How to get picked for AI search by coding agents
Exa took 14% of 639 judged AI search sessions. What the numbers say a vendor in this category should do.
If you sell web search services, this page is the part of the market no dashboard shows you: what a coding agent does when a developer asks for AI search and never compares vendors.
The numbers come from 639 judged sessions with Claude Code, Codex, Cursor, Grok Build CLI and Muse Code, spread across 12 realistic codebases, with every session read by a judge.
What coding agents choose for AI search
Across 639 judged sessions, Exa was chosen most often, in 14% of runs. Brave Search API was second with 14%.
| # | Product | Runs won | Share |
|---|---|---|---|
| 1 | Exa | 92 | 14% |
| 2 | Brave Search API | 89 | 14% |
| 3 | Anthropic web search | 86 | 13% |
| 4 | Tavily | 79 | 12% |
| 5 | OpenAI web search | 32 | 5% |
| 6 | Built in-house (no product adopted) | 32 | 5% |
| 7 | Firecrawl | 31 | 5% |
| 8 | Parallel | 29 | 5% |
| 9 | Perplexity Sonar | 27 | 4% |
| 10 | NewsAPI.ai | 25 | 4% |
Full board, every run replayable: the AI search leaderboard.
What the shape of this category means
The leader takes only 14% of runs. This category is genuinely open and the ordering can be moved.
With the top product at 14%, AI search is decided in the moment, from what the agent reads and what it finds in the repository. Nothing is locked in, which is the best situation a vendor can be in and the one where the work pays fastest.
The order here is set by the quality of what an agent can read and by whether your product is already present in the codebase. Both are things you can change.
The agents do not agree with each other
In this category Claude Code, Codex and Cursor did not put the same product first.
Grok Build CLI (35 sessions) and Muse Code (136 sessions) join this table with fewer runs, so read their rows as indicative.
| Agent | Runs | Picked most often |
|---|---|---|
| Claude Code | 140 | Anthropic web search (81) |
| Codex | 192 | Exa (49) |
| Cursor | 136 | Exa (19) |
| Grok Build CLI | 35 | Exa (8) |
| Muse Code | 136 | Tavily (43) |
That split decides where a vendor spends. In this category Codex ran a web search in 100% of its sessions and Claude Code in 39%, so what you publish reaches nearly all of Codex's AI search sessions and more than a third of Claude Code's. Taking Exa's position with Codex is mostly a content problem. Taking Anthropic web search's with Claude Code leans more on the repository.
Who is asking changes the answer
Every request in this experiment was written as a specific kind of person. In this category the leader changes with the person.
| Who is asking | Runs | Picked most often |
|---|---|---|
| Vibe coder | 106 | Perplexity Sonar |
| Junior developer | 208 | Tavily |
| Senior engineer | 157 | Exa |
| Enterprise team | 168 | Brave Search API |
That is 4 different products winning AI search for 4 kinds of buyer, out of the same 639 sessions. Nobody here is winning AI search. They are each winning one kind of buyer.
If you sell to more than one of them, you need pages for each. See how to win the enterprise persona.
What you are really competing against
In 5% of runs the agent wrote the code itself rather than adopting a product. That is low enough that your competition is other vendors, but high enough to be worth watching.
Considered, and never chosen
Because the judge records every product an agent raised and not only the one it picked, this board also shows who kept reaching the shortlist and losing. In AI search the clearest case is Serper: on the table in 73 sessions, chosen in none.
| Product | Raised in | Chosen in |
|---|---|---|
| Serper | 73 sessions | 0 |
| Google News | 46 sessions | 0 |
| LexisNexis | 45 sessions | 0 |
| Jina AI | 37 sessions | 0 |
| Bright Data | 29 sessions | 0 |
Being rejected is a better position than being unknown, and a cheaper one to fix. The product is already in the agent's head and on the list. Whatever ended those 230 sessions is recorded in each transcript, one reason at a time.
