ExpressVPN today published a study of 102 search queries showing that Google Trends returned different "worldwide" scores for the same query, the same period and the same settings depending on whether the request came from an IP address in the United States, Japan or India. Across one year of weekly data, the three countries agreed on the exact score only 39.9% of the time.

In Short

Google Trends is supposed to give one answer when someone asks how popular a search was around the world, but a VPN company asked the same questions from three countries and often got three slightly different answers. That matters to marketers, journalists and researchers who use Trends to plan campaigns, spot seasonal peaks or back up stories, because the number on their screen may depend on where their computer appears to be. Differences shrank over a five-year view and were usually small, but the study shows that a Trends chart is one sample among several possible ones, not a fixed figure.

What ExpressVPN published

The research appeared on ExpressVPN's website on October 21, 2026, under the title "Consistency in Google Trends: How results can vary across IP addresses." It is credited to the company's "Research team," with no individual authors named, and the page lists a reading time of 30 minutes. ExpressVPN sells VPN subscriptions, and the study was carried out using its own servers, a detail that matters when reading the report's closing recommendations.

The premise is simple. Google Trends does not report raw search volumes. According to ExpressVPN, citing Google's own documentation, the tool "displays relative interest based on an anonymized, categorized, and aggregated sample of actual search requests." Each query is scored from 0 to 100 relative to its peak within the chosen period and place. Because the figure is drawn from a sample, ExpressVPN set out to test how much that sample shifts when only one variable changes: the country from which the request is made.

The company was careful to limit its claims. "Our observations apply only to the specific sample and timeframe included in this study, and we don't recommend extrapolating the results to broader conclusions about global data consistency," the report states. In a separate limitations section, it goes further: while the findings may suggest that location affects the reproducibility of online data, "our sample and methodology are insufficient to draw this conclusion." The study, it adds, "wasn't designed to determine exactly why those differences occur or how broadly they apply."

How the study was built

The query list came from Google's own Year in Search roundups. ExpressVPN combined every category that appeared in those annual lists between 2020 and 2025, filtered the pool to terms it judged "neutral, non-political, and unlikely to trigger censorship or regional filtering," and then randomly selected 100 queries. It added two industry terms of its own, "VPN" and "Antivirus," bringing the total to 102. The resulting list mixes actors such as Joaquin Phoenix and Pedro Pascal, athletes such as Mike Tyson, foods such as tanghulu and mango pickle, songs, sports tournaments and landmarks.

Each query was run twice: once as a search term and once as the matching Google Trends topic. Topics, in Google's definition as quoted by ExpressVPN, are "a group of terms that share the same concept in any language." The topic for each query was the top suggestion Google Trends offered for the term, and the report acknowledges that this selection was manual, so "a small element of human judgment was involved."

The geographic setting in Google Trends stayed at "worldwide" throughout the main study. "The only variable we changed per query? The VPN location from where the search was made," according to the report. ExpressVPN connected to servers in what it describes as "the three countries with the highest Google traffic worldwide: the U.S., Japan, and India," a ranking the web page sources to a third-party site, World Population Review, rather than to Google.

Data was exported as CSV files from a logged-out Chrome Incognito window connected to an ExpressVPN server in each country, then merged with a Python script. The report gives two windows: 52 weeks and 260 weeks. The interactive chart on the page dates the one-year window from April 13, 2025 to April 5, 2026. The report does not state the exact days on which the main exports were pulled, whether the three countries were queried at the same moment in the main study, or how many devices were used for it.

A match was recorded only when a query returned the identical weekly score in all three countries. Any difference, even of one point, counted as a mismatch. Values that Google Trends exports as "<1" were replaced with 0.5, and blank cells were set to 0.

The headline numbers

Over the one-year window, 102 queries across 52 weeks produced 5,304 weekly comparisons for terms and the same number for topics. Terms matched across all three countries in 2,493 of them, or 47.0%. Topics matched in 1,738, or 32.8%. Combined, the match rate was 39.9%, which means six in ten weekly comparisons returned at least one differing score.

Stretching the window to five years improved agreement. Terms matched in 17,731 of 26,520 weekly comparisons (66.9%), and topics in 14,698 (55.4%). ExpressVPN characterizes this as an increase of around 21 percentage points; the exact gains in its tables are 19.9 points for terms and 22.6 points for topics.

