Do off-site mentions help once an AI agent is reading about you?

What did the first large test of this question find?

It found that being readable beat being talked about. Researchers at a company that sells agent-readiness scores for websites sent AI agents buyer questions about 1,056 real businesses (Finder and colleagues (opens in a new tab)).

Each question named the business and asked about its pricing, features or setup. The agents could search the web and open pages before answering. In all, the team ran 37,927 of these research runs across four different agent setups.

They split businesses into two groups: sites an agent could read easily, and sites it could not. The groups were matched on fame, on how much the AI already knew about the brand, and on two stand-ins for off-site optimization. One counted mentions on third-party review and comparison sites, Wikipedia and Reddit; the other scored how findable the business was.

With readability held fixed, neither had a confirmed link to how the answer was built, whether the business was recommended, or what the run cost.

What happens when an agent cannot read your site?

It answers anyway, using other people’s pages. In about 99% of runs that hit a dead end on a site, the agent still produced an answer from whatever it found elsewhere.

When the site was readable, the agent answered from the business’s own pages alone in 78% of runs, against 56% when it was not. As readability fell, agents ran more web searches, from 1.8 to 4.5 per run on average. Each extra search is another chance to pick up information the business never published.

The answers changed in tone too. When agents could not read the site, they were 4.4 times more likely to say they could not access the business. Readable businesses were clearly recommended in 20% of answers, against 11% for the rest.

Two AI judges from different companies had to agree before an answer counted as a clear recommendation.

Does readability change what the agent gets right?

Yes, mainly by reducing what it leaves out. The researchers graded answers about 131 businesses against facts captured from each site. The table compares answers about the same business, agent and question.

Answers built from elsewhere were 3.7 times more likely to contain none of the facts the buyer asked for. The main failure was omission, not invention: facts never mentioned rose from 29% to 45%, while facts stated wrongly rose only from 4% to 6%.

One caveat matters. Compared by group rather than by source, accuracy differed by under two points, 51.7% against 49.8%, which was not a reliable difference. Reading the site helped; being in the readable group made reading more likely but did not guarantee it.

So do off-site mentions not matter at all?

They matter, but at an earlier step. This study only asked about businesses by name, so it tested what happens after a brand is already on the agent’s radar.

Getting named in an open question is different. In our brand study, we counted the independent sites naming a brand in the pages an assistant cited. Each tenfold increase went with 4.7 times the odds of the brand being recommended.

A large vendor-authored study of 102 brands also found off-site pages dominate open questions. Only 2.9% of citations pointed at the tracked brand’s own website (Kumar (opens in a new tab)). The agent study’s authors frame it the same way: being findable helps a business get named; a readable site decides what the answer says once the agent looks.

How ready are most websites for agents today?

Not very. In our study of the top 10,000 websites, we asked each homepage for a Markdown version, a plain-text format agents read easily. Only 3.2% of live sites returned one.

Some sites also shut agents out on purpose. In our crawler study, 15.2% of top sites with a readable robots.txt file, the file that tells crawlers what they may visit, blocked OpenAI’s training crawler. A smaller 7.4% blocked the crawler ChatGPT uses for search.

Why does it matter that your site agrees with itself?

Because AI answers tend to repeat whatever your own sources say, including their contradictions.

In our study of business facts in AI answers, we compared phone numbers in AI answers with each business’s Google profile. Where the profile number did not appear on the business’s own website, 30.6% of the numbers AI engines gave differed from the profile. Where it did appear, 1.6% did.

What should you do about it?

Split the work into getting named and getting read, and fund the second as seriously as the first.

1. Check whether AI agents can read your key pages. Load them with scripts turned off; if pricing, features or setup steps disappear, agents may miss them too.
2. Review your robots.txt and bot protection. Make sure you are not blocking the search and user-requested agents you want to reach you. For Google’s case, see whether blocking Google-Extended hurts AI Overviews.
3. Publish the facts buyers ask about in plain text: prices, plans, features and setup steps, on pages an agent can reach in one or two clicks.
4. Keep facts consistent across your own site and profiles, so an agent finds one answer.
5. Keep earning independent coverage, but do not expect it to fix what agents say once they read about you.

If you want help making your site readable to AI agents, see our generative engine optimization service.

What does the research not tell us yet?

No study has improved a site’s readability and measured the change in agent answers before and after.

• The main study is by a vendor that sells agent-readiness scoring, and it used its own scoring tool to define readable sites.
• Off-site work was measured by only two rough stand-ins. Richer measures of reputation might behave differently, and one showed a weak, unconfirmed hint of helping recommendation.
• The two groups were not matched on how much content their sites publish, which the authors flag as a gap.
• The businesses skewed toward software and commerce, in English, with fact-lookup questions only.
• How readable sites perform in open category questions, where no business is named, was not tested.

Frequently asked questions

Do mentions on other sites help AI agents recommend my business?

Not once the agent is researching you by name, in current evidence. In a study of 1,056 businesses, two stand-ins for off-site optimization had no confirmed effect, while site readability did.

What happens if an AI agent cannot read my website?

It answers from other pages. In about 99% of runs that hit a dead end, the agent still answered, and businesses with hard-to-read sites were clearly recommended 11% of the time, against 20%.

Are AI agents more accurate when they read my site?

Yes. Answers built from the business’s own site got 48.3% of the asked facts right, against 34.3% for answers built from other pages.

Is PR still worth it for AI search?

Yes, for getting named. In our brand study, more independent coverage went with 4.7 times the odds of being recommended per tenfold increase.

Sources

• Finder, Elovic, Shalev and Yosef (2026), AX is the New AEO (opens in a new tab), arXiv:2609.34951.
• Kumar (2026), Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines (opens in a new tab), arXiv:2606.20065.
• Underneath (2026), Do Wikipedia and schema make AI assistants recommend a brand?
• 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?