Search APIs for AI apps mostly promise the same thing: give us a query, get back clean, structured content your model can read. That promise is easy to make and hard to compare. Two tools can return the same number of results, with the same tidy formatting, and still leave a model with very different raw material to work with.
So, instead of measuring latency or counting tokens, we ran a test: after each tool finishes searching, what can the LLM actually say?
The test case #
We picked the 2026 global oil crisis as the subject. It's a useful test case because the headline is easy and the substance isn't. Anyone can summarize it as an energy supply shock. Understanding it properly means getting into shipping restrictions, emergency reserves, refinery and fuel shortages, industrial supply chains, and wildly uneven effects from one country to the next.
Then we made the test harder by using one broad query:
2026 global oil crisis
No separate searches for diesel, aviation fuel, LNG, or petrochemicals. I wanted to see what each service surfaced on its own, without being steered toward any particular consequence. Whatever each tool found would become the entire evidence base for the briefing that followed.
Each provider returned 10 search result records. Jina ran with its default search configuration, and RAG Web Browser was set to a maximum of 10 results. The two JSON outputs then became research material for two separate LLM briefings.
The synthesis task
Both briefings used the same prompt, with web search and outside knowledge disallowed. The core instruction:
Using only the supplied research material, write a concise but informative briefing on the 2026 global oil crisis as of October 11, 2026. Help an informed reader understand what caused the crisis, how it developed, and its most significant consequences. The prompt asked for four sections: overview, key developments, consequences, and complexities and outlook. It capped length at 600 to 800 words and required specific figures, source citations, and a clear line between established facts and uncertain forecasts.
Crucially, it never named the developments or industries the answer should cover. That was the point. Whatever showed up in the briefing got there because the search tool found it.
| Element | Setup |
|---|---|
| Products | Jina AI Search API vs. Apify RAG Web Browser |
| Exact query | 2026 global oil crisis |
| Searches | One query per service |
| Result limit | 10 records per service |
| Synthesis | Same briefing prompt, using only the supplied evidence |
| Target answer length | 600 to 800 words maximum |
| Comparison | Source discovery, extraction completeness, reliability, evidence quality, briefing quality, and efficiency |
Results #
Apify produced a slightly stronger analytical briefing, but both services found a substantial shared set of relevant sources, and both briefings were credible.
Apify's decisive addition was a recent September analysis from the Center for Strategic and International Studies (CSIS). That single source let its briefing explain how oil and gas markets adapted to the disruption, rather than just cataloging what the crisis had affected.
Jina's briefing gave more concrete examples of societal impacts, and its extraction of the International Energy Agency's policy tracker included detailed country-by-country tables that Apify missed entirely. Jina's bigger problem was reliability: three of its 10 search results failed outright with HTTP 401 or 403 errors.
Retrieval figures
| Measure | Jina AI Search | Apify RAG Web Browser |
|---|---|---|
| Search result records | 10 | 10 |
| Unique result URLs | 10 | 10 |
| Substantive, usable content records | 6 | 8 |
| Explicit HTTP failures | 3 | 1 |
| Extracted content characters | 253,621 | 372,137 |
| Relative text volume | 1.00× | 1.47× |
Where the result sets overlapped
Both services found these six sources:
| Shared source | Role in the research |
|---|---|
| Wikipedia: 2026 Iran war fuel crisis | Broad background, chronology, and country-level consequences |
| IEA: 2026 Energy Crisis Policy Response Tracker | Government actions and energy-policy responses |
| Brookings: The timing of the impending crude crisis | Strait of Hormuz flows, emergency supply, and the limits of oil-market buffers |
| European Central Bank: Energy shock | Why oil and gas prices responded less dramatically than some expected |
| CNN: The global oil crisis is turning into an everything crisis | Petrochemicals, plastics, manufacturing, and medical supply chains |
| Gedeth: Global Oil Shock 2026 | Implications for global businesses and trade |
Where RAG Web Browser and Jina AI differed
The remaining results are where the two tools parted ways.
Jina found Forbes, Reuters, Reddit, and YouTube. None of them delivered a substantive readable article or transcript in this run. Forbes and Reddit returned 403 errors, Reuters returned 401, and the YouTube result was mostly site navigation. Four promising sources, zero usable evidence.
Apify found CSIS, which turned out to be an exceptionally useful article, plus another economic-impact piece whose crawl failed with HTTP 500. That was Apify's single failed request.
Overall comparison
| Dimension | Jina | Apify | Edge |
|---|---|---|---|
| Relevant source discovery | Strong shared set, but inaccessible additional results | Strong shared set plus extra information | Apify |
| Extraction reliability | Three blocked results | One failed crawl | Apify |
| Content compactness | Smaller overall. Condensed available information efficiently | More page and navigation text | Jina |
| Analytical source quality | Brookings, ECB, IEA, CNN, but missed some Wikipedia content | Same core sources plus CSIS | Apify |
| Downstream societal coverage | More country-specific examples | Comparatively general | Jina |
| Downstream market analysis | Good | More quantitative and interconnected | Apify |
| Content completeness | Missed sites due to blocking, shorter Wikipedia capture | Same sources, complete retrieval, fewer blocks | Apify |
Costs and efficiency #
Both platforms offer free credits, and this test used only a small fraction of either allowance. But the two allowances work differently in a way worth knowing before you build on top of them.
Jina gives new users a one-time allocation of 10M API tokens. Apify gives you $5 in platform credits every month, renewing indefinitely.
| Measure | Jina | Apify |
|---|---|---|
| Content characters | 253,621 | 372,137 |
| Usable content records | 6 | 8 |
| Total charged | 68,642 API tokens out of 10M | $0.026 platform credits out of $5 |
On raw efficiency, Jina produced a smaller output, while Apify retrieved more content with fewer errors. One point in Jina's favor: it doesn't charge tokens for failed requests, so those three blocked results cost nothing.
Conclusion #
In fairness, this was a narrow comparison, and both platforms do considerably more than what they measured.
Jina offers additional services including Embeddings and Reranker. Apify is a marketplace of tools and an automation platform with hosting, scheduling, proxies, and monitoring. RAG Web Browser is one of nearly 100,000 Actors published there.
One query, one topic, one run. Results on a different subject, or on the same subject a week later, could look different. Blocking in particular is volatile: the sites that returned 403s here may well be accessible tomorrow.
That said, RAG Web Browser comes out ahead in this test, but Jina didn't perform badly. Its results were competitive and well-suited to the task, and both services identified the same six core sources.
Apify's biggest advantage was finding the CSIS analysis that Jina missed. That gave the LLM stronger evidence on emergency reserves, alternative trade routes, shifts in Chinese demand, LNG redistribution, and Europe's gas storage vulnerabilities. Apify also hit fewer extraction failures, which matters more than it sounds: a blocked source isn't just a missing document, it's a missing argument the model can never make.
Jina wasn't consistently weaker at extraction either. It found an IEA policy tracker article with extensive country-level tables that Apify missed, and several of its extractions from shared sources were more compact.
The broader point applies regardless of which tool you pick. When an LLM researches a topic for you, the quality ceiling of its answer is set long before the model starts writing. It's set by what the search layer managed to reach.
Note: This benchmark reflects a single test run in October 2026, using one topic, one LLM (GPT 6 Astra Instant), and default settings for both providers. Results can vary with different queries, result limits, or extraction settings. Both platforms offer free tiers and support agent integration via MCP.