"Projecting trends over longer periods seems to provide less variable worldwide trend scores, no matter where the search is coming from," the report states. "That said, mismatches still appeared in a notable share of results in the 5-year study." A third of term comparisons and 44.6% of topic comparisons still disagreed over five years.

That topics performed worse than terms is one of the less intuitive findings. A topic pools related searches across languages, and a wider pool might be expected to produce steadier numbers. In this dataset it did not. ExpressVPN concludes that "the broader subject coverage used for topics didn't appear to produce more uniform scores across locations."

Most disagreements were small

ExpressVPN tested whether one-point differences were inflating the mismatch count. It sorted every comparison into three groups: exact match, small variation (scores within one point of each other), and large variation (more than one point apart).

Period Query type Exact match Within 1 point More than 1 point apart
1 year Terms 47.0% 31.7% 21.3%
5 years Terms 66.9% 23.7% 9.5%
1 year Topics 32.8% 32.0% 35.2%
5 years Topics 55.4% 25.6% 19.0%

Counting anything within a point as a match lifts consistency to 78.7% for terms and 64.8% for topics over one year, and to 90.6% and 81.0% over five years. Even with that tolerance, at least one in five comparisons still differed in at least one of the samples, according to the report. For topics over one year, large variations (35.2%) were slightly more common than exact matches (32.8%).

The "<1" question

Google Trends exports very low interest as the string "<1" while its on-screen chart displays the same points as 0. ExpressVPN re-ran the exact-match comparison three ways: treating "<1" as 0.5, leaving it as text, and setting it to 0. The first two produced identical counts. Treating it as 0 added a handful of matches where one country returned "<1" and another a true zero. The change amounted to 0.06% of one-year comparisons and 0.16% to 0.18% of five-year comparisons, which the report calls "near-negligible."

The imputation table does contain a small arithmetic inconsistency with the main results table. It lists 2,812 one-year term mismatches against 2,811 in the main table, and 8,792 five-year term mismatches against 8,789, with similar one-to-three-unit gaps for topics. The totals implied by the imputation table therefore exceed the 5,304 and 26,520 comparisons stated elsewhere by between one and three. The report does not explain the difference, and it does not change any percentage at the precision given.

Where the disagreement sits

The most striking pattern is geographic. ExpressVPN compared countries two at a time and averaged, for each week, how many of the 102 queries returned the same score.

Period Query type U.S. and Japan U.S. and India Japan and India
1 year Terms 48.67 48.44 98.38
5 years Terms 69.44 68.53 100.20
1 year Topics 34.40 34.21 97.58
5 years Topics 57.52 56.85 100.06

Japan and India were almost interchangeable. Requests from those two countries returned the same worldwide score on nearly every query in most weeks. Requests from the United States did not. Over one year, the U.S. matched either Asian country on fewer than half the terms and roughly a third of the topics each week, and according to ExpressVPN, "the U.S. pairs didn't reach the same level of alignment as Japan and India in any single week." Two of the five-year averages sit slightly above 100 on a 102-query scale, which is mathematically possible but leaves almost no room for weekly mismatches.

The report offers no explanation for why a U.S. IP address would be served a different worldwide sample from Japanese and Indian ones. It does not speculate about data centers, caching or sampling partitions, and it explicitly disclaims any attempt to identify causes.

Which queries held steady

Globally recognized people, events and places matched most often. The term "Catherine, Princess of Wales" matched in every week but one across both studies combined. Tournaments such as the FIFA Club World Cup, the UEFA European Championship and the ICC Champions Trophy also ranked near the top.

A stable term did not guarantee a stable topic. "Lovely - Billie Eilish" matched in 51 of 52 weeks as a term but in only 7 weeks as a topic. "Tokyo Station, Tokyo, Japan" matched in all 260 weeks of the five-year study as a term; its topic matched in 38.

The weakest performers fell into two groups: queries with very little search activity in at least one country, and topics with broad labels such as "Drink," "Food," "Dog breed" and "Video game." The most extreme example in the report is "MASHA ULTRAFUNK - HISTED," which on one day in the one-year window scored 74 when requested from India and 1 when requested from both Japan and the U.S. Other queries showed discrepancies of plus or minus 10 points on certain days, according to ExpressVPN.

Four follow-up tests

To check whether hardware or the "worldwide" setting itself was driving the results, ExpressVPN took 10 queries at random from the original 102 and added France and the U.K. as VPN locations, choosing one predominantly English-speaking and one non-English-speaking European country.

