Perplexity bought attention at Indian scale. In July 2025, the AI company partnered with Airtel to give the telecom operator’s 360 million customers a free 12-month Perplexity Pro subscription, creating one of the largest AI growth experiments of the past year.
New redemptions ended on January 16, but the promotion had already transformed Perplexity’s reach. The company recorded 56 million downloads in India during the seven months the offer remained available to new users — more than nine times the total from the previous seven-month period.
Perplexity’s app downloads reached 5.9 million in India in July 2025, a 625% increase from the previous month. Monthly active users more than doubled to 8.9 million in July before reaching a peak of 22 million in October, giving the company a striking burst of visibility in a country with more than a billion internet subscribers and over 700 million smartphone users.
The giveaway created a spike, not permanent momentum
The promotion’s afterlife is less spectacular. Perplexity’s downloads in India between February and July 2026 were estimated at 3.3 million, down more than 90% from the preceding six months. Monthly active users stood at nearly 14 million in July 2026, down 37% from the October peak.
That decline does not mean the offer failed. Free access brought millions of people into the product, and a portion continued using it after the promotion ended. Abe Yousef described the result plainly: “While the time-sensitive nature of this promotion would naturally lead to a decline in adoption after the offer period, ongoing usage has remained resilient.”
Revenue data gives that resilience a more concrete shape. Perplexity’s in-app purchase and subscription revenue in India between February and mid-August 2026 rose about 60% from the period when the Airtel offer was available to new users.
From July 18 through August 12, average daily in-app purchase revenue in India ran 9% above the previous 30 days and 27% above the average for the first half of 2026. The free subscription ended as a customer-acquisition stunt, but it also left behind a larger pool of people willing to pay. Even promotional users occasionally become customers. Annoying for anyone hoping the funnel would remain theoretical.
The numbers still reveal only part of AI use
Usage figures show who downloaded an app, opened it, or paid for access. They do not explain what people ask AI systems to do, and the available research offers only a narrow view of that missing picture.
The AI Observatory, a project involving MIT Trustworthy AI Research, analyzed 85,633 conversational turns across 24,521 conversations from seven datasets collected between 2023 and 2025. The datasets covered 5,000 users interacting with 52 different models.
That is a useful sample, but it remains small beside the data analyzed by Anthropic and OpenAI: 1 million Claude conversations and 1.5 million ChatGPT conversations, respectively. The AI Observatory’s data also likely underrepresents sensitive uses because it comes from voluntarily provided sources, while AI companies typically do not share their chat data for analysis.
Still, the patterns are clear enough to challenge the idea that people use every chatbot in the same way. People turned to Grok and Gemini more often for information retrieval, with Grok proving especially popular for news and politics; Anthropic drew more coding work, Gemini saw social and roleplay uses, and ChatGPT handled homework assistance.
Conversations within WildChat grew longer and more elaborate over time, with prompt tokens, response tokens, and conversation turns all rising. Exchanges labeled as sensitive became less frequent over time, though the dataset’s voluntary nature limits what that trend can establish.
The model behind a conversation also mattered. ChatGPT exchanges powered by GPT-3.5 were shorter, while GPT-4 conversations were longer and more iterative.
Those findings place Perplexity’s India numbers in the right frame. A download surge can measure distribution, a user peak can measure curiosity, and revenue can measure some lasting value — but none explains the work people hand to an AI system.
Shayne Longpre put the broader problem in one sentence: “No single company report tells the whole story.” Reuel made the stakes harder to ignore, warning against “completely operating in the wild and making these really consequential decisions without knowing what’s actually happening beyond those company narratives.”
Based on
