The AI agent reads your site before it recommends you. If it can’t, it answers anyway.

When an AI agent cannot read a business's website, it does not give up: it answers anyway, from whatever else it finds, and that business is recommended about half as often. A study published on 6 October 2026 shows it with two traces of the same question, put to the same agent. A buyer asks what they can try for free.

On twilio.com, the agent runs one search, opens five of Twilio's own pages, from the trial page to its documentation, and answers from them. Seven steps. On hashicorp.com, the agent is blocked twice by the site's bot protection. It turns to web search seven times, wanders through documentation fragments and a blog post, and ends its journey on two competitors' blogs. It still gives an answer, a confident one.

The study calls the quality that separates those two sites agent experience, or AX: how well a site serves the AI agent that arrives to fetch and read it. Its authors argue that AX now matters more than being talked about on other sites. That claim challenges part of the AEO and GEO playbook, including work agencies like KAIS do, which is a good reason to read it closely.

What did the study measure?

The study, arXiv:2609.34951, ran 37,927 AI agent journeys on 1,056 real businesses, split into two matched groups that differ mainly in how readable their sites are to agents. Its authors, Ido Finder, Assaf Elovic, Gad Shalev and Liad Yosef, work at ora research. One point matters before any number: ora sells the agent-readiness ranker the study uses to sort sites into groups. The authors state it themselves and present the score as their instrument, not as an independent measure.

Half the businesses were classed as agent-ready, with an accessibility score of at least 0.65 out of 1, and half as not agent-ready, at 0.50 or below. The score checks, among other things, how much content survives a fetch without JavaScript, whether bot controls let user-triggered agents through, and whether pricing and documentation pages can be reached. Each agent-ready business was paired with a similar one in the same sector and popularity band. The pairs were matched on fame, on how much the model already knew about the brand, and on two proxies for AEO: how widely the brand is mentioned on third-party sites, and a findability score.

Each business received three buyer questions, on pricing, features and setup, run three times on four different agent stacks. These ran on Anthropic's claude-sonnet-4-6 and claude-haiku-4-5 with Anthropic's own search, and on OpenAI's gpt-5.4 and gpt-5.4-mini with the Tavily search API. Every question named the business's domain. The study therefore measures what happens once an agent is already looking at a given business, not whether that business gets surfaced among competitors.

Do AI agents still answer from what they learned in training?

Barely: in the four agent stacks tested, only 7% to 10% of the finished answer came from the model's training knowledge, whether the site was readable or not. The rest came from pages and search results fetched during the journey.

The authors also traced the shift across OpenAI releases, putting 90 buyer questions to each model with web tools optional. GPT-4.1 still built about half of its answer from memory, 52%. By gpt-5.6, that share had fallen to 14%. The study used an AI judge to make that call, and the authors note that memory still handles stable facts in a plain chat. But for questions about prices, plans and setup, what the agent reads during the journey now decides most of the answer.

What happens when an agent cannot read a business's site?

It answers anyway, and the answer is built without the business: in about 99% of journeys that hit a dead end on a site, the agent still answered, from other sources. For agent-ready businesses, 78% of the finished answer came from their own pages and 12% from web search. For the others, their own pages fell to 58% and web search doubled to 25%. In other words, 42% of what the buyer reads about a hard-to-read business comes from someone else.

The mechanism shows up as a steady slope. Across the range of accessibility scores, the average number of web searches per journey rose from 1.8 on the most readable sites to 4.5 on the least readable. Each extra search is another chance for the answer to rely on outdated pages, competitors' comparisons or aggregators. On less readable sites, agents were also blocked by the site 2.1 times as often (33.5% of journeys against 15.8%), took 23% more turns and 15% longer to answer.

Are readable businesses recommended more often?

Yes: answers clearly recommended agent-ready businesses 20% of the time, against 11% for the others, a ratio of 1.9. Two AI judges from different vendors rated each answer from 0 to 4 without knowing the business's group. An answer counted as a clear recommendation only when both gave the top score.

The gap held on all four agent stacks, even though their baselines differ widely: the claude-code stack clearly recommended 5% of the time against 2%, the gpt-5.4 stack 36% against 20%. It also held in all twelve business categories, from 1.2 times in customer service to 2.5 times in IT infrastructure, and was largest where buyers compare vendors before committing.

The language of the answers shows what happens instead. When the agent could not read the site, its answer was 4.4 times more likely to admit it could not access the information, 3.0 times more likely to vouch for the business through second-hand sources, 1.8 times more likely to stay vague and 1.4 times more likely to send the buyer off to check for themselves. An answer that opens with "the site returned an error" is not a recommendation, whatever follows.

Are answers more accurate when the agent reads the site?

Yes, for the same business and the same question: answers built from the business's own site got 48.3% of the asked facts right, against 34.3% for answers built from the web, 41% more. The researchers captured each site's published facts with a full browser render, then graded 31,127 facts in 2,499 answers about 131 businesses.

The failure is not invention but omission. Facts stated wrongly rose only from 4% to 6% when the answer came from the web. Facts never mentioned rose from 29% to 45%. A web-built answer was 3.7 times more likely to contain none of the facts the buyer asked for. The effect was largest on pricing, +64%, present on features, +18%, and absent on setup, where the answers mostly lived in documentation neither channel reached.

