Perplexity supercharges Pro Search with multi-step reasoning

Standard AI search is great for simple facts, but it usually falls apart the moment you need it to solve a complex problem or do actual research. It runs a single search, skims the top links, and gives you a surface-level summary.

Perplexity is rolling out a major overhaul to its flagship Pro Search feature, turning its engine from a basic answering tool into a multi-step research assistant.

Multi-step reasoning and dynamic follow-ups

The upgraded Pro Search can now tackle complex queries by breaking them down into step-by-step goals. Instead of just firing off a single query, it analyzes initial results and dynamically triggers follow-up searches to dig deeper into the topic.

Users can watch the entire thought process unfold in real time, showing the exact steps the system took to reach its conclusion alongside a broader set of sources.

Alongside the logic revamp, Perplexity also gave its computational chops a serious boost. Code execution has been accelerated for data analysis and debugging tasks.

To fix the classic hallucination problem with math, Perplexity integrated Wolfram|Alpha directly into Pro Search, offloading advanced mathematical questions to a dedicated computational engine.

Free users still get limited daily access to test out the upgraded feature, while Perplexity Pro subscribers get significantly higher limits and expanded model options.

Why it matters

We think this is the direction AI search has to go to stay useful. The novelty of single-shot AI answers has worn off, and professionals need depth and factual rigor.

For lawyers reviewing case law, developers debugging messy scripts, or analysts building market overviews, this upgrade saves real time. Letting the AI decide its own follow-up queries eliminates the tedious loop of manually rephrasing searches.

Integrating Wolfram|Alpha is particularly smart. Large language models are notoriously bad at precise calculations because they predict tokens rather than compute math. Handing that work to a specialized engine solves the problem cleanly without waiting for frontier models to miraculously become calculator-accurate.

The main risk here is error compounding. When an agent chains multiple searches together, getting misled by bad source material early on can send the subsequent steps down a rabbit hole. Perplexity's challenge won't just be search depth, but how well it filters noise when conducting multi-turn autonomous research.