · 7% training knowledge not agent-ready 58% your pages · 7% external · 25% web search · 10% training knowledge share of the finished answer, by evidence source ora.ai/research In ~99% of the journeys that hit a dead end on a site, the agent answered anyway, built from whatever it found elsewhere. ### Web search rises as readability falls The less an agent can read your site, the more it searches the web: 2.6× from the most readable sites to the least. Every search is another chance the agent gets fed information you never published, or stopped updating : other people's pages, stale mirrors, whatever the ranker serves. Web search is provided by [Tavily](https://tavily.com/) on openclaw and eve; Claude stacks use built-in search. The ambiguous middle band of scores (0.50–0.65) is excluded by design. ### Web searches per run, by readability average per agent run · 1,056 domains 2.6× more web searches ← not agent-ready ora accessibility score agent-ready → ora.ai/research #### breakdown · web searches by harness and domain category Search reliance is a stack personality. openclaw navigates by search and claude-code barely searches, but losing readability pushes every stack, and every category, the same way. Web searches per run, agent-ready → not: by harness | stack | ready | not ready | lift | |---|---|---|---| | claude-agent-sdk claude-sonnet-4-6 | 0.4 | 0.9 | 2.4× | | claude-code claude-haiku-4-5 | 0.1 | 0.3 | 3.5× | | openclaw gpt-5.4-mini | 6.9 | 10.4 | 1.5× | | eve gpt-5.4 | 1.2 | 1.9 | 1.6× | by domain category · steepest 5 | category | ready | not ready | lift | |---|---|---|---| | Consumer | 1.9 | 3.7 | 2.0× | | IT Infrastructure | 2.0 | 3.9 | 1.9× | | Development | 1.7 | 3.2 | 1.8× | | Artificial Intelligence | 1.8 | 3.1 | 1.7× | | Sales & Marketing | 2.2 | 3.6 | 1.6× | 04 / cost & performance ## Blocked sites cost the agent's operator more per answer On sites that are not agent-ready, runs take 23% more turns (5.5 → 6.8), get blocked by the site 2.1× as often, and take 15% longer to reach an answer (48s → 55s). The extra turns and retries are billed to whoever runs the agent. Cost per run understates it. A 403 is cheap; the expensive part is that only 56% of runs on not-agent-ready sites end with an answer grounded in the business, against 78% of runs on agent-ready sites. Spread the cost of the failed runs over the answers that did use your site and the price is +64% per grounded answer , averaged across all harnesses using Claude and GPT. ### Every grounded answer costs the agent's operator up to 93% more when the site isn't agent-ready cost per answer grounded in the business · agent-ready → not agent-ready · per stack openclaw gpt-5.4-mini · +93% · $0.135 → $0.260 · claude-agent-sdk claude-sonnet-4-6 · +93% · $0.112 → $0.216 · claude-code claude-haiku-4-5 · +57% · $0.020 → $0.031 · eve gpt-5.4 · +11% · $0.091 → $0.101 +64% averaged across stacks ora.ai/research 05 / accuracy ## The common failure is a missing fact Every answer is graded fact by fact against ground truth captured from each business's own site: 31,127 asked facts across 2,499 answers and a sampled 131 of the 1,056 businesses. Wrong facts are rare in both conditions. The failure that grows when the agent can't read the site is an answer that contains none of the facts the user asked for . It is fluent and confident, so hallucination metrics and AEO dashboards do not flag it. ### Answers built off-site are 3.7× more likely to contain none of the facts the buyer asked for read from the site 7% from the web / other sites 25% share of answers with none of the facts the buyer asked for · 3.7× higher off-site · paired within domain, harness, and question type 2,499 graded answers · 131 businesses ora.ai/research ### Wrong facts stay rare either way. Missing facts grow from 29% to 45%. the fate of every asked fact · by where the evidence came from stated right half right stated wrong never mentioned read from the site 40% stated right · 27% half right · 4% wrong · 29% never mentioned from the web / other sites 28% stated right · 21% half right · 6% wrong · 45% never mentioned 31,127 asked facts graded against each business's own site ora.ai/research ### The source predicts the accuracy Answers built from the business's own pages are 1.4× as accurate as answers built from web search and write-ups on other sites, and the gap is widest on pricing questions ( 60% vs 37% ), where secondhand information ages fastest. Grouping answers by how much of their evidence came from the company's own pages, accuracy climbs from 41% to 56% . ### Accuracy climbs +37% as the evidence comes from your own site by share of evidence read from your site ← none evidence read from your site all → 2,425 answers ora.ai/research #### breakdown · accuracy by question type, empty answers by stack accuracy by question type · built from the site vs from the web pricing · 60% · 37% · +64% · features · 46% · 39% · +18% · setup · 24% · 24% · even empty answers by stack · built from the site vs from the web claude-code claude-haiku-4-5 · 15% · 70% · 4.8× · claude-agent-sdk claude-sonnet-4-6 · 11% · 21% · 1.9× · openclaw gpt-5.4-mini · 7% · 9% · 1.2× · eve gpt-5.4 · 7% · 7% · 1.0× read from the site from the web or other sites ### Pricing pays. Setup never surfaces. Prices have one fresh source, your own pages, so reading them pays. Setup is harder: two-thirds of its facts never surface, and when stated they are equally right from either source. What moves setup is findability: sites where agents could ground setup answers score 31% vs 18%, from both sources alike. 