Do on-site AEO signals predict appearing in AI answers? We measured 62 businesses
No, not in this study. Across 62 local businesses in Japan, our on-site AEO score (structured data, llms.txt, crawler access and similar signals) showed no relationship with how often Perplexity named or cited a business when asked for recommendations: Spearman ρ = 0.03 (95% CI −0.23 to 0.28, p = 0.834). With 62 businesses the study could reliably detect only correlations of about |ρ| ≥ 0.35, so a weaker relationship cannot be ruled out.
Published 2026-10-28. Data measured 2026-10-24. Nurevo makes AEO tools, including the score tested here; see "Our conflict of interest" below.
Summary
- Question: do businesses whose websites score higher on Nurevo's AI-readability score (0 to 100) appear more often in AI answers to questions that do not name them?
- Answer: not supported. ρ = 0.03 overall, and still 0.03 after adjusting for industry and city.
- Overall, businesses appeared in 21.0% of answers to discovery questions; 25 of the 62 appeared at least once.
- No individual part of the score (structured data, llms.txt, coverage and so on) showed a relationship either, after correcting for multiple tests.
- What we take from it: structured data helps AI read a business's facts correctly, but on its own it does not decide whether a business is mentioned. Other factors, very likely including what other sites say about the business, matter more. That last point is a hypothesis this study did not test.
Why we ran this study
AEO (answer engine optimization) advice, including ours, says that making a website easy for AI to read (schema.org structured data, an llms.txt file, letting AI crawlers in, server-rendered content) helps a business show up in AI answers. Nurevo sells a score built on exactly those signals. We wanted honest evidence of whether the score tracks what owners actually care about: being named by AI. We fixed the hypothesis, the analysis and the decision criteria before collecting any answers, and committed to publishing the result either way.
Method
- Sample: 72 local businesses with their own official websites, chosen by web search (portal sites excluded) across 5 industries (dental clinics, restaurants, beauty salons, tax accountants, lodging) and 4 Japanese cities (Tokyo, Osaka, Fukuoka, Sapporo). 10 were excluded because the score could not be computed for their site, leaving 62 for analysis.
- Score: each site's home page, robots.txt and llms.txt were scored with the same code as Nurevo's public score: structured data (30 points), coverage (25), legibility (20), llms.txt (15) and consistency (10), with a gate for crawler access, edge blocking and server rendering (failing it caps the score at 15).
- Engine: Perplexity, through its Agent API with the "fast" preset (one web search, then an answer), the same adapter Nurevo's paid plan uses for AI visibility. One engine only; see the limitations.
- Questions: three "discovery" questions per business that name the area and the kind of business but not the business (for example, "Recommend a dentist in Umeda"), shared by every business of the same industry in the same area, plus one "named" question asking for a business's address and opening hours. Each question was asked three times.
- Outcome: a business "appeared" in an answer if its name was in the answer or its own domain was among the cited sources. The appearance rate is the share of its discovery answers in which it appeared.
- Detector check: in answers to the named questions the business's name was detected 84.3% of the time, above the 70% threshold we set for trusting the detector.
- Analysis: Spearman rank correlation with a permutation p-value and bootstrap confidence interval; partial correlation with industry and city as fixed effects; a permutation test within strata; and a logistic regression with industry and city fixed effects and a quasi-binomial correction, because repeated answers to the same question are not independent. The six component tests are Holm-corrected.
The criterion for "supported" was set in advance: ρ ≥ 0.3, p < 0.05, a 95% CI above zero, and a positive adjusted correlation. Answer text was not stored; only whether a name appeared, whether the domain was cited, and the citation position.
Results
The verdict is "not supported". Every test of the score against appearance gives a correlation close to zero:
| Test | Result | p-value |
|---|---|---|
| Score vs appearance rate (Spearman ρ, primary) | 0.03 (95% CI −0.23 to 0.28) | 0.834 |
| Same, adjusted for industry and city (partial correlation) | 0.03 | 0.845 |
| Same, permutation test within strata | 0.03 | 0.829 |
| Score vs name appearing (discovery questions) | −0.04 | 0.782 |
| Score vs own site cited (discovery questions) | 0.18 | 0.149 |
| Score vs own site cited (named questions) | −0.03 | 0.813 |
| Logistic regression with industry/city fixed effects, quasi-binomial: odds per +1 SD of score (18.81 points) | ×1.01 (95% CI 0.62 to 1.64) | 0.972 |
The largest coefficient, 0.18 between the score and the business's own site being cited, is not distinguishable from zero (p = 0.149) and should not be read as a finding.