Test 1: devices and countries together

Three machines, labeled "PC1," "PC2" and "Laptop," each pulled daily scores through servers in all five countries for a 90-day window from January 12, 2026 to April 12, 2026. Averaged across 91 days, France, the U.K. and India all returned the same score on all three devices for more than 9 of 10 queries. Japan averaged 6.21 for terms and the U.S. 5.09.

The device split was uneven. In the U.S., PC1 and PC2 agreed on 9.93 queries a day, but each matched the Laptop on only about 5.1. In Japan, the same pattern appeared in milder form. ExpressVPN's reading is that "in some locations, one device appeared to receive a different Google Trends sample than the other two." The report's summary describes device influence as "minor to negligible," a characterization that sits uneasily with a U.S. laptop that disagreed with two desktops on half the queries each day. The 5.89 figure the summary gives for the U.S. is an average across the five columns of the device table, mixing term scores, topic scores and pair scores.

Across countries the gap was wider. On a typical day, only 2 to 3 of the 10 queries returned the same score from all five countries at once. Pairwise, France and the U.K. averaged 8.37 matching queries out of 10 and India and Japan 8.41, while pairs including the U.S. averaged 5.1 and pairs of one Asian and one European country averaged 4.4. The U.S. aligned least with the U.K. among its pairs (4.58), despite both being predominantly English-speaking.

Test 2: same server, three devices

With all three devices connected to the same server in the U.K., then Japan, then the U.S., and running the same query at the same time, the scores lined up closely. Across 1,620 weekly observations per score type, covering 54 weeks from April 2025 to April 2026, the three devices returned identical scores in 81.73% of cases for terms and 81.98% for topics. Within a three-point range, agreement rose above 99%. The summary cites a correlation of 0.999998 or higher.

Test 3: a fixed region, different origins

This test addressed the possibility that "worldwide" was simply too broad a setting. ExpressVPN fixed the Google Trends region to Japan and changed only the server location among Japan, the U.K. and the U.S., again over 12 months of weekly data. If the reported region alone determined the sample, all three origins would see the same chart.

They did not. Terms matched exactly in 43.15% of cases and topics in 17.78%. Within three points, terms reached 75.56% and topics 52.59%. "This was one of the clearer patterns observed in our dataset," the report states. In other words, a user in London and a user in Tokyo looking at the Japan chart for the same query and week could see different numbers.

Test 4: changing the time window

The last test compared views of the same queries across three granularities: hourly (7-day view), daily (90-day view) and weekly (12-month view). The 7-day chart generally aligned with the tail of the 90-day chart. The longer comparison did not hold. According to ExpressVPN, "weekly values in the 12-month chart didn't appear to be a direct roll-up of daily values from the 90-day chart." The report gives no match percentages or sample counts for this test, describing it in qualitative terms only, and concludes that Google Trends "may use different sampling, scaling, or grouping methods depending on the selected timeframe."

What the study does not establish

Several gaps limit how far the findings travel. The main study covers three countries, a finite set of 102 queries drawn from Google's most popular searches, and data that ExpressVPN itself describes as "a snapshot." Re-running the same queries later "might produce slightly different results, even under identical conditions," the report acknowledges. Low-volume niche terms, the kind many advertisers track, are underrepresented by design, although the study's own results suggest those are the least stable.

ExpressVPN also states that "We have no way of knowing the criteria Google uses to select the categories it features in its YiS," so the starting pool reflects editorial choices by Google that the researchers could not inspect. The two terms the company added, "VPN" and "Antivirus," sit in its own commercial category; the interactive tool shows that "VPN" matched across all three locations in 10 of 52 weeks.

Then there is the commercial framing. The conclusion recommends running searches from different locations and states that "A VPN gives you the option to submit queries through different IP addresses, which would generally give you a more varied dataset." That is a recommendation from a VPN vendor. Nothing in the data shows that pooling samples from several countries produces a more accurate worldwide figure, only a different one; the report does not test which of the three countries returns the "correct" score, and Google publishes no ground truth against which to check.

Google has not responded to the study in the material provided, and the report does not say whether ExpressVPN contacted the company.

Why this matters for marketers

Google Trends is a staple of seasonality planning, content calendars and pitch decks. Agencies use it to time budget increases ahead of holiday peaks, and PR teams cite its percentage swings in press releases. PPC Land has reported on that reliance repeatedly: on October 15, an agency analysis that put UK "Black Friday" search interest down 36% since 2019 circulated without absolute volumes or a full methodology. Studies of that kind rarely state the country from which the underlying data was pulled, or on which day.