06 / what to do about it ## The fix is a site agents can read Current AEO practices shape what models learned about you in training and the mentions search engines pick up across the web. AX is about making your own site accessible, readable, and usable for the agent already on it. Every measure in this study moves the right way when a site is agent-ready. We think the arbitrage is closing. Tactics that shape what gets said about you elsewhere (seeded threads, listicles on other sites, citation placement) lose leverage each time agents go to the source instead of reading about it. What is left is SEO so agents can find you, and AX so they can reach your site, read it, and use it. Start with the two resources below to see what your site needs to become agent-ready. score your site ora.ai The agent-readiness ranker behind this study's accessibility score: 100+ checks, per-layer breakdown, free browser score. the standard agentready.org The open specification for agent readiness (find, read, act), with the practices ranked by measured agent behavior. 07 / what we didn't measure ## Agents that check their own connectors first Every journey in this study starts the same way. The agent searches, finds your site, and reads it. That's how most agents work today. We don't expect all of them to work that way a year from now. Assistants are building their own front door to businesses. - june Microsoft, Google and Hugging Face, with partners including GitHub and Snowflake, published a draft of Agentic Resource Discovery (ARD), an open spec that lets an agent search a registry for tools, MCP servers and other agents before it calls anything. - july OpenAI replaced ChatGPT's app directory with a plugin directory, now the main place to find what ChatGPT can connect to. - september Meta launched Muse with connectors for email, restaurants, payments and Shopify stores. Muse can connect to any service with a public API using credentials the user provides. - september Instinct raised $1B for a personal agent that plugs into your email, messages and apps and works from there. Why does this matter for what we measured? If an assistant checks its own connectors and registries first, the web becomes the fallback. Muse already works this way: it uses a service's API when one exists and opens a browser only when none does. A business with a connector could win the answer, the recommendation and the transaction without the agent ever opening its site. A business without one might never be considered, however readable its site is. We've already seen a preview. When Amazon blocked Muse, Shopify merchants stayed reachable through a declared checkout path. We haven't measured this, so we're not claiming it. It's our next study. We want to know how often an agent with a relevant connector skips the web entirely, whether a connector or ARD listing changes how often a business gets recommended, and whether agents actually read registry metadata when they choose. If the answer is "often", the find layer of AX grows to include the assistant's own catalog, alongside search. 08 / conclusions ## What we conclude Five findings hold across the four Claude and ChatGPT harnesses and every business category we tested. Together they describe a market where training data no longer decides who gets recommended. 01 Training knowledge is no longer where the answer comes from. Agents now ground answers in the live web: frontier models search about nine times out of ten when a fact is fresh, and in our traces training knowledge supplies just 7 to 10% of the finished answer, readable site or not. Optimizing what the model knows about you optimizes a shrinking slice. 02 Readability is the largest single lever we measured. Agent-ready businesses are recommended 1.9× more often on average and up to 2.6× (14% vs 5% on claude-agent-sdk) depending on the harness. When the agent can read the site, 78% of the finished answer is built from the business's own pages, against 58% when it cannot. 03 When the agent can't read you, it builds the answer without you. Blocked agents do not give up. They go to web search, whose share of the answer doubles from 12% to 25%. The answer built off-site is 3.7× more likely to contain none of the facts the buyer asked for, yet it reads fluent and confident, so nothing flags it. And accuracy follows the source: an answer grounded in your own pages is 1.4× as accurate. 