A Politecnico di Milano preprint reports substantially more concentrated visible source exposure on Baidu Chinese than Google Chinese. Auditing Source Exposure in Baidu and Google AI Search was submitted September 21, 2026, at 10:57 UTC; its arXiv record states acceptance at WAC @ EMNLP 2026. This is an academic audit, not a search-engine update.
The paper analysed 1,998 observations from 500 English MS MARCO queries and translated Chinese versions, collected in Milan between June 20 and July 10. Overview rates were 97.6% for Google Chinese, 97% for Google English, 82% for Baidu Chinese and 4% for Baidu English.
Host-level Gini coefficients were 0.895 for Baidu Chinese and 0.500 for Google Chinese. The leading host accounted for 27.1% versus 6.6% of host occurrences, with Baidu-owned subdomains prominent. Aggregate host-set Jaccard overlap was 3.2% between the Chinese settings and 0.8% between Baidu Chinese and Google English.
These are visible-host measurements, with repeated hosts deduplicated per answer. Collection location, logged-in Google state and translated queries limit generalisation. Semantic similarity does not establish accuracy or citation support, and aggregate overlap does not describe each matched answer’s sources.
Do not assume a translated strategy transfers
Our editorial interpretation is that international GEO requires separate observation in each intended market. The study does not prove that a website visible on Google will be invisible on Baidu. It does provide a reason to test that assumption rather than treating translation as sufficient evidence of distribution.
A business entering another language market should first define the audience’s information needs and search context. Direct translations can support comparisons, while locally developed queries can reveal different expectations. Keep those query groups distinct so a report makes clear whether it is comparing equivalent wording or actual market demand.
For a proposed audit, inspect both the answer and its visible references. A useful summary can mention a business without citing its website; a cited page may appear without the brand being discussed. Those outcomes deserve separate records because they answer different questions about exposure.
Measure activation, sources and concentration separately
Begin by recording whether an answer appears at all. Then count the distinct visible hosts and examine how often each appears across the tracked set. Define how subdomains are handled before calculating comparisons: several services under one owner can look diverse at one level and concentrated at another.
Report the measurement rule alongside the result. A share of host occurrences should not be relabelled as a share of users or answers. An aggregate set comparison should not be used to claim that every pair of responses draws on different sources. These distinctions help a commercial report remain faithful to what was actually observed.
Concentration can also guide a closer editorial review. Ask what kinds of sources appear, whether they address the query and whether their claims are current. Frequent appearance alone does not establish authority or reliability. Source prominence and source quality require different checks.
Build a market-specific visibility baseline
Our recommendation is to repeat observations under documented conditions and maintain a stable query group. Record device, location, account context, collection time and what parts of the source interface were inspected. When conditions change, annotate the comparison rather than presenting the movement as an unexplained platform trend.
For businesses, use that baseline to identify where accurate owned information and independent coverage are missing or inconsistent. Prioritize useful local-language material and genuine distribution opportunities supported by the observed results. Avoid assuming that activity inside a platform’s ecosystem guarantees selection.
The practical lesson is to evaluate cross-market AI visibility through evidence gathered for that market. Similar wording in two answers may reassure a reader about the message, but a publisher still needs to understand whether its material receives exposure and which sources mediate that exposure.