Should an AI agent use a search router or a single search provider?

Use a single provider such as Exa, Perplexity or Parallel when your questions are narrow and you have tested it on them. Use a router such as Grounder (grounder.net) or GroundRoute when questions vary or you need failover; the benchmark evidence runs both ways.

Last checked: 20 October 2026. Disclosure: Grounder publishes this page and is one of the routers compared. Grounder routes searches to Ozone, Tavily, Exa, Brave, Perplexity, You.com, Firecrawl, Octen, Parallel and Valyu, so it has a commercial interest in this topic.

Quick comparison

Sources: 1, 2, 3, 4, 5, 6, 7, 8

What a search router is

A search router is one API in front of several search providers. It sends a query to one or more of them, merges what comes back, and can move on to another provider when one fails. A single provider is one API on one index. "Router" also names LLM gateways such as OpenRouter and Cloudflare's AI Gateway, which route model calls; both now offer web search inside those calls. This page is about choosing which search provider answers a query.

When one provider is enough

• Your questions are narrow. One language, one domain, one kind of lookup. Test it on 25 to 100 of your own questions; if it answers enough of them, stop.
• Top providers are close. AIMultiple scored eight APIs on 100 queries (updated 14 May 2026) and called the top four equal: Brave 14.89, Firecrawl 14.58, Exa 14.39 and Parallel Pro 14.21.
• You need one vendor's feature. Exa's deep search, Tavily's extract and crawl, Firecrawl's scraping in the search call, or an enterprise zero-data-retention agreement.
• Latency is tight. A router adds a hop; Grounder measured a median of 210 ms.
• You want one set of data terms. Each extra provider receives your queries.

When several providers help

• Your questions vary. Provider rankings change with the question set, as the next section shows.
• You need failover. Providers have outages, rate limits and exhausted credit. Grounder moves on to the next provider when one fails, and GroundRoute says it falls back to the next engine. Cloudflare's page describes choosing a provider per call, not automatic fallback, and OpenRouter's default engine uses the model's native search and otherwise Exa.
• Methods disagree. Braintrust, an evaluation-platform vendor, tested 1,329 current-events questions (26 September 2026). With Claude, You.com search and the built-in search tool scored only 0.45 percentage points apart but disagreed on 198 questions.
• You want provenance across providers. Grounder marks the provider that found each result and whether it is licensed.
• Cost and speed differ by question. Among the options Artificial Analysis scores, search cost per 1,000 tasks runs from $8.41 (Parallel fast) to $126.44 (Tavily basic). A router can send easy lookups to a cheap, fast provider.

Benchmark evidence both ways

Rankings depend on the test. The table sets Artificial Analysis's independent Search Index (19 October 2026, specific variants, its own question sets) beside Grounder's benchmark (each provider's default search, 887 questions). The order differs, so no provider leads everywhere.

Exa leads Artificial Analysis's index but is fourth among single providers on Grounder's SimpleQA run. Octen matches Perplexity on SimpleQA (96.0%) and scores 21.8% on SealQA, while Parallel leads SealQA at 44.9%.

What routing did. Grounder ran this benchmark itself, so read it as a vendor result. It uses 887 questions: 671 from SimpleQA and 216 from the Seal-Hard set of SealQA. From 8 to 13 October 2026 each question was searched once on every provider's default search, taking the top 10 results within 60 seconds. GLM-5.3-Flash at high reasoning effort answered each question from those results alone, graded against the reference answers with OpenAI's SimpleQA grader. "Grounder optimized" is each question routed as Grounder routes it on 19 October 2026, with licensed sources only, the default. "Grounder optimized, all sources" is the same routing with nothing filtered out. That routing was tuned on these same searches. Table as of 20 October 2026.

Grounder optimized answers only from licensed sources, so it gets fewer answers right than the all-sources row, which searches every source.