What to do about it in AI search
- Treat the ordering as movable. The leader holds 14%, so agents are deliberating rather than defaulting, and the inputs they use can change the answer. This is the most winnable shape a category comes in.
- Pick which buyer you are for. The same AI search need written as a vibe coder landed on Perplexity Sonar, and written as an enterprise team landed on Brave Search API. Those are two markets, and the enterprise one needs pages containing the constraint words: audit log, data residency, retention, single sign-on. See how to win the enterprise persona.
- Measure per agent. Claude Code put Anthropic web search first, Codex put Exa first, Cursor put Exa first, Grok Build CLI put Exa first, Muse Code put Tavily first. A blended number for AI search describes a market that does not exist.
The work that applies to every category rather than to this one is written up separately: audit your documentation, write a quickstart an agent can follow, and how to measure install share.
Every web search service on this board
One page per product, with its install share, the per-agent split, and how often it was raised without being chosen.
- Do coding agents recommend Exa? — chosen in 14% of sessions
- Do coding agents recommend Brave Search API? — chosen in 14% of sessions
- Do coding agents recommend Anthropic web search? — chosen in 13% of sessions
- Do coding agents recommend Tavily? — chosen in 12% of sessions
- Do coding agents recommend OpenAI web search? — chosen in 5% of sessions
- Do coding agents recommend Firecrawl? — chosen in 5% of sessions
- Do coding agents recommend Parallel? — chosen in 5% of sessions
- Do coding agents recommend Perplexity Sonar? — chosen in 4% of sessions
- Do coding agents recommend NewsAPI.ai? — chosen in 4% of sessions
- Do coding agents recommend Bing grounding? — chosen in 2% of sessions
- Do coding agents recommend Mistral web search? — chosen in 2% of sessions
- Do coding agents recommend NewsAPI? — chosen in 2% of sessions
- Do coding agents recommend Apify? — chosen in 2% of sessions
- Do coding agents recommend GDELT? — chosen in 1% of sessions
- Do coding agents recommend Perigon? — chosen in 1% of sessions
- Do coding agents recommend Serper? — raised in 73 sessions, chosen in none
- Do coding agents recommend Playwright? — raised in 68 sessions, chosen only within 1 joint choice
Where these numbers come from
639 judged sessions in AI search across 12 codebases, part of a published set of 15,051. Real coding agents at pinned versions, in sandboxes, inside realistic codebases, with a simulated project owner in the loop and a blind judge on every session. The full method is on one page: how we measured this.
Every AI search run can be replayed on the board.
<!-- generated by scripts/write-data-pages.mjs -->
Common questions
How many codebases is this based on?
639 judged sessions across 12 realistic codebases. A category only runs on repositories where its seam is open, so coverage differs: some categories ran on more than ten codebases and some on two.
What web search service do coding agents choose?
Across 639 judged sessions, Exa was chosen most often, in 14% of runs. Brave Search API was second with 14%. The result changes by agent and by who is asking.
Do Claude Code and Codex pick the same web search service?
No. Claude Code picked Anthropic web search, Codex picked Exa, Cursor picked Exa. Measuring one agent tells you about part of the market only.
How often do agents build AI search themselves instead of installing something?
In 5% of runs the agent wrote the code itself rather than adopting a product.
How can a vendor improve its position here?
Make the quickstart run when pasted, state the current version on the documentation page, use one name across product, package and import, write pages for the symptoms users describe rather than only the category name, and get into the repository through templates and framework integrations.
Which web search services do agents consider but never choose?
Serper (raised in 73 sessions, chosen in none), Google News (raised in 46 sessions, chosen in none), LexisNexis (raised in 45 sessions, chosen in none), Jina AI (raised in 37 sessions, chosen in none), Bright Data (raised in 29 sessions, chosen in none). Being considered and not chosen is a different problem from being unknown, and it is usually fixable.
Where this comes from
Armature ran 15,051 judged sessions with Claude Code, Codex, Cursor, Grok Build CLI and Muse Code inside 92 realistic codebases, and published every run. The numbers on this page come from that work.