Newscatcher CatchAll API: I Tested the Recall-First Web Search Built for AI Agents

Newscatcher CatchAll API: I Tested the Recall-First Web Search Built for AI Agents

I asked the Newscatcher CatchAll API one stubborn question: find every warehouse fire in Europe this quarter. Not the top ten. Every one. A normal search API hands you a ranked page of links and quietly drops the long tail, which is fine for "best pizza near me" and useless for an enumeration job. Roughly 15 minutes later, in Base mode, CatchAll came back with regional-press and trade-publication events that the ranking-based APIs in our best-ai-search-apis-2026 test never surfaced. Each result arrived as one structured JSON object, not a link. That gap is the whole story.

Here's the short version before we get into the wiring.

Key Takeaways

• CatchAll is a recall-first web search API: it returns structured event records, not a ranked SERP.
• Newscatcher reports 79.8% recall and F1 0.705 across 32 queries; the homepage rounds this up to 86%.
• Two modes: Lite (seconds, roughly 100-result cap) and Base (asynchronous, around 15 minutes, no cap).
• Best for enumeration tasks (compliance, competitive intel, supply-chain monitoring); it is not a SERP rank tracker.

Recall-first search optimizes for finding everything, where ranking-first search optimizes for ordering the few things it surfaces. Keep that one sentence in your head and the rest of this review clicks into place.

What Is CatchAll? (Recall-First Web Search, Explained)

CatchAll is a recall-first web search API from Newscatcher: instead of a ranked list of the top ~10 links, it returns a deduplicated set of structured event records, each one a single JSON object with citations and extracted entities. Recall-first means the goal is completeness, finding every relevant event, not ordering a short list by relevance.

Newscatcher frames it with a blunt intuition pump: if 200 valid events exist and your system surfaces 5, your recall is 2.5%. For "what's the best laptop," that's fine. For "every regulatory action against EU fintechs this year," a 2.5% answer is worse than useless because you can't see what you're missing.

That's the category gap CatchAll targets, and it's worth understanding even if you never sign up. The YC launch pitch calls it a "recall-first web search API," and you can read the positioning straight from Newscatcher's CatchAll Web Search API product page. The honest framing: a SERP API answers "what should I read first?" CatchAll answers "what is the complete list?"

How CatchAll Works: The Retrieval + Validation Pipeline

CatchAll runs a five-stage pipeline: query planning rewrites your prompt into multiple retrieval angles, large-scale retrieval scans 50,000+ pages per job, the Leiden algorithm clusters related pages into single events, an LLM validates each cluster against your query, and the survivors come back as structured JSON. Recall comes from the scan; precision comes from the validation.

Let's break those down quickly:

1. Query planning. Your one query becomes several retrieval prompts covering different phrasings and event types, so "warehouse fire" also catches "blaze at logistics depot."
2. Large-scale retrieval. Newscatcher says a single job pulls 50,000+ pages at around 10,000 pages per minute with no result cap, reaching regional press, trade pubs, and regulatory filings that mainstream SERPs bury.
3. Clustering. The Leiden algorithm groups densely-connected pages into communities. In plain terms: 30 articles about the same Rotterdam fire collapse into one event, not 30 rows.
4. LLM validation. Each cluster gets scored against your query, and irrelevant ones are dropped. This stage costs real LLM calls, so in a production agent stack you'd want to route those through an LLM gateway to control spend and fallback.
5. Structured output. One JSON object per validated event, with a dynamic schema.

Newscatcher reports indexing more than 2 million real-world events daily, with new events landing in under hours. The technical write-up, "The Architecture of Completeness," covers the Leiden and validation internals if you want the deep cut.

This is where my warehouse-fires test lived. I ran the enumeration query in Base mode, the job took roughly 15 minutes, and the value showed up exactly where the pipeline promises it: regional and trade sources, clustered into discrete events, that a ranked SERP would have buried below the fold or skipped entirely.

Key Features: Monitors, Watchlists & Event Extraction

Beyond one-off search, CatchAll ships Monitors and Watchlists. Monitors are scheduled re-runs (minimum hourly) that return only the new, deduplicated events since the last run. Watchlists filter by entity, with a 1-10 relevance score and entity resolution that matches the same company across languages and jurisdictions.

Monitors turn a one-off search into a standing watch: each run returns only the events that are new since the last. That's the difference between "search the web today" and "tell me the moment something changes."

One honest caveat on the "real-time event monitoring API" framing: "real-time" here means hourly re-runs, not sub-second streaming. If you need millisecond push notifications, this isn't that. For a compliance team that checks twice a day, hourly is more than enough. The Company Watchlist is the standout for competitive intelligence, since it resolves "Acme GmbH," "Acme Inc," and "Acme Holdings" into one tracked entity instead of three noisy ones.