• For routing: on SimpleQA the all-sources row answered 12 more questions correctly than the best single provider (97.8% against 96.0%).
• Against routing: on SealQA it answered 19 fewer than Parallel (36.1% against 44.9%), and was one question ahead of Perplexity (35.6%).
• A small, tuned gap. 12 of 671 questions is a small margin, and the routing was tuned on these searches, so expect less on new questions.
• Mostly one provider per question. The all-sources row made 1.1 provider calls per question, so most questions went to a single provider.
• The default row trades answers for licensing. Licensed sources only is what every Grounder account gets unless it turns that off, and the default row answers only from licensed sources. Every provider except Ozone scored higher than it on both benchmarks.

Method and data: Grounder's benchmark and GET https://api.grounder.net/v1/demo/results.

Sources: 5, 6, 9, 10, 11, 12, 13, 14

How to test whether routing helps you

1. Collect 25 to 100 real questions with the answers you expect.
2. Hold the model, prompt, result limit and maximum number of search calls constant, as Braintrust advises.
3. Run each question through two or three single providers and one router.
4. Record correct answers, tokens, median search latency and cost, and compare cost per correct answer.

The routers

A single provider needs no section: see the table above and "When one provider is enough". Building your own layer means owning failover, deduplication of near-identical URLs, result normalization, logging of which provider answered, and each vendor's data terms.

Cloudflare Web Search API

• What it is: one endpoint that sends a query through Cloudflare's AI Gateway to Ceramic.ai (default), Exa or Linkup, as you choose on each call.
• Pricing: each provider's list price with no markup: Ceramic.ai $0.25 per 1,000 requests, Linkup $5, Exa $7; a 5% fee applies to credits bought through Unified Billing.
• Strengths: one response shape for three providers, switched with one parameter; search logged beside model traffic.
• Limits: open beta; no published rate limits; up to 10 results per call; no domain or date filters.
• Pick it when: you use AI Gateway and want to swap among these three providers.

Sources: 1, 2, 3, 14

OpenRouter web search

• What it is: web grounding inside OpenRouter's LLM API, with engines native, Exa, Parallel, Perplexity or Firecrawl. A beta server tool lets the model decide how often to search.
• Pricing: Exa $7 per 1,000 requests, Parallel $1 to $5, Perplexity $5, Firecrawl $5 plus $12.50 per 1,000 results, plus model tokens.
• Strengths: one key and one citation format across engines.
• Limits: results arrive as model-cited text, not a result list; it is not a standalone search API; retention follows each engine's policy.
• Pick it when: you already buy models through OpenRouter.

Sources: 4, 15, 16

GroundRoute

• What it is: a separate product from groundroute.ai, unrelated to Grounder. One search tool over Serper, Brave Search, Exa, Tavily, Firecrawl and Perplexity, with caching and fallback.
• Pricing: no subscription or minimum. With your own keys you pay $0 on cache misses and share cache savings 50/50; the managed plan passes through engine cost on misses with the same share.
• Strengths: three cache tiers (private exact, pooled across tenants, semantic) and a lower bill when queries repeat.
• Limits: it routes among general web engines on cost and a quality bar. Its pages cite a 170-query benchmark of its own; the pages we read list no licensed-content source or retention terms.
• Pick it when: your main goal is a lower search bill.

Sources: 5, 17

Dogpile Fetch MCP

• What it is: a web search API and MCP server from System1, announced on 23 July 2026. Its press release says it classifies each query and sends it to a backend, and that a Deep mode splits harder questions and removes duplicates.
• Pricing: not checked.
• Strengths: a different design: a classifier in front, and a mode that searches several times for one question.
• Limits: we could not fetch the release and have not tested the product. Its claim that one index surfaces about half of what seven search APIs found together is System1's own analysis.
• Pick it when: you want a metasearch design and can verify it on your own questions.

