How Perplexity Is Changing the Way AI Searches the Web

Perplexity, an AI search company, has announced a new way to organize how its search system works. Instead of using one big system that handles all searches the same way, the company is letting its AI software build search instructions on the fly, customized for whatever question you ask. The company announced this shift on June 1.

Think of the old way as a single fixed recipe for making soup—you follow the same steps for every pot. The new way lets the AI write its own recipe each time, choosing which ingredients and methods work best for that particular meal.

What Changed

Under this new approach, called Search as Code, the different parts of Perplexity's search system become like building blocks. Instead of sending every question through the same pipeline, the AI can pick and arrange these blocks in different ways depending on what it's trying to find.

The system needs to handle thousands of searches every second across Perplexity's different products. In some cases, a single task might run hundreds or thousands of search operations in a matter of minutes. The old single-recipe approach was starting to feel like a bottleneck.

Now the AI can write code that says: "For this question, I need to search the web like this, then check academic databases like that, then organize the results this other way." Each search gets its own tailored plan.

How It Works

The AI generates these search instructions—the code—on the spot, then runs them in what's called a sandbox. A sandbox is basically a locked-off digital space where the code can run safely without risking damage to the main system. Think of it like a controlled laboratory: the experiment can happen, but if something goes wrong, it only affects that one room.

This approach gives the system more flexibility—different questions get different treatment. It also adds safety, since the code runs in a protected environment.

Why This Matters

The shift from one-size-fits-all search to customized search follows a pattern we have seen before in software engineering. Decades ago, most of corporate computing ran on large mainframe machines that did everything the same way for everyone. As companies needed more flexibility, those mainframes gave way to smaller, specialized systems that could be mixed and matched. We are seeing the same evolution now with AI search.

Many AI systems today use something called retrieval-augmented generation, or RAG. The basic idea: when you ask a question, the AI finds relevant information and then generates an answer based on what it found. The problem is that these systems often treat search like a black box—they ask for information in one fixed way, regardless of the question. Perplexity's approach lets the AI adapt its search strategy based on what it is trying to learn.

This is particularly useful for AI agents—software systems that can work through complex tasks on their own, step by step. If an AI needs to gather information about a topic from many different angles, it can now adjust its search methods as it goes, rather than being locked into one predetermined path.

Trade-offs to Consider

Programmable search introduces some genuine complications. The system now has to generate search code before it can run it, which takes a bit of time. That extra time might be worth it if the custom search finds the right information faster, but it is a balancing act.

When the AI generates code, there are also security concerns. You want to make sure bad actors cannot trick the system into doing something harmful. The sandbox helps, but creating these protected spaces that are both flexible and secure is tricky.

A monolithic search system—the old way—can be fine-tuned to handle the most common searches very quickly. A programmable system might do specialized searches better but behave more unpredictably overall, depending on what the AI decides to do.

What Comes Next

Perplexity has also released Perplexity Labs, a platform that lets AI systems tackle longer, more complex research tasks. It includes tools for deep web searching, running code, and creating charts and images. When you combine programmable search with these extended research capabilities, you get AI that can really dig into a problem over time, adjusting its approach as new information arrives.

The broader technology industry may be watching this closely. If search becomes something the AI can program on demand, it raises a question: what other parts of AI systems could become programmable in similar ways. The same idea might eventually apply to how AI generates answers, reasons through problems, or manages information flow.

For companies building AI applications that need to find information in sophisticated ways, Perplexity's approach offers a different path than the traditional RAG systems most teams use today. It could be especially valuable for applications that have to adapt their search methods based on what they discover along the way.