The Structured JSON Output (With a Real Example)

Every result is one event as one JSON object: a cluster_id, a title, a relevance score, an entities array, and a source_citations array. No HTML scraping, no link list to parse. This is what makes CatchAll a genuine structured web search API rather than a SERP wrapper.

Here's a truncated event from my warehouse-fires run, lightly cleaned but real in shape:

{
  "events": [
    {
      "cluster_id": "evt_8f21a",
      "title": "Fire at logistics warehouse near Rotterdam",
      "relevance": 9,
      "entities": [
        {"name": "Rotterdam", "type": "location"},
        {"name": "Maasvlakte", "type": "facility"}
      ],
      "source_citations": [
        {
          "url": "https://...",
          "publisher": "regional trade press",
          "published_at": "2026-..."
        }
      ]
    }
  ]
}

Notice the source_citations array points at a regional trade outlet, exactly the kind of source a ranking API deprioritizes. Because each event is already structured, you can drop the validated records straight into a RAG pipeline or store and embed them in a vector database without a scraping or cleanup step in between. That saved-step is the quiet productivity win.

How Do You Call CatchAll in Python? (Code Quickstart)

You get an API key, POST your query to /v3/search with the x-api-token header, and parse the events array. That's the whole loop. Here's a minimal Python call:

import requests

resp = requests.post(
    "https://api.newscatcherapi.com/v3/search",
    headers={"x-api-token": "YOUR_API_KEY"},
    json={"query": "warehouse fires in Europe", "page_size": 10},
)
events = resp.json()["events"]
for ev in events:
    print(ev["title"], "—", ev["relevance"])

Pro tip and gotcha in one: Lite mode returns in seconds but caps around 100 results, while Base mode is asynchronous and takes roughly 15 minutes for a deep job. For Base, you submit and poll rather than block on a single call, so design your agent to fire-and-check, not wait. If you're wiring CatchAll into an agent, you'd typically call it as a tool via function calling. Confirm the exact request params and the Lite-versus-Base flag against the CatchAll docs before you ship; the auth header is x-api-token.

How Much Does CatchAll Cost? Is There a Free Tier?

Pricing is usage-based and pay-per-validated-record, roughly $0.10 per record, and zero results means zero charge. There's a free tier of about 2,000 credits on signup plus around 10 searches a month, no card required, so you can run a real enumeration test before committing.

That zero-results-zero-charge model matters for enumeration work: a query that legitimately has no matching events doesn't burn budget. On the common question of whether Google's search API is free, native Google and Bing search aren't this. They return ranked links, not validated structured events, and Bing's Search API is being retired, which is part of why independent indexes are having a moment.

Real-World Use Cases

CatchAll fits any job where missing one item is the failure mode. Compliance and regulatory tracking lean on Monitors plus its regulatory-filing coverage. Competitive intelligence runs on the Company Watchlist. Supply-chain monitoring is the warehouse-fires pattern, watching for disruption events. Market research uses enumeration over trade press.

The pattern across all four: you're building a complete list, then acting on it, often inside an automated web search API for AI agents workflow. A few concrete shapes:

• Compliance: standing Monitor on enforcement actions in your sector, hourly.
• Competitive intel: Watchlist on three rival entities, resolved across their legal names.
• Supply chain: enumeration of disruption events (fires, strikes, recalls) near your supplier facilities.
• Market research: one-shot Base-mode scan of every product launch in a niche this quarter.

The Benchmarks: Is CatchAll Really 3x Better Than Exa?

In Newscatcher's own March 2026 benchmark of 32 queries, CatchAll reports F1 0.705 and 79.8% recall (4,807 events), winning 27 of 32 queries against Exa Websets, Parallel AI FindAll, and OpenAI Deep Research. Newscatcher describes that as roughly 3x more relevant events than the field. Every number here is the vendor's own.

Now the honest part. Newscatcher's own rigorous benchmark says 79.8% recall; the homepage rounds it to 86%. We'll quote the lower number. The 79.8% figure comes from the dated, detailed 32-query product-page table, while the 86% headline is a rounder claim from a different cut on the homepage and a blog post. Both are Newscatcher's. I lead with the lower one because quoting a vendor's own more-conservative internal number is the trust move a marketing page can't make. Either way, the directional finding held up in my testing: recall is genuinely higher than the ranking-first tools. For the full field, see how CatchAll ranks against 12 other AI search APIs in our best AI search APIs roundup, and check the raw table yourself on Newscatcher's product page.

Honest Limits: What CatchAll Is NOT For

CatchAll has four real limits the marketing pages bury, and you should weigh them before building. It is not low-latency, not uncapped in its fast mode, not yet plug-and-play for agents, and not a rank tracker. None are dealbreakers, but each rules out a use case.

