Comparing the AI-native search APIs -
Data enrichment has become an essential part of our internal stack -
Every person added to our databases is automatically enriched with profile information: LinkedIn, company, job title, work history, location, etc.
That data then gets fed into scoring algorithms, notifications, data visualizations, and other internal products.
I’ve been using Exa blindly for the lookups -
Which - when reflecting - is pretty interesting because when I first tried the product it wasn’t that impressive. I felt like my queries kept coming up with 0 or weak results.
But I gave it another try after their latest round announcement. And I became surprisingly loyal to it; the product obviously has hugely improved.
However, I realized last week that I hadn’t really done the analysis on the best option.
So I ran two tests:
Profile data enrichment
General research
Profile data enrichment
For the first test, I gave each provider the names and companies of 135 people and asked it to find the correct LinkedIn profile.
Surprisingly, everyone did a good job with this.
There was no meaningful difference in quality between Exa, Firecrawl, Brave, and Parallel. At least for straightforward people search, no one had that big of an edge.
Firecrawl was the cheapest provider, followed by Brave. Parallel was the most expensive.
Which was slightly annoying, because I was rooting for Parallel in every test. I’m a huge fan of Parag (the company’s CEO) - we overlapped at Twitter, and he’s a super genius. But at the same time, they just announced a massive Google integration & partnership, so maybe cheap LinkedIn lookups aren’t exactly their highest priority haha.
General Research
For the second test, I wanted to test on something broader.
I created 25 realistic queries across five categories, and every provider received the same queries:
News (e.g. recent US export controls on advanced AI chips to China)
Company (e.g. which venture firms led Anthropic’s most recent funding round)
Technical (e.g. Postgres pgvector HNSW index tuning for approximate nearest neighbor search)
Discovery (e.g. startups building developer tools for evaluating LLM output quality)
Factual lookups (e.g. who is the current CEO of Twitter/X and when did they start)
Claude then graded the returned links from 0 - 10 based on relevance. The results were anonymized, and the labels were rotated so neither the brand name nor its position could influence the score.
Exa and Perplexity were the winners.
They had the strongest results overall and (according to Claude) had an edge in open-ended research questions for finding the right source. It’s worth noting Perplexity was a bit slower.
But again, strong results across the board.
Conclusion
All the providers had strong results for both tasks. Which - to me - was actually just a sign of how powerful all these tools are.
For straightforward profile enrichment, the results were similar enough that price is the only differentiator. For broader research, Exa and Perplexity ranked the highest.
The main caveat to all of this is that I’m not trying to publish the definitive search benchmark. I was interested in testing the specific things we use these tools for - specifically extracting profile information - and making sure I wasn’t using Exa blindly.
There are broader and more rigorous benchmarks. Search Arena compares complete AI search experiences, and AIMultiple ran a larger benchmark directly across eight search APIs. AIMultiple’s results were slightly different. Brave had the highest raw score; Firecrawl, Exa, and Parallel were all close to one another. Parallel performed particularly well on real-time and comparative queries.
But anyway, TL;DR: they all work and are incredibly powerful & useful. I’ll probably keep Exa as our default.
I’m a General Partner at Chapter One, an early-stage venture fund that invests $500K - $2M checks into pre-seed and seed-stage startups.
If you’re a founder building a company, please feel free to reach out on Twitter (@seidtweets) or Linkedin (https://www.linkedin.com/in/jamesin-seidel-5325b147/).