Both Vendors Won Their Own Benchmark
Exa published a comparison page. Exa won. Parallel published a comparison page. Parallel won. Same two products. Two different verdicts. That should tell you exactly what a vendor benchmark is worth when you're deciding what goes inside your business. Here's the part nobody sells you: the choice between them isn't a technical one. It's an operational one, and you can make it without knowing what a vector embedding is.
There are only two shapes
Every AI assistant that answers questions about the live web needs something to fetch the web for it. Those somethings come in two shapes.
A retriever. You ask, it hands back the most relevant pages, fast. Your AI reads them and answers. Picture a research assistant who puts the right five documents on your desk in half a second and says nothing else.
A researcher. You give it a brief, it goes away, visits a dozen sources, cross-checks them, and returns a structured answer with citations. Picture a junior analyst who takes the brief and comes back with a report.
Exa is the first shape. Parallel is the second.
Neither is better. They're different jobs.
Sort your own questions
Forget the products for a minute. Here are things a business actually asks:
"Find the five most current compliance resources on this question." "Here's my best customer's website — find twenty companies that look like them." "My customer-facing assistant needs to answer without an awkward pause."
Those are retriever jobs.
"Find every Australian accounting firm meeting these six conditions, with verified contacts and evidence." "Build a competitor landscape where every claim is backed by a source I can check." "Watch this regulator's site and tell me when the rules change."
Those are researcher jobs.
Read both lists again. The pattern isn't technical.
The retriever jobs are the ones where a human is sitting there waiting. The researcher jobs are the ones where a human would have burnt two days.
That's the decision line. Not speed versus quality. Who's waiting, and how long is acceptable.
The expensive mistake
It isn't picking the wrong vendor.
It's putting a two-minute research job in a path where a customer is watching a spinner. Or paying deep-research prices for a lookup a fast search would have nailed for a fraction of it.
Most businesses that get burnt here didn't buy the wrong tool. They pointed a good tool at the wrong job.
If search quality is genuinely business-critical, run both: the retriever on every interactive path, the researcher only on the high-value jobs where depth and citations earn their wait. That's not indecision. That's matching the tool to the job.
The test that beats every comparison page
Write down the 30 to 50 questions your business actually needs answered from the web. Real ones, pulled from real work. Not invented for the test.
Sort them into "someone is waiting" and "someone would have spent a day".
Then run both tools across the whole list and score four things:
Did it get the answer right? Was the source one you'd be comfortable citing to a client? How slow was it at the slow end — not the average, the worst tenth? What did it cost once you include page fetches and retries?
That takes about a week. It will beat every comparison page on the internet, because it's the only one run against your business.
The bit underneath
Notice what that week actually produces. Not a vendor decision. A written list of the questions your business needs answered, sorted by urgency and value.
Most businesses can't produce that list. That's the real finding.
The tooling question is downstream. The question upstream is which jobs in your business are repeatable enough, and documented well enough, to hand to a machine at all.
Smart AI in a broken process is still a broken process.
If you don't have that list of 30 questions yet, that's the work to do first. The AI Discovery Workshop is fixed price, fixed scope, and you leave with it written down.