Sources: 8

Where Grounder fits

Grounder (grounder.net) is a web search API and MCP server. It sends each query to the provider, or pair of providers, most likely to answer it well, merges and deduplicates the results, and marks each result with the provider that found it (found_by lists every provider that returned the URL) and whether it is licensed. Its 10 search providers are Ozone, Tavily, Exa, Brave, Perplexity, You.com, Firecrawl, Octen, Parallel and Valyu. You can pin a search to one provider, name up to 10 providers to search and merge, and cap time, tokens, provider calls and sources per search. Not every search goes to all 10 providers: by default a search goes to one provider or a pair, and fanning out to providers you name is something you set. Across 816 searches recorded on 7 and 8 October 2026, each routed and then sent to all 10 providers, Grounder added a median of 210 ms to the provider calls it waited for; half fell between 128 and 266 ms.

• Licensed results by default: every signed-up account returns only licensed results unless it turns that off. A search can return fewer results than requested. The benchmark rows above show the cost: 14.3% and 12.0% correct with licensed sources only, against 97.8% and 36.1% with all sources.
• Limits: search only, with no fetch, extract or crawl. Search records are kept indefinitely and there is no zero data retention. New search keys are no longer made, so new accounts connect over MCP; the HTTP API and the drop-in endpoints in Tavily's and Brave's response shapes accept only keys made earlier. A provider that is out of credit or down can fail a search; the fallback setting, on by default, moves on to the next provider.
• Cost: your first 1,000 searches are free, up to 100 a day. No card needed. 120 requests per minute per key. Own provider keys can be added for the nine providers that take one; that provider then bills you directly.
• Connect: claude mcp add --transport http grounder https://api.grounder.net/mcp. Other clients: the MCP server guide. More: About Grounder.

Sources: 6, 7, 12, 18

Grounder and robots.txt

Allowed sources only is off by default. When it is on, Grounder returns a result only from a site whose robots.txt allows the provider that found it and does not opt out of search or AI answers. A result that cannot be checked in time is left out, so a search can return fewer results.

Sources: 6

How we compared

We compared on providers behind the option, who chooses the provider, price, free tier, failover, result format and data handling, from each product's own pages checked on 20 October 2026. Benchmark evidence comes from Artificial Analysis's Search Index ↗ (independent), AIMultiple's agentic search test (third party, updated 14 May 2026) and Grounder's own benchmark, which Grounder ran. The protocols differ, so scores are not comparable across them. We found no independent measurement of router overhead or routing gains for any product, including Grounder.

Sources: 9, 10, 12, 19

FAQ

Do I need a search router?

Not if one provider answers your test questions well and you can live with its outages. A router helps when your questions vary, when you need failover, or when you want each result to show which provider found it.

What is the difference between a search router and OpenRouter?

OpenRouter is an LLM gateway that routes model calls; its web search feature adds search to those calls through Exa, Parallel, Perplexity, Firecrawl or the model's native search. A search router chooses among search providers and returns search results.

Does a search router improve accuracy?

Sometimes. In Grounder's benchmark, routing with all sources scored 97.8% on SimpleQA against 96.0% for the best single provider, but 36.1% on SealQA against 44.9% for Parallel. The default setting, with licensed sources only, scored 14.3% and 12.0%. The routing was tuned on those same searches, so expect less on new questions.

Does Grounder send every search to all 10 providers?

No. By default Grounder sends a search to one provider or a pair. You can name up to 10 providers for a search; then it searches those and merges the results.

How much latency does a router add?

Grounder measured a median of 210 ms across 816 searches recorded on 7 and 8 October 2026, with half between 128 and 266 ms. We found no published overhead figure for Cloudflare's, OpenRouter's or GroundRoute's search.

Is GroundRoute the same as Grounder?

No. GroundRoute (groundroute.ai) is a separate product that routes six engines to cut cost. Grounder (grounder.net) routes 10 providers and marks the provider and licensed status of each result.

Next step

Grounder's first 1,000 searches are free, up to 100 a day, no card needed. Connect over MCP at https://api.grounder.net/mcp or read the Docs.