ChatGPT's traffic slump might just be summer break

ChatGPT's historic honeymoon phase has officially cooled off.

According to new data from analytics firm Similarweb, August marked the third consecutive month of worldwide traffic declines for OpenAI's viral chatbot.

Users aren't just visiting less, they're also hanging around for shorter stints. Data shared by Reuters shows average time spent on the site dropped from 8.7 minutes in March to just 7 minutes in August.

The summer slump and the back-to-school bounce

The bleeding, however, appears to be slowing down.

Global visits dropped by roughly 10% in both June and July. But in August, that decline leveled off to 3%.

In the U.S. alone, web visits actually ticked up by 0.4% in August. Worldwide unique visitors also saw a tiny bump of roughly 0.3%, inching up from 180 million to 180.5 million.

So why the summer drop? Similarweb senior insights manager David F. Carr pointed to a very relatable culprit: summer break.

A huge chunk of early traffic came from students writing essays, prompting professors to panic and a Princeton student to launch the AI detector GPTZero. As Carr noted, traffic stabilized late in August right as American students headed back to classrooms.

Hype exhaustion or a dumber bot?

Before Meta's Threads stole the crown in July, ChatGPT held the record as the fastest-growing consumer app ever, crossing 100 million users in just two months.

Beyond homework, it quickly found real footing in the workplace for coding, research, and scheduling.

Yet the traffic slump coincided with growing frustration over model quality. In July, power users complained that GPT-4 felt noticeably dumber.

A paper from researchers at Stanford and UC Berkeley backed those worries up, finding that GPT-4 suffered severe performance drops in specific areas, including prime number identification, where accuracy plummeted from 97.6% to just 2.4% over three months.

OpenAI did not immediately respond to requests for comment on the numbers.

Why it matters

We think this traffic dip marks the natural end of the tourist phase for generative AI.

For everyday users and knowledge workers, ChatGPT is no longer a party trick. It is a utility. If you rely on it for production code or analysis, wild swings in model accuracy are a massive operational risk. The Stanford and Berkeley findings highlight the biggest headache for enterprise adoption: unpredictability.

For the AI race at large, the data reveals a fragile foundation. When summer vacation alone can shave 10% off your global traffic, your product is still functioning more like an academic crutch than an indispensable workplace OS.

OpenAI proved it can build the fastest-growing toy on earth. Now it faces the much harder task of keeping users when novelty wears off, which starts with proving that its flagship models aren't quietly getting worse behind the curtain.