GPTZero barely grew this summer, but a smaller AI detector nearly tripled its traffic

Published
Watermarks are coming to AI text. Who still needs a detector? A figure holds a large magnifying glass over a sheet of paper stamped with a seal.

Pangram’s estimated visits rose 195% from May to September, GPTZero’s rose 4% and 19 of 23 AI humanizer sites lost visits. What that leaves open.

AI text is starting to carry watermarks. In August, Anthropic said new Claude models would mark their output, TechCrunch reported. On October 5, OpenAI said ChatGPT and Codex text would get an invisible watermark in the European Union.

If you were thinking about building an AI detector, that sounds like bad news. We pulled the company and keyword histories stored in Night Watcher to see what the market looked like through September.

Pangram, one of five detector brands we compared, went from an estimated 1.08 million monthly visits in May to 3.19 million in September. The two largest, GPTZero and ZeroGPT, each ended 4% higher than they started. And of 23 AI “humanizer” sites, 19 lost visits.

Our histories end in September. That is after Anthropic’s announcement and before OpenAI’s. They cannot show what either one changed, and they cannot separate the watermark news from the start of the school year.

Pangram nearly tripled. The two largest ended 4% higher.

We compared five detector brands with at least 30,000 monthly brand searches in September. Broader writing products that include a detector, such as QuillBot, Grammarly and Turnitin, are not in the group.

Monthly visits indexed to May 2026 equals 100: pangram.com ends at 295, gptzero.me at 104, zerogpt.com at 104, originality.ai at 83 and copyleaks.com at 74.
Source: Night Watcher. Research retrieved November 16, 2026. Domain visits across traffic channels, not users or revenue. The index compares growth rates; the table below shows size.
Tracked domainMay 2026 visitsSeptember 2026 visitsChange
pangram.com1,079,7523,190,346+195.5%
gptzero.me16,839,82017,547,481+4.2%
zerogpt.com26,339,09827,453,503+4.2%
copyleaks.com3,790,9202,796,881−26.2%
originality.ai2,395,9571,993,918−16.8%

Pangram grew fastest from the smallest base. ZeroGPT still had more than eight times its estimated visits in September, and GPTZero more than five times.

The flat lines hide a deep summer dip. GPTZero’s estimated visits fell 40% between May and July and then recovered. Ending 4% up is a little better than the median for sites that size, which was −3%. But searches for “gptzero” were 49% higher in September than in May, so level traffic came against stronger demand.

Pangram was not the only detector site to grow. A much smaller one, getsolved.ai, went from about 190,000 visits to 636,000. Winston AI fell 25%.

The traffic history cannot tell us why any of them moved. We have not tested any detector’s accuracy, and visits are not a measure of it.

The category follows the school year. Pangram’s searches kept climbing.

Searches for “ai detector” were an estimated 5 million in September 2026. They were also 5 million in September 2025. “gptzero” was 550,000 in both months.

In between, the category rises and falls with the school calendar. “ai detector” dropped to 1.83 million in July before recovering.

Four keyword histories from May 2025 to September 2026. “ai detector” and “ai humanizer” fall each summer and return each September to the same level. “pangram ai detector” climbs from 390 to 33,100. “turnitin ai detector” reaches 135,000 in September 2026. Each panel uses a separate vertical scale.
Source: Night Watcher. Values are rounded monthly search estimates. Each panel has its own vertical scale, and related queries must not be added together.

“pangram ai detector” rose from 1,300 searches in September 2025 to 33,100 a year later, about 25 times as many. It rose in June and held in July while the category fell by more than half, then more than tripled by September.

That is still a small number next to the giants. But no other brand in the comparison came close to that growth: “copyleaks” was up 22%, “gptzero” and “zerogpt” were level, and “originality ai” was down 18%.

“turnitin ai detector” was up 23% year over year at 135,000. And “ai image detector” more than tripled between August 2025 and August 2026, from 33,100 to 110,000. Its history ends a month earlier than the others.

Nineteen of 23 humanizers lost visits

Humanizers are the tools that rewrite AI text to read as human, or to pass a detector. We found 23 of them in our data with at least 200,000 estimated visits in May or September and an estimate for both months.

Nineteen had fewer visits in September than in May. The median change was −35%. Undetectable.ai, the largest of the 23 in May, fell from 3.99 million to 2.65 million.

The school calendar does not explain that. May and September are both term months, and searches for “ai humanizer” were level between them. A larger phrase, “humanize ai”, was 18% lower than a year earlier.

Four grew. Two of them, gpthumanizer.ai and humanizeai.co, multiplied from bases under 100,000 visits. The biggest by far was cleverhumanizer.ai, which went from 1.03 million visits to 5.93 million and ended as the largest of the 23. Its own site says the tool is free with no monthly word cap.

Add it in and the group’s combined visits fell 7%. Leave it out and they fell 32%. Combined visits add separate domain estimates, so read those as a direction and not a head count.

Two gaps matter here. The largest humanizer in our data in May, humanizeai.pro, has no September estimate and is not in the 23. It had 9.9 million visits in May and 6.1 million in August. And verva.com shows no visits in May and about 260,000 in September; a change from zero cannot be calculated, so it is left out too.

What watermarks leave open

OpenAI’s own post sets out the limits. The watermark covers the EU only at launch. The detector is open to approved researchers and expert organizations, not the public. In OpenAI’s tests, replacing a quarter of the words in a passage cut detection from about 92% to 17%. And the post says the absence of a watermark “does not prove human authorship.”

A teacher, editor or hiring manager who needs an answer today still has to get it somewhere else. That is consistent with detector demand holding level. It does not prove watermarks will fail.

What we would look at if we were building here

A general text detector is a crowded bet. The category’s searches are level year over year, the two largest brands had 85% of the visits among the five we compared, and a challenger is growing quickly.

A humanizer looks worse. The median one lost a third of its visits between May and September, and the one that added the most visits gives its product away.

The narrower questions look more open. Who has to act on a detection result, and what does a wrong answer cost them? An editor paying freelancers and a school handling an appeal have different needs. The rise in image-detector searches is another lead.

That is a direction to investigate, not a validated opportunity. Before building, we would sit with five people who make that kind of decision and watch what they do when a result comes back uncertain. If they already have a process that works, move on. If the same gap keeps appearing, test whether closing it earns a paid pilot.

To keep watching this group, our ranking of AI detection and humanizer companies by monthly growth is public.

How we checked the numbers

We retrieved stored records from Night Watcher on November 16, 2026. The detector comparison uses estimated monthly visits for May through September 2026 for five detector brands with at least 30,000 monthly brand searches in September and an estimate for every month. It is a selected comparison, not every detector.

The humanizer comparison uses the tracked sites we classified as selling the rewriting of AI text to read as human or to pass detection, with at least 200,000 estimated visits in May or September and an estimate for both months. We classified them by hand. Twenty-three qualified. Two more large ones have no September estimate; both had fewer visits in August than in May. One site with no visits recorded in May is also left out.

All traffic figures are modeled domain visits across channels. They are not first-party analytics, users or revenue. The median comparisons use the 8,708 sites we track with at least 20,000 estimated visits in May and an estimate for both months.

For search demand, we inspected the 260 highest-volume keyword histories from a targeted search and compared selected phrases from May 2025 to September 2026. We excluded duplicate spellings and did not sum overlapping queries.

We did not test any detector or humanizer. Nothing here measures accuracy, false accusations or whether a tool works.

Jordan Cole

Creator of NightWatcher AI. Specializes in data-driven insights for AI product development, market validation, and competitive analysis.