How Exa actually works

I went looking for how Exa actually works because something about it kept nagging me. The web has been indexed the same way for twenty-five years, and suddenly a small company is re-indexing it, at enormous cost, for a reader that is not human. That reader is the AI agent. And the way you build a search engine for a machine turns out to be genuinely different from the way you build one for a person.

Here is what I found, and where the story is real versus where it is a landing page.

The core idea: predict the link, not the word

Every large language model is trained on one deceptively simple task: predict the next word. Exa's insight was to change the target. Instead of predicting the next word, they trained a neural network to predict the next link.

Picture a sentence like "Found an amazing article I read about the history of Rome's architecture:" followed by a hyperlink. Exa took billions of examples like that, hid the link, and trained a model to guess which page belonged there. Their premise, in their own words, is that "the way people talk about a link is a great indicator of both the link's content and its quality." Do that a few billion times and you get a model that, at query time, lets you search "as if you're about to share the link you want."

That is a real conceptual shift. Keyword search matches the words in your query against the words on the page. Exa does not care about the words. It turns every document into an embedding, a long list of numbers that captures its meaning, and turns your query into an embedding too, then finds the documents whose meaning sits closest. They call the approach Neural PageRank. It is closer to describing what you want than typing what you hope is on the page.

One honest caveat, because it matters: Exa likes to frame this against Google as "Google is just keywords." That is a convenient oversimplification. Google has used neural matching (RankBrain, BERT) for years. The real contrast is not neural versus keyword. It is that Exa built its whole stack, from the training objective to the output format, around a machine reader instead of a human one.

What it takes to build this

This is the part that surprised me most. Understanding meaning at web scale is not a clever prompt. It is a hardware problem.

Exa runs its own cluster, which it calls, with a pun I enjoy, the Exacluster. Eighteen nodes of eight GPUs each is 144, and "Exa" means ten to the eighteenth. In total it is 224 GPUs: 144 NVIDIA H200s and 80 older A100s, roughly 26 terabytes of GPU memory. They say they spent about five million dollars on it, and they use it in an almost violent way: embedding billions of pages while simultaneously training a reranker, checkpointing to storage as they go.

The Exa 2.0 embedding model, they say, was trained for over a month straight on the H200 half of that cluster. The embeddings are then served from a vector database they wrote themselves, in Rust, with custom clustering and compression. This is not a company gluing together off-the-shelf parts. They are paying, in electricity and months, for a model of the web's meaning.

Now the numbers you should hold at arm's length. Exa says its index holds eighty billion documents, tracks 1.4 trillion URLs, and calls itself "the world's largest independent index." None of that is independently audited. It is a landing page. Tellingly, the founder himself, when he is not writing marketing, softens it to "one of the largest." And the claim I saw repeated most, that the index "refreshes every page every minute," did not survive checking. What actually exists is on-demand live crawling when you ask for fresh content, which is a different and more modest thing. When you read numbers about search indexes, assume they are the best possible framing until a third party measures them.

What happens when you search

The flow itself is clean. Your query gets embedded by the same model that embedded the documents. The system does an approximate nearest-neighbor lookup, finding the closest meanings without comparing against all eighty billion vectors one by one, because that would be impossibly slow. The default mode now blends neural and keyword retrieval automatically and picks per query. If you read older Exa material telling you to set the mode to "neural," that is now legacy advice.

The part that is actually about agents

Here is where the machine-reader thing becomes concrete, and where I think the real product is.

A normal search engine returns ten blue links and a two-line snippet, because a human is going to click and read. That is useless to an agent. An agent cannot click. It needs the content, cleanly, in its context window, without a wrapper of HTML and cookie banners and navigation.

So Exa's most interesting endpoint is not search at all. It is /contents, which takes a URL and returns the page as clean markdown, stripped and readable, handling JavaScript-rendered pages and PDFs. It can return the full text, or extractive "highlights," the few passages most relevant to your query, which they claim are ten times more token-efficient than the whole page. It can even return structured JSON matching a schema you define. Their own framing is the giveaway: "LLMs just want dense information. We train models that take full webpages and condense them into just the tokens an LLM needs."

That sentence is the whole thesis of the company. Search for humans optimises for the click. Search for agents optimises for the token.

So how good is it, really

I could only find one genuinely independent benchmark, from AIMultiple, across eight agent-search APIs. Exa came third, in a statistical dead heat with Brave, Firecrawl, and Parallel. Competitive, clearly good, not the runaway leader the marketing implies. A separate open benchmark from a competitor put others ahead on raw question-answering accuracy, which is exactly the kind of source you read with one eyebrow raised.

The honest summary is this. The mechanism is real and elegant. The engineering is serious and expensive. The scale and freshness numbers are unverified marketing. And on the one neutral scoreboard that exists, it is one of several strong options, not a different species.

What this means when you are buying search for your agents

If you are giving an agent access to the web, the lesson underneath Exa is the thing to internalise: you are not buying a search box, you are buying a reader. Judge it on that. A few questions worth more than the marketing:

• Does it return content or links? An agent needs the text, clean, not a snippet it has to go fetch and scrape itself.
• How token-efficient is the output? Every result you pour into a context window has a cost and a distraction tax. Extractive highlights beat full dumps.
• Can you get structured output? If you have to parse prose back into fields, the tool is still built for a human.
• How fresh, honestly? Not the headline number. Ask whether it fetches live when you need it, and what that costs in latency.
• What does it cost per call, at your real volume? Agents make many calls, not one. The pricing that looks trivial at ten queries is your whole margin at ten thousand.

The specific winner will have changed by the time you read this. The shift will not. The web spent twenty-five years being organised for people who click. It is now being quietly rebuilt for machines that read. Exa is the clearest look I have found at what that actually takes.