One result points the other way, and the authors report it. Compared group against group rather than answer against answer, accuracy differed by less than two points, 51.7% against 49.8%, which is not significant on 131 businesses. Readiness makes reading the site more likely, but does not guarantee it.

Does AEO still matter once a site is readable?

At the step this study measures, its two AEO proxies showed no effect of their own, while site readability predicted grounding, recommendation and cost. With readability, popularity and sector held fixed, how widely a brand was mentioned on third-party sites had no link to any of the three outcomes. The findability score showed one weak hint on recommendation, below the authors' bar for confirmation.

That result has a precise scope, and it is narrower than the paper's title. Every question named the business, so being surfaced among competitors was never tested. The authors themselves describe answer-engine optimization as two steps: search optimization to be surfaced, and agent experience for the reading that follows. They also note that agents reach most pages through a web search, so what ranks still decides much of what an agent sees. Their conclusion is not that being found no longer matters, but that it is not enough.

How strong is the evidence?

The main contrasts are large and consistent, but the study comes from a vendor and leaves one confound open. The table reports the headline comparisons as published.

The authors list their limits. The groups were not matched on how much each site publishes: a business that blocks agents may also publish less, which would widen the gap on its own. Only a before-and-after study, improving a site and measuring again, would settle that. The cohort leans towards software and commerce, mostly in English, and the questions are factual lookups. Recommendations and answer sources were judged by language models. And the readiness score that defines the groups is the authors' own product.

What should a business check on its own site?

Whether an agent sent to the site comes back with the facts a buyer asked for. What follows is the KAIS reading of the study, turned into checks a business can run. The study scores readiness as a whole; it does not isolate the effect of each item below.

First, whether the content exists without JavaScript. Most agents do not run scripts by default, so a page assembled in the browser after loading can be nearly empty to them. The test is simple: fetch the page's raw HTML and see whether prices, plans and key facts are in it.

Second, whether the site lets agents in. In the HashiCorp trace, the pricing page returned a bot block twice. A firewall or bot-protection rule can refuse an agent that robots.txt allows, a gap described in Your robots.txt says yes. Your firewall says no.

Third, whether the facts buyers ask for are written on reachable pages. Pricing showed the largest accuracy gain when agents read the site, setup the smallest, because setup facts were buried in documentation agents rarely reached. A price that only appears in a quote form, or a feature list rendered as an image, leaves the agent to look elsewhere.

Fourth, whether each page is easy to read once fetched. Clear sections with headings that describe them help the right passage be found, as a Notre Dame study showed earlier this month, discussed in Page structure helps AI find your content.

How can a business measure what agents do on its site?

By running the buyer's questions through an agent and reading the trace, not only the answer. The study's method is observable from the outside: did the agent open the business's pages, was it blocked, how many times did it search the web, which sites did it end on, and did the answer hedge.

That adds a layer to AI visibility measurement. KAIS keeps questions that name the brand out of its visibility score, because a question that names a brand always finds it. This study shows that those questions still deserve a separate test. Once a buyer names the business, the outcome depends on whether the agent can read the site. A drill-down test with branded questions, run a few times on two or three agents, shows whether the site answers for itself or leaves the job to others.

Where does this sit in the KAIS AI Visibility Chain?

The study sits at the second and third stages of the KAIS AI Visibility Chain, Be Retrieved and Be Understood, applied to the business's own site.

It also bears on Be Trusted. When an agent cannot read a business's site, it relies on pages the business does not control. The authors cite research in which editing a single Reddit or Wikipedia page with as few as 13 words got planted text cited in 38% to 51% of deep-research reports. A readable first-party site is the one source a business can vouch for. And the chain's last stage, Be Recommended, is where the study finds its largest gap, measured with the same caution as in Your brand is cited by AI. Is it recommended?.

Being found opens the door. Being read decides the answer.

The study does not retire off-site visibility. It moves the decisive moment one step later. An agent that has found a business still has to read it, and when it cannot, it does not wait. It asks someone else.

Twilio's answer came from Twilio. HashiCorp's ended on its competitors' blogs. Both answers sounded sure of themselves. Only one was built by the business it described.

Sources

• Finder, Elovic, Shalev and Yosef — AX is the New AEO, ora research, arXiv:2609.34951v3 [cs.AI], 6 October 2026. Study design and covariate balance (section 4, Table 1), agent stacks (Table 2), training-knowledge share (section 3, Figure 1), traces (Figure 3), recommendation (section 5.2, Figures 4–6, Table 3), answer composition and searches (section 5.3, Figures 7–8), cost (section 5.4), accuracy (section 5.5, Table 5, Figure 10), AEO control (section 5.6, Table 6), summary (Table 7) and limitations (section 7). The authors work for ora, whose agent-readiness ranker defines the study's groups. Figures from other sources (the Reddit and Wikipedia poisoning study, Mathias Biilmann's definition of AX) are cited as reported by the authors. All figures come from the paper; none were reproduced by KAIS.