AX is the new AEO

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.

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.

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.

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.

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.

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.

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.

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.

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.

• juneMicrosoft, 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.
• julyOpenAI replaced ChatGPT's app directory with a plugin directory, now the main place to find what ChatGPT can connect to.
• septemberMeta 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.
• septemberInstinct 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.

1. 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.
2. 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.
3. 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.
4. 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.
5. 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. The analysis code and a sample of the derived data are released at github.com/ora/research.

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; 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.

09 / sources

@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},
}