Do AI-Cited Pages Name the Recommended Businesses? Checking 175 Answers
Published
This article was first published in Korean by Intrix Lab on October 7, 2026.
Often, they do not. As of 2026-09-28, we fetched 612 pages cited in 175 AI answers and checked whether the businesses named in the answers appeared anywhere on those pages. For Model A, 24% of business names appeared nowhere on the cited pages; for Model B, 42%. This means a link next to an answer is difficult to trust as evidence for that recommendation.
What did we compare?
A study published on arXiv in September 2026 argued that citations and actual supporting evidence diverge as AI condenses multiple documents into an answer. It claimed that checking against the original sources raises the share with insufficient support to as much as 88%. We used answers collected by Intrix to check whether something similar happens with Korean-language business recommendations.
We checked whether a business named in an answer appeared on at least one page cited by that answer. This is a more lenient standard than the original study's sentence-level assessment. Across 180 research answers covering 6 target businesses, we identified 760 citation links and could read 612 of them. We fixed the assessment criteria before making the comparison.
Did business names appear on the cited pages?
| Business names absent from all cited pages | Share |
|---|---|
| Model A | 69/285 (24.2%) |
| Model B | 238/570 (41.8%) |
Our predefined threshold was 30%. Model B exceeded it, supporting the claim in our data that "citations cannot be considered to support the answer's content." Model A fell below the threshold. Model A generally recommended businesses that appeared on the pages it cited.
Where did Model B get the business names?
It appears to have drawn many of them from memory. Of 214 business mentions in Model B's answers to requests for online store recommendations, 140 were absent from the cited pages. In the same study, Model B named the target brands at a rate of 10.7% even when it answered without searching. Finding pages to cite and recalling business names operate separately in Model B.
What about numbers?
The share of unverified numbers, such as prices, durations, and counts, was higher: 76% for Model A and 50% for Model B. However, these numbers included values the models calculated themselves and general advice such as "look at reviews from the past few months." Mixed-format expressions such as "2만7천 원" (27,000 won) were sometimes only partially read. We therefore do not interpret the numerical results solely as evidence of a mismatch with sources.
What should businesses check?
First, links attached to an answer are not enough to conclude that AI has described your business correctly. Many recommended names are absent from those links. Second, the route by which a business gets named differs by model. For Model A, your name needs to appear clearly on cited pages. For Model B, reputation and external mentions accumulated across multiple sources function as memory.
To check which pages AI uses as evidence and which businesses it recommends in your industry, you can start with a first measurement.
How far can these results be trusted?
The study covered 6 target businesses and answers to Korean-language questions submitted through APIs. We fetched the pages after the answers were generated, so their content may have changed in the meantime. We could not read some pages. Because we checked for names "anywhere" rather than using sentence-specific citation locations, the unverified share could be higher under the original study's method.
Cases where the same business was written differently in English and Korean may have been counted as unverified. We do not extend these results to other industries or consumer app interfaces.
Frequently asked questions (FAQ)
Q. If AI includes a link, doesn't that mean the information is on that page?
Often, that was not the case in this comparison. Of the businesses named by Model B, 42% appeared nowhere on the pages cited by the corresponding answer.
Q. Which model is more accurate?
For business names, Model A aligned more closely with the pages it cited. The unverified share was 24% for Model A and 42% for Model B. However, these results cover 6 target businesses, so we do not generalize them to the models' overall accuracy.
Q. What should we do if our business is recommended but our website is missing from the supporting links?
AI may have recalled your business from memory or another source. Recording which questions lead to recommendations and which pages support them can help you notice if that route stops working.