How much of what we measure about the global Internet actually reflects the underlying architecture of the network itself, and how much is merely an artifact of the tools we use to collect those measurements? This fundamental question sits at the heart of modern internet research. Academics, security analysts, and policy researchers frequently rely on Virtual Private Networks (VPNs) to gather data through connections anchored in different countries and economies. This methodological approach commonly operates under the underlying assumption that any VPN provider operating within the same economy will provide a roughly equivalent, homogenous view of the local digital landscape.
However, groundbreaking new research reveals that this long-held assumption may be fundamentally flawed. A recent study shows that switching between different commercial VPN providers can dramatically alter which specific servers a web browser connects to, as well as the geographic economies in which those servers are actually located—even when the connection originates from the exact same country and city. These provider-dependent variances can easily lead academic studies and policy analyses to significantly overestimate or underestimate how much web traffic genuinely stays within a given domestic economy, how much crosses international borders, and which global economies appear to host the critical infrastructure serving users.
The implications of these findings extend far beyond academic curiosity. It means that a research study utilizing one particular VPN provider might confidently conclude that a nation’s web traffic is largely served by domestic infrastructure, while another study using an alternative provider could make that exact same economy appear heavily dependent on foreign-hosted infrastructure. Consequently, the choice of a measurement tool can shape not just raw data collection, but also major policy conclusions regarding data localization mandates, cross-border data exposure, and the true geographic distribution of the global Internet.
How researchers tested whether VPN choice changes Internet measurements
To investigate the true impact of VPN selection on internet measurements, researchers developed an advanced system known as VPN-Scope. This custom-built platform was designed to visit identical sets of websites through multiple different commercial VPN providers at nearly the same exact time, utilizing server connections located in the same economy and, whenever technically possible, the exact same city.
The comprehensive study covered four major commercial VPN providers—ExpressVPN, NordVPN, ProtonVPN, and Surfshark—across 14 distinct economies spanning Europe and the Americas. To ensure a representative sample of web traffic, the researchers selected 1,000 websites per economy, blending globally ranked websites with locally popular platforms and official government web portals. Each of these selected websites was scheduled for 25 distinct visits through each of the four providers over a continuous three-week testing period conducted in March and April 2025.
For every single recorded visit, the researchers meticulously tracked the exact servers contacted as the web page loaded, which included the backend servers supplying crucial embedded resources such as style sheets, scripts, fonts, and tracking elements. Following data collection, the team examined the geographic locations and hosting organizations of these servers. Additional specialized experiments were conducted to investigate Domain Name System (DNS) resolution behavior and to identify the precise physical locations serving content through the various VPN routes. Together, these rigorous tests helped the researchers connect the observed data discrepancies to the underlying technical mechanisms producing them.
How much VPN providers changed the measured Internet infrastructure
The profound extent of these discrepancies became immediately clear when analyzing measurements gathered from Chile. In the Chilean trials, just 9.5 percent of the analyzed web requests routed through ExpressVPN were associated with server IP addresses geolocated directly within Chile. By sharp contrast, the corresponding domestic shares were vastly higher when routing through other providers, reaching 62.7 percent through Surfshark and 64.2 percent through NordVPN. The cross-border distribution across other destination economies shifted just as drastically; for instance, Brazil accounted for 57.1 percent of destinations when measured via ExpressVPN, compared to a mere 2.4 percent through NordVPN.
These stark differences emerged despite measuring the exact same set of websites from the exact same source economies during the same time frame. The results underscore that a VPN provider cannot be treated as a mere implementation detail or a simple utility for acquiring an IP address in a target country. Changing the provider fundamentally alters the geographic conclusions that a researcher would naturally draw from otherwise comparable measurements.
A researcher relying on NordVPN data would observe that the vast majority of analyzed requests are associated with domestic servers inside Chile, whereas a researcher utilizing ExpressVPN would observe the exact opposite outcome. For public policy studies intended to inform data-localization regulations, that discrepancy changes the apparent national reliance on domestic versus foreign infrastructure, even though these surface-level network measurements alone cannot establish where user data is ultimately stored at rest.

Similar divergences appeared throughout the researchers’ DNS experiments. When comparing two different providers operating within the same economy, the sets of returned IP addresses differed in approximately 12 percent of comparisons for initial website root domains. For dependency domains—the third-party services responsible for supplying scripts, fonts, analytics trackers, and other modular page components—the rate of disagreement rose sharply to approximately 31 percent. Looking exclusively at the primary website domain typed into a browser address bar can therefore obscure a substantial portion of the underlying network variation.
Despite these differing IP addresses, the divergent paths did not usually translate to entirely different hosting companies. When evaluated at the organizational level, the compared DNS answers shared at least one common hosting organization in 98 percent of initial-domain comparisons and 96 percent of dependency-domain comparisons. This represents an organizational overlap rather than an identical set of hosting entities. The broader finding is that VPN choice exerts a much stronger influence on our geographic picture of where content is served than on our understanding of which specific organizations are hosting it.
What researchers can control to make VPN measurements more reliable
While some sources of experimental variation are difficult to eliminate entirely, others can be substantially mitigated through rigorous and careful experimental design. Researchers can independently verify the true physical locations of VPN exit nodes, ensure that connections are compared from the same metropolitan area whenever feasible, and maintain strict consistency across browser settings and measurement timing schedules. Furthermore, researchers must thoroughly document the specific provider, connection configurations, and network protocols used, rather than describing a measurement dataset loosely as originating "from Italy." The specific DNS service handling website-name lookups also requires careful validation, as selecting the same public resolver service in different configurations does not guarantee that every VPN provider will actually route queries through it.
Other significant differences, however, stem from autonomous decisions made by the underlying networks themselves. Content Delivery Networks (CDNs) strategically distribute cached copies of web content across numerous physical locations globally, and they routinely select entirely different edge servers based on the specific network path a request originates from. Network routing protocols can also direct two distinct connections to entirely different physical facilities even when they utilize the exact same destination IP address.
In targeted tests conducted by the research team, VPN connections originating from Milan successfully reached content-serving locations situated in either Milan or Frankfurt. Sharing a starting city—and even sharing an identical destination address—was therefore insufficient to guarantee a shared physical destination. These routing decisions depend heavily on complex inter-provider network agreements and traffic engineering practices that researchers cannot standardize simply through browser adjustments or VPN settings alone.
Taking these comprehensive findings into consideration, the study suggests that the commercial VPN provider must be treated as an integral component of the experimental design, rather than merely treated as a transparent tool for obtaining a network presence in a targeted economy. Researchers should explicitly report which provider and connection parameters were utilized, independently verify where the VPN exit node is physically located, and, when geographic accuracy is critical to the research question, consider replicating measurements across multiple distinct providers.
Even these rigorous methodological precautions cannot eradicate all observed variance. Underlying DNS infrastructure behavior, CDN traffic steering mechanisms, and dynamic network routing can still cause two separate VPN connections originating from the same economy, or even the same city, to reach entirely different servers or physical data center locations. The significance of this variation ultimately depends upon the specific research questions being asked.
The study’s extensive measurements indicate that high-level conclusions regarding the specific economy an endpoint resides in, whether network traffic appears to remain domestic, and which physical server replica actually serves the content are exceptionally sensitive to the choice of VPN provider. Conversely, conclusions regarding which hosting organizations are involved in delivering the content tend to remain far more stable across different providers.
For academic and policy studies examining cross-border data flows, regulatory data localization, or the broader geographic distribution of global Internet infrastructure, any empirical result obtained through a single VPN provider should be carefully interpreted as representing just one perspective of the network, rather than a provider-independent picture of an entire national economy.