Google itself has been nudging the tool toward heavier professional use. It opened an alpha test of a Google Trends API in July 2025, offering a rolling 1,800-day window and consistent scaling across requests so that separate pulls can be joined without rescaling, a feature that answers one of the problems ExpressVPN's granularity test describes. In January 2026, Gemini began suggesting search terms inside the Trends Explore page, and that interface reached mobile web in April. None of those changes addressed the sample a given user receives.

Google's guidance has consistently framed the data as approximate. When it published Search Central documentation on Trends for content creators in October 2024, it described the data as a random sample of search queries. In an August 2024 tutorial, a Google Trends engineer recommended comparing topics with topics and terms with terms, and warned that frequent drops to zero indicate a sample too small to rely on. ExpressVPN's finding that topics were less consistent than terms across countries adds another dimension to that advice, though it comes from a vendor rather than from Google.

Sampling noise in Trends is not a new complaint, and repeated pulls of the same query have long been known to wobble. What ExpressVPN adds is a structured comparison that holds everything constant except the requester's country. The result points at a variable most users never think about: Google Trends appears to treat IP geolocation as an input even when the requested region is fixed.

A pattern across Google's data products

The study lands during a difficult period for anyone who measures Google from the outside. In October, PPC Land reported that Google was returning wrong results on 11% of DemandSphere's daily rank checks, and that Nozzle had counted swapped results on 8% of the search pages it collected. Those cases involve deliberate countermeasures against automated scraping, which is a different mechanism from sampling noise, and nothing in ExpressVPN's report suggests Google treats Trends requests from VPN servers as hostile. The common thread is narrower: figures that look definitive on a screen depend on conditions the user does not see. Every rank tracker, every keyword research tool that leans on Trends curves, and every analyst who exports a CSV inherits that dependency.

First-party Google reporting has had its own reliability issues. Search Console spent about ten and a half months over-reporting impressions because of a logging error before Google acknowledged it in April 2026. AI visibility tracking faces a parallel challenge: practitioners have found that location and personalization change the answers AI assistants give, so a single tracked prompt rarely represents what users actually see.

What changes in practice

For practitioners, the most direct implications concern documentation and comparability rather than any single number. A Trends figure pulled from a U.S. office and one pulled by a colleague in Mumbai may not be the same figure, even for an identical "worldwide" query. Two exports covering different time ranges may not be stitchable. And topic-level data, which many teams prefer because it pools spelling variants and languages, was the least consistent across locations in this sample.

The five-year results offer some reassurance. Over longer windows the U.S. pairs climbed from about 48 to about 69 matching terms out of 102 per week, and most disagreements in every sample were within a single point. ExpressVPN's own summary is measured: "The fluctuations we observed in our dataset were relatively small for most queries." For broad directional questions, whether interest in a product is rising or falling, the tool appears to hold up. For precise percentage claims about short windows, niche terms or broad topics, this study suggests the number depends in part on where the request originated.

Timeline

Summary

Who: ExpressVPN's research team, which is not individually named, studied Google Trends using the company's own VPN servers in the United States, Japan and India, with France and the U.K. added for follow-up tests.

What: Across 102 queries drawn from Google's Year in Search lists, the same "worldwide" Google Trends query returned the exact same weekly score from all three countries only 39.9% of the time over one year (47.0% for terms, 32.8% for topics), rising to 66.9% and 55.4% over five years. Japan and India almost always agreed; the U.S. frequently did not. Fixing the region to Japan still produced different scores depending on the requesting country, while devices on the same server mostly agreed.

When: The study was published today, October 21, 2026. The one-year dataset covers April 13, 2025 to April 5, 2026, and the follow-up device test ran from January 12 to April 12, 2026.

Where: The research was published on ExpressVPN's website, using data from Google Trends requested through VPN servers in the U.S., Japan, India, France and the U.K.

Why: Google Trends reports a sampled, relative index, and the study tests how stable that sample is when only the requester's location changes. The findings matter to marketers, journalists and researchers who rely on Trends for seasonality, content planning and published statistics, though ExpressVPN states its sample and method are insufficient to draw general conclusions, and the company sells the VPN service it recommends for pulling data from several locations.