• Base mode is asynchronous (~15 minutes per job). Wrong tool for a chatbot that needs an answer in two seconds.
• Lite mode caps around 100 results. Want deep recall fast? You can't have both; deep recall pays the latency tax.
• No official MCP server yet. You wrap the REST endpoint yourself. If you want it as a native agent tool, you'd wrap the REST endpoint as an MCP server, the same way we build the MCP servers we already use.
• It's not a SERP or rank-tracking tool. It won't tell you where you rank on Google. Different job entirely.

This section is the part no first-party page will write for you. If the async latency or the missing MCP server kills your use case, better to learn it here than after integration.

CatchAll Alternatives & When to Choose Them

CatchAll wins on raw recall for enumeration, but it isn't the right call for every search job. Here are seven real alternatives, each with the honest "pick this instead" condition. No strawmen.

The quick heuristic: enumeration and monitoring point to CatchAll, conversational RAG points to Tavily, semantic discovery points to Exa, and rank tracking points to SerpAPI.

How Techsy Uses Recall-First Search in Agent Builds

At Techsy, we ship AI agents for B2B clients, and recall-first search slots in cleanly for enumeration and monitoring jobs: think a compliance agent that needs the complete list of enforcement actions, not the top five. We'll reach for a recall-first API like CatchAll there, and honestly reach for Tavily or Exa when the task is conversational RAG or semantic lookup instead. Picking the wrong search primitive is one of the most common agent-build mistakes we fix. Want a hand choosing? Get a free consultation.

Frequently Asked Questions

What is the Newscatcher CatchAll API?

CatchAll is a recall-first web search API from Newscatcher. Instead of a ranked list of links, it returns structured event records, one JSON object per real-world event, each with source citations and extracted entities. It's built for AI agents, enterprise research, and monitoring tasks where finding every relevant event matters more than ordering a short list.

How is CatchAll different from a normal (SERP) search API?

A SERP API ranks and returns the top handful of links, optimizing for "what should I read first." CatchAll optimizes for completeness, scanning 50,000+ pages per job and clustering them into deduplicated structured events. You get one object per event with citations and entities, not an HTML results page you have to scrape and parse yourself.

Is CatchAll really 3x better than Exa Websets?

Newscatcher reports it on their own March 2026 test of 32 queries: CatchAll at 79.8% recall and F1 0.705 versus Exa Websets at 19.6%, winning 27 of 32 queries, which they frame as roughly 3x more relevant events. Note the homepage rounds recall up to 86% from a different cut. All figures are the vendor's own; treat them as attributed, not independently audited.

How much does CatchAll cost? Is there a free tier?

Pricing is usage-based and pay-per-validated-record, around $0.10 per record, with zero charge when a query returns no results. The free tier gives roughly 2,000 credits on signup plus about 10 searches a month, no card required. That's enough to run a real enumeration test against your own use case before committing budget.

How fast is CatchAll?

It depends on the mode. Lite returns in seconds but caps at roughly 100 results. Base is asynchronous and takes about 15 minutes per job, with no result cap, for deep enumeration. For Base jobs you submit and poll rather than block on one call, so it's wrong for anything that needs a sub-second answer like a live chatbot.

Does CatchAll have an MCP server?

Not an official one yet. To use it as a native agent tool today, you wrap the REST endpoint yourself, the same pattern covered in our MCP guide. It's a thin wrapper around a single POST to /v3/search, so building a small MCP server around it is straightforward if your stack already speaks the protocol.

What are Monitors and Watchlists?

Monitors are scheduled re-runs, minimum hourly, that return only the new deduplicated events since the last run, turning a one-off search into a standing watch. Watchlists filter results by entity with a 1-10 relevance score and resolve the same company across languages and jurisdictions. Together they cover compliance tracking and competitive intelligence without re-querying the full web each time.

Can I use CatchAll for SEO rank tracking?

No. CatchAll returns validated structured events, not search-engine rankings, so it won't tell you where your page sits on Google. For rank tracking, SERP features, or mirroring exactly what Google shows, use SerpAPI or Serper instead. CatchAll and rank trackers solve genuinely different problems despite both touching "web search."

What use cases is CatchAll best for?

Enumeration and monitoring tasks where completeness matters: compliance and regulatory tracking, competitive intelligence, supply-chain disruption monitoring, and market research over trade press. The common thread is that missing a single relevant event is the failure mode, which is exactly what recall-first search is designed to prevent. For conversational RAG or semantic lookup, a ranking-first tool fits better.

Bottom line: recall-first is not ranked, and that's the point. CatchAll trades latency for completeness, and for enumeration jobs that's the right trade. Run the free tier on your own hardest query before you decide, since you can try CatchAll's free tier with no card. If the async wait or missing MCP server is a dealbreaker, an alternative from the table above will serve you better.