The short version
- Across 37,927 agent sessions on 1,056 real businesses, agent-ready businesses were clearly recommended 1.9 times as often, in a study by ora research (opens in a new tab).
- When the site was readable, 78% of the answer came from the business’s own pages; when it was not, 58% did.
- Answers built from the business’s own site got 48.3% of the asked facts right, against 34.3% for answers built from elsewhere.
- Few sites are set up for agents: only 3.2% of top websites served a Markdown version on request in our study of 5,902 sites.
What did the study find?
Agent-ready businesses were clearly recommended almost twice as often. The gap held in every agent setup and every business category tested.
Finder and colleagues at ora research ran 37,927 agent sessions. An AI agent here means an assistant that searches the web, opens pages and reads them before answering. Each session was a buyer question about one business: its pricing, its features or how to set it up.
The 1,056 businesses were split into two groups by how readable their sites were to agents. The groups were matched on fame, on how much AI systems already knew about each brand, and on how often the brand was mentioned on other sites. Two AI judges from different companies rated every answer, and an answer counted as a clear recommendation only if both gave it the top score.
What does “agent-ready” mean?
It means an agent can fetch your pages and read the content as text, without being blocked. The study used ora’s own scoring tool to measure it.
The score checks things such as how much content survives without JavaScript, the code that builds many pages in the browser after they load. It also checks structured data, Markdown versions of pages, whether pricing and documentation are reachable, and whether bot controls let in agents acting for a user. The authors write that across nearly 100,000 sites they have scored, fewer than 1% earn the tool’s top grade.
Our own measurements show how early this is. In our study of the top 10,000 websites, 45.8% of homepages carried JSON-LD structured data, a machine-readable description of the page. In our crawler study, 15.2% of top sites blocked OpenAI’s GPTBot from the whole site.
What happens when an agent cannot read your site?
It answers anyway, using other people’s pages instead of yours. In about 99% of sessions that hit a dead end on a site, the agent still answered.
For agent-ready businesses, 78% of the answer came from their own pages and 12% from web search. For the others, own-site content fell to 58% and web search rose to 25%. The share from the AI’s built-in knowledge stayed small either way, at only 7–10% of the answer.
The agents also searched more. Across the readiness scale, average web searches per session rose from 1.8 on the most readable sites to 4.5 on the least. In one example in the paper, an agent blocked twice by a software company’s site ended its answer on two competitors’ blogs.
The tone of answers changed too. When agents could not read a site, they were 4.4 times as likely to say they could not access the business. They were 3.0 times as likely to vouch for it using secondhand sources.
Are answers less accurate without your website?
Yes, mostly because facts go missing, not because they are made up. Comparing answers about the same business, site-built answers were 41% more accurate.
Site-built answers got 48.3% of the asked facts right, against 34.3% for answers built from other sources. Answers built elsewhere were 3.7 times as likely to contain none of the facts the buyer asked for. Wrong facts rose only from 4% to 6%, while facts never mentioned rose from 29% to 45%. Our guide on whether agents invent or omit facts looks at where the wrong ones come from.
| Fate of each asked fact | Built from the business’s site | Built from elsewhere |
|---|---|---|
| Stated wrong | 4% | 6% |
| Never mentioned | 29% | 45% |
Pricing showed the biggest gain from reading the site, and setup questions showed none.
Our own studies point the same way. In our business facts study, AI answers gave a phone number that differed from the Google profile 30.6% of the time when the profile number was missing from the business’s own website. When it was there, the figure was 1.6%.
Does the effect hold across agents and industries?
The direction held everywhere, but the size varied a lot. Baseline recommendation rates differed sevenfold between agent setups.
The lowest, Claude Code, clearly recommended 5% of the time; the highest, eve on GPT-5.4, 36%. Each recommended agent-ready businesses more. Across the twelve business categories, the lift ranged from 1.2 times in customer service to 2.5 times in IT infrastructure.
The study also tested whether being talked about elsewhere mattered once readability was known. Across all 1,056 businesses, neither of the two measures of off-site mentions predicted recommendation, while readability did.
That is one study’s result. In our brand entity study, independent coverage in the pages assistants cited was the strongest predictor of recommendation we measured.
How much should you trust this study?
Treat it as strong early evidence with clear caveats. The authors sell the scoring tool they used, and they say so.
- Vendor study. ora built the readiness ranker and reports its scores “as the study’s instrument rather than as an independent measure.”
- Content depth not matched. A business that blocks agents may also publish less. The authors say the recommendation gap “cannot rule this out.”
- AI judges. Recommendations were rated by two AI systems, not by people, though the two agreed within one point on 88% of answers.
- Accuracy on a subset. Facts were graded for 131 businesses. Between the two groups as a whole, accuracy differed by under two points (51.7% against 49.8%), which was not a clear difference.
- Scope. Mostly software and commerce businesses, in English, asking about business facts.
What should you do about it?
Make sure an agent that reaches your site can read it, then keep earning the visits through search. Practical steps:
- Load your key pages with JavaScript turned off. If pricing or product facts disappear, agents probably cannot see them.
- Check robots.txt and bot-protection settings for the AI crawlers and user-triggered agents you want to admit. In our crawler study, none of the 123 top-site pages ChatGPT cited that day was closed to its search crawler. Opting out of Google’s AI training is a separate call, weighed in whether blocking Google-Extended hurts AI Overviews.
- Publish prices, plans and core facts as plain text on your own site, in one consistent version.
- Add an Organization record in structured data, and consider offering Markdown versions of key pages.
- Keep doing search work. Agents still reach most pages through a web search, as the authors note.
If you want an outside audit of how agents read your site, see our generative engine optimization service.
What does the research not tell us yet?
No one has yet fixed a site’s readability and measured its recommendations before and after.
- Cause and effect. The study compared different businesses, so publishing depth or product quality may explain part of the gap.
- Independent replication. The only large study comes from a vendor using its own scoring tool.
- Other agents and tasks. Four agent setups were tested; buying, booking and other actions were left for later work.
- Which fix matters most. The score bundles many checks, so the effect of Markdown, structured data or bot settings alone is unknown.
Frequently asked questions
Do AI agents recommend businesses whose websites they can read?
In the largest study so far, yes: agent-ready businesses were clearly recommended 20% of the time, against 11% for others. The study was run by ora research, which sells the readiness tool it used.
Does blocking AI bots hurt my brand in AI answers?
It can shift the answer to other sources. When agents could not read a site, only 58% of the answer came from the business’s own pages, against 78% when they could.
Do AI agents make up facts about businesses they cannot read?
Mostly they leave facts out. In the ora study, wrong facts rose only from 4% to 6%, while facts never mentioned rose from 29% to 45%.
How many websites are ready for AI agents?
Very few by any measure. Only 3.2% of top websites served Markdown to agents in our study, and ora reports fewer than 1% of the sites it scored earned its top grade.
Sources
- Finder, Elovic, Shalev and Yosef (2026), AX is the New AEO (opens in a new tab), arXiv:2609.34951.
- Underneath (2026), Do websites serve Markdown to AI agents? 2026 data
- Underneath (2026), Which AI crawlers do top websites block? 10,000 sites, 2026
- Underneath (2026), Do AI answers match a business’s Google profile?
- Underneath (2026), Do Wikipedia and schema make AI assistants recommend a brand?