04 Blocking agents is paid for by the agent's operator, then by you. On not-agent-ready sites, runs take 23% more turns, hit a block 2.1× as often, and each grounded answer costs the operator up to 93% more. Every harness is under pressure to route around sites that behave this way. 05 The AEO arbitrage is closing. Tactics aimed at the model's training knowledge (seeded threads, listicles on other sites, citation placement) lose leverage each time retrieval moves closer to the source. Starting August 8, 2026, site: queries went from near zero to 23–24% of ChatGPT's background searches. AEO is now SEO plus AX. SEO affects how you surface in search, and AX decides what the answer says about you. Read the [full paper](https://arxiv.org/abs/2609.34951) . The analysis code and a sample of the derived data are released at [github.com/ora/research](https://github.com/ora/research/tree/main/ax-is-the-new-aeo) . the experiment ## How we designed the research We ran the study across 1,056 live business sites. Real buyer questions went to four agent harnesses spanning Claude and GPT models: claude-agent-sdk, claude-code, openclaw, and eve. The two GPT stacks searched the web through [Tavily](https://tavily.com/) ; the Claude stacks used their built-in search. That comes to 37,927 agent journeys and 140,000+ page fetches from those sites, each journey run in its own isolated environment. Beforehand we scored every site for agent readiness. We then paired each agent-ready business with a not-agent-ready one in the same industry, matched on brand fame, how much the model already knew about the brand, and how widely it is cited on sites it doesn't own, which we measured with Tavily searches. The two groups are indistinguishable on all three, so the lift holds with them equal: the 1.9× recommendation gap is the controlled number. Sites were scored on August 24, 2026, and the journeys ran in late August and September 2026, on claude-sonnet-4-6, claude-haiku-4-5, gpt-5.4-mini, and gpt-5.4. Full experimental design is in [the preprint](https://arxiv.org/abs/2609.34951) . 09 / sources 01 [AX is the New AEO](https://arxiv.org/abs/2609.34951) Ido Finder, Assaf Elovic, Gad Shalev, Liad Yosef · ora research, september 2026 · arXiv:2609.34951 02 [The Agent Readiness Standard on GitHub](https://github.com/agentready-org/standard) open specification and data · agentready.org · the practices ranked by measured agent behavior 03 [ora / research · ax-is-the-new-aeo](https://github.com/ora/research/tree/main/ax-is-the-new-aeo) published data and analysis code for this report 04 ora ranker the agent-readiness ranker that splits the 1,056 businesses in this study · 100+ checks, per-layer breakdown 05 [Tavily: The web access layer for AI agents](https://tavily.com/) Tavily, 2026 · the web-search API for agents behind the openclaw and eve stacks, and the source of the third-party citation control 06 [Is SEO dead? How ChatGPT models search for product recommendations](https://maestra.ai/blogs/is-seo-dead) TUM and maestra.ai, 2026 · 1,650 runs, 370k search results · share of Business-tier runs whose searches are anchored to the current year: GPT-5.2 ~6%, 5.3 53%, 5.5 87% (Plus-tier runs 60% / 79%) 07 [Look It Up: Analysing Internal Web Search Capabilities of Modern LLMs](https://arxiv.org/abs/2511.18931) Sahil Kale, arXiv:2511.18931, 2025 · figures from v2 (august 2026) · on recency-dependent queries, search-invocation rates of 91.0% (Claude Haiku 4.5) and 93.9% (Claude Sonnet 4.6), vs 87.5–92.1% for GPT-5-mini / GPT-5; end-to-end accuracy still caps near 66–71% 08 [Reddit Citations Are Dropping in ChatGPT](https://promptwatch.com/data/reddit-citations-are-dropping-in-chatgpt) Promptwatch, august 2026 · Reddit's share of ChatGPT Search citations, 3.83% (jul 18 to aug 7) to 0.52% (aug 14 to 17) · labeled provisional by the authors; trailing share has since recovered to about 1.5% 09 [ChatGPT's Reddit Citations Collapsed 95%: The Mechanism Explained](https://www.qwairy.co/blog/chatgpt-reddit-citations-collapse-august-2026) Qwairy, august 2026 · site: queries rose from near 0% to 23–24% of ChatGPT's background searches from august 8, redirecting citations from discussion platforms to official sites · the 95% is measured on Qwairy's own fixed panel of brands, before aug 8 vs after aug 14; the 86% we cite is Promptwatch's, on a different panel and window 10 [Introducing the Agentic Resource Discovery specification](https://commandline.microsoft.com/agentic-resource-discovery-specification-ard/) Microsoft, june 2026 · draft open spec developed with Google, Hugging Face, GitHub, Snowflake and others · [spec](https://agenticresourcediscovery.org/spec/) 11 [Plugins in ChatGPT and Codex](https://help.openai.com/en/articles/20001256-plugins-in-chatgpt-and-codex) OpenAI Help Center · as of july 9, 2026 the app directory is migrated to the plugin directory 12 [Amazon blocked Meta's Muse. Then Shopify wired it into every store.](https://thenewstack.io/amazon-meta-muse-block/) The New Stack, september 2026 · Muse launched september 8; Amazon blocked it, and Shopify enabled agentic checkout with Shop Pay through a declared payment path 13 [Viral AI agent Instinct raises $1B Series C at a $10B valuation](https://techcrunch.com/2026/09/28/viral-ai-agent-instinct-raises-1b-series-c-at-a-10b-valuation/) TechCrunch, september 28, 2026 cite this paper Finder, I., Elovic, A., Shalev, G., & Yosef, L. (2026). AX is the New AEO. arxiv.org/abs/2609.34951 ### BibTeX ``` @misc{finder2026axnewaeo, title={AX is the New AEO}, author={Ido Finder and Assaf Elovic and Gad Shalev and Liad Yosef}, year={2026}, eprint={2609.34951}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2609.34951}, } ``` # Footer The standard for agent experience. © 2026 [era labs](https://eralabs.ai/) Product - Scan a site - [Journey](https://journey.ora.ai/) - [WebMCP audit](https://webmcp.ora.ai/) - [Directory](https://directory.ora.ai/) Resources - Methodology - Research - Blog - API Docs Company - About - Contact - Privacy - [@oradotai](https://x.com/oradotai)