What is
SuperSearch?

Ten pillars of superintelligence

There are ten spaces in which a system can be intelligent, and in every one of them we are far from the upper bound. The list below is the clearest map of that territory anyone building these systems has shown in public, and it is worth walking through, because one of the ten is not like the others.

Ten spaces of intelligence

We are far from the upper bounds of intelligence.

1. 01Visual intelligence
2. 02Language intelligence
3. 03Knowledge
4. 04Reasoning
5. 05Computational speed
6. 06Physical intelligence
7. 07Social intelligence and morality
8. 08Creative intelligence
9. 09Meta-cognition
10. 10Survival and replication

Under pillars 2, 3 and 4 sits one requirement, the last mile of knowledge. Whatever has to be known must be reachable, wherever it sits.

1. Even the most powerful models cannot know everything. They need search for current, accurate knowledge of the real world.
2. Every AI system needs trusted knowledge from the web before it can act.
3. The knowledge layer is the prerequisite for every pillar above it. You cannot simulate what you do not know.

Introducing SuperSearch

That is the idea. It is the first pillar stated as a capability rather than a component. Two things are about to be true at once. Agents will search the web a thousand times more than people do. That is Exa's estimate.¹ The web they search will also change faster than it ever has. Agents will be writing to it as well as reading from it. No model can hold that in its weights. The world moves faster than a training run. No single index can hold it either. No index has ever held the web. What is needed is a machine that keeps track of all of it. The first thing that machine has to do is ask everything that knows.

Any one index is half the web

This would matter less if one engine held the web. None ever has. In 1995 Erik Selberg and Oren Etzioni built MetaCrawler at the University of Washington because "each service provides an incomplete snapshot of the Web," and measured that no single service returned more than 45 percent of the pages people went on to open.² In 2005 Dogpile commissioned researchers at Pittsburgh and Penn State to compare the first pages of Google, Yahoo, MSN and Ask across 12,570 queries: 84.9 percent of results appeared on only one engine, and 1.1 percent on all four.³ The 2007 replication found 88.3 percent unique to one engine.⁴ A 2022 study put the top-ten overlap between Google and Bing "always under 32 percent."⁵

The indexes built for agents are no different, and now there is a measurement. In July 2026 Dogpile ran one hundred real agent queries through seven search APIs: Exa, Tavily, Firecrawl, Brave, Bing, Serper and Claude's built-in search. Any single one "missed roughly half of what the others could find."⁶ The same study found that three of the seven contained each other's results at 83 to 86 percent in every direction, "Google in three trenchcoats," and that Claude's built-in search rides on Brave. Seven APIs were four independent indexes.⁶ Nobody inside any one of those APIs could have seen that. It is visible only by asking several and comparing.

So the retrieval step of every automated research loop, as built today, reads about half of what is known, cannot tell which half, and cannot tell when two of its sources are one source.

The field

SuperSearch is possible now because, for the first time since the 1990s, there are many engines to ask, and they want to be asked by programs. Exa, Brave, Parallel, Perplexity, You.com, Linkup, Ceramic, Keenable, Seltz and TinyFish each describe an index of their own, several of them started in the last eighteen months by people who built Google's, Yandex's and Cuil's.⁷ Firecrawl's Alexandria, announced in September, gathers its own indexes with licensed sources and the live web behind one call.⁸ Context.dev offers one API over several kinds of web data so that teams do not have to "combine multiple data vendors" themselves.⁹ SerpApi and Serper sell Google's results page as JSON. Cloudflare's Web Search API, launched in October, routes to Ceramic, Exa and Linkup.¹⁰ Dogpile itself is back, as an MCP server.

Every one of these is a view of the web, and the views disagree. Some disagree because they are independent, which is the signal. Some agree because they are the same index resold, which is noise dressed as corroboration. Each of them, on its own, is one pillar's worth of a single engine. SuperSearch is the layer above them that turns many views into coverage, their disagreement into evidence, and their behaviour into a record. The more independent indexes exist, the more it is worth, which is the opposite of how a single engine's value behaves.

Jevlin, the first product

The first product built toward SuperSearch is Jevlin. It is a fan-out search. One call asks every engine at once and returns one list. Every result says which engines found it. Every engine leaves a receipt. It finds what no single engine finds alone. Each of its three properties answers one failure of a single index, and each one is only possible because the reader is a program.

Coverage: one call, every engine. The half of the web a single index misses is in the other indexes. Asking them all in parallel replaces the retry loop, which costs latency and tokens on every turn, with one call that costs neither. This is the part metasearch had.

Evidence: attribution on every result. This is the part metasearch hid, because a person wants one ranking. A program wants the seams, because agreement between independent indexes is evidence. Reciprocal rank fusion "consistently yields better results than any individual system"; ensembles of retrievers "substantially improve" over any single one; on a 2025 hallucination benchmark the fused retriever gave "the lowest hallucination rate."¹¹ Attribution is also how a program tells real agreement from one index in three coats, which the Dogpile study could only discover by hand.

Audit: a receipt per engine. Which engine answered, how fast, what it returned, what it cost. Nobody built this before, because nobody needed it: the person was the audit. A research loop has to check for reward hacks before it treats a result as progress, because the judge is an attack surface and a stronger optimizer finds its blind spots faster than a weaker one. A search layer the agent cannot audit is the same kind of blind spot, and the loop will find it. The receipt is the judge for retrieval: it says one engine timed out and two returned nothing before the agent writes its confident answer.