A three-week review compared paid advertising activity with other campaigns and examined server infrastructure, device configurations and repeated activity
Based on a three-week analysis of paid clicks for 47 advertisers running ads in ChatGPT in September 2026, a firm that sells software to advertisers for identifying and filtering suspected invalid ad traffic has shared some findings with the media.
First, the analysis classified 0.1%–34% of measured clicks as invalid, depending on the advertiser. The 34% figure was the highest advertiser-level rate reported, not a rate for all 47 advertisers or all ChatGPT ads. The release did not identify how many advertisers had rates near that figure or provide click counts for individual campaigns.
Second, 15 servers associated with one hosting provider generated 318 clicks on ads from five advertisers during the observation period, according to the analysis. The release reported activity on each day of the three weeks. It classified the clicks as automated but did not publish the evidence behind its assertion that none came from a person.
Other findings
Third, four hosting and proxy networks in the analysis had reached 25 of the 47 advertisers and accounted for half of the clicks classified as invalid. The analysis did not disclose the total number of invalid clicks or the count attributed to those networks. Also:
- Clicks associated with data center and proxy infrastructure occurred at 3.6 times the rate recorded in Google Ads campaigns run by the same advertisers. Neither underlying rate nor the click totals were provided
- Clicks associated with device configurations classified as impossible occurred at 10 times the rate recorded in those Google Ads campaigns. The release did not specify the configurations, underlying rates or false-positive rate
- On the account described as worst hit, nearly 90% of flagged clicks were attributed to server infrastructure. The number of flagged clicks on that account was not disclosed.
- The analysis methodology involved using more than 200 per-click signals, including network infrastructure, device configurations, automation signatures and repeated activity. It did not publish its classification rules, independent validation results or billing records showing whether flagged clicks were charged and remained unadjusted. The sample was described as 47 enterprise advertisers, but selection criteria, markets, sectors, campaign sizes and exact observation dates were not supplied.
According to Miguel Lopes, Chief Product Officer, TrafficGuard, the firm that shared the findings, “Some campaigns we analysed were largely clean, but others were paying for clicks that had no chance of converting.” The press release did not include billing or conversion data that would allow readers to assess that characterization.