By score band
Nurevo shows the score as red, yellow or green. Most sites in the sample were red, and none reached green, so this study says nothing about very high scores.
| Band | Businesses | Mean appearance rate | Appeared at least once |
|---|---|---|---|
| Red | 54 | 20.8% | 22 |
| Yellow | 8 | 22.2% | 3 |
| Green | 0 | - | - |
By component of the score
None of the parts of the score was related to appearance. After Holm correction every p-value is 1.000. These tests were exploratory.
| Component | Spearman ρ with appearance rate |
|---|---|
| Structured data (schema) | 0.08 |
| Coverage | 0.12 |
| Legibility | −0.17 |
| llms.txt | −0.02 (odds ratio 1.41, 95% CI 0.79 to 2.52) |
| Consistency | 0.06 |
| Gate (crawler access, edge, server rendering) | 61 passed: 21.3% appearance; 1 failed: 0% |
By industry and by city
Appearance varied by industry and by city more than by score. These are descriptive figures from small groups, not tested differences.
| Group | Businesses | Mean score | Appearance rate |
|---|---|---|---|
| Tax accountants | 13 | 27.92 | 10.3% |
| Lodging | 13 | 21.08 | 23.9% |
| Dental clinics | 13 | 40.15 | 25.6% |
| Beauty salons | 8 | 24.25 | 8.3% |
| Restaurants | 15 | 24.20 | 30.4% |
| Osaka | 15 | 26.27 | 15.6% |
| Sapporo | 17 | 29.41 | 9.8% |
| Tokyo | 15 | 25.53 | 31.1% |
| Fukuoka | 15 | 29.27 | 28.9% |
Limitations
- Cross-sectional: one measurement of many sites shows association, not cause. Whether improving a site changes its appearance needs a before-and-after study of the same sites.
- Statistical power: with 62 businesses, a correlation weaker than about |ρ| = 0.35 would often be missed. Businesses in the same area and industry share the same answers, so the effective sample is smaller still. "Not found" here does not mean "does not exist".
- One engine, one week: Perplexity with the "fast" preset, measured on 2026-10-24. ChatGPT, Gemini, Google's AI Overviews and Copilot use different search indexes and models; these results do not transfer to them.
- Unobserved confounders: reviews, brand recognition, advertising, years in business and listings on other sites were not measured. Only industry and city were adjusted for.
- Sample: businesses were found through web search, so they lean toward businesses that are already easy to find. No site scored green. 5 industries and 4 cities in Japan may not represent other industries, regions or languages.
What it means for a business website
Structured data, llms.txt and crawler access still do a job: they let an AI read a business's facts (address, opening hours, phone number, services) directly and correctly instead of guessing. That job is worth doing, and it is a different job from being recommended. In this study, a better-read site was not a more-mentioned one.
Our working hypothesis is that sources outside the website (review and map listings, local media, directories, other sites that mention the business) weigh heavily in whether an AI names a business when it answers a recommendation question. This study did not measure those sources, so this is a hypothesis to test, not a result.
In practice: keep the on-site basics in order because they make what AI says about you accurate, and measure whether AI actually mentions you rather than inferring it from a readability score.
Our conflict of interest, and what we are changing
Nurevo makes AEO tools, and the score tested here is one of our products. A result that the score does not predict appearance is not one we would have chosen, and we are publishing it because the criteria were set before the data came in and because owners should know what a score can and cannot tell them.
What we are doing about it: we do not present the AI-readability score as a predictor of appearing in AI answers. We measure appearance directly instead: the Pro plan measures how often Perplexity names or cites a site (beta), and we are adding a small monthly appearance check for Free and Standard sites. We are also developing a second version of the score that blends measured AI visibility with the on-site signals, rather than relying on the on-site signals alone.
The names and websites of the businesses measured are not published.
FAQ
Does structured data make AI mention my business?
In this study of 62 local businesses on Perplexity, no relationship was found between on-site signals such as structured data and how often a business was mentioned (ρ = 0.03). Structured data helps AI read your facts correctly, but on its own it did not predict being mentioned.
Does llms.txt help a business appear in AI answers?
We found no relationship in this study: ρ = −0.02 between the llms.txt component and the appearance rate (Holm-corrected p = 1.000). The study was small and used one engine, so this is not proof that llms.txt never matters.
Does this mean AEO is pointless?
No. On-site AEO makes what AI says about a business accurate. This study suggests it is not sufficient to be recommended, and that being mentioned should be measured directly rather than inferred from a readability score.
Which AI engine was measured?
Perplexity only, through its Agent API with the "fast" preset, on 2026-10-24. The results do not transfer to ChatGPT, Gemini, Google's AI Overviews or Copilot.