Which LLM web search API fits your agent or RAG pipeline?
Start with the output, not the vendor list. If your agent needs fresh candidate URLs, buy search results. If it must quote facts from the pages, add page extraction. If you want the model to return a cited answer in one step, use a grounding tool and evaluate the answer itself. An LLM web search API can mean any of those three things, but they do not have the same cost or evidence quality.
For this October 6, 2026 update, we reviewed the official product and pricing pages linked below and calculated the one-query example from their stated units. We did not run a paid cross-provider answer test, so there is no invented ranking for relevance, latency or accuracy here. The decision method is designed for your own queries and acceptance rules.
What does an LLM web search API actually return?
Your pipeline usually has three jobs. Discovery finds candidate URLs, titles and snippets. Retrieval fetches the content behind selected URLs. Grounding makes an answer that cites evidence. A service can cover one, two or all three, and a vendor's “search” label does not tell you which fields are included in the base price.
| Output contract | What you receive | What you still build or check | Good starting fit |
|---|---|---|---|
| Search results / SERP | Ranked URLs, titles, snippets and sometimes richer search blocks | Fetch page content, remove duplicates, judge source trust, cite evidence | Agents that choose which pages to open; rank or visibility tools |
| Retrieved pages | Page text or markdown for selected URLs | Find relevant passages, handle stale or blocked pages, synthesize with citations | RAG pipelines whose bottleneck is clean source content |
| Grounded answer | Model-generated response with cited sources | Check whether citations support each claim, control model and token cost | Fast question-answering flow with a measurable answer rubric |
A snippet is evidence of what a result card said, not proof of everything on the linked page. A retrieved page can contain the fact you need, but an LLM can still cite the wrong passage. A generated answer may look polished while citing a page that does not support the sentence. That is why the three stages need different acceptance tests.
Seven options and the job each does best
Serpent API for structured search results. Use it when your agent needs engine-specific result cards, URLs and snippets, or when you already own the retrieval and answer steps. Quick Web is one billing unit per call; Deep Web uses one unit per requested page. The published new-credit rate card sets Default at $0.60 per 1,000 Web units, then lower new-credit rates after a qualifying deposit. A search result alone is not full page text or a model answer. Check the response fields against your agent's required contract before integrating.
Brave Search for an independent search index. Its Search plan lists $5 per 1,000 requests, $5 in monthly free credits, and result URLs/snippets including an LLM-context option. Its Answers product is separate: $4 per 1,000 queries plus token charges according to the same plans page. Choose Search when you want to handle retrieval and synthesis yourself; evaluate Answers separately if you want a generated response.
Tavily for a search-oriented agent step. The credit documentation lists pay-as-you-go at $0.008 per credit, one credit for basic search and two for advanced. Extraction costs one basic or two advanced credits for every five successful URLs. That makes Tavily attractive when your agent wants a coordinated search and content workflow, but a “query” and a “credit” are not the same unit once you turn on advanced search or extraction. Price the exact options you request.
Exa for semantic discovery and page text. Its pricing page starts search at $4 per 1,000 and describes text/highlight content options; it also lists a monthly free credit allowance. Exa can suit a research agent that needs semantic matches and readable content, but do not assign its search-only starting rate to a workflow that also asks for extracted text. Test source coverage for your domain and the real content fields you need.
Firecrawl when retrieval is the hard part. Its pricing says Search uses two credits per ten results and a basic Scrape uses one credit per page. The free tier lists 1,000 credits per month; Hobby lists 5,000 credits for $19 monthly, or $16 per month billed annually. It can turn chosen pages into usable text, but a search result and a scrape are separate operations. The pricing page also says a page error such as 403 or 404 can still use a scrape credit, so log page outcomes.
OpenAI web search when the model should answer. The official Responses web-search guide describes a model using web search and returning clickable URL citations. That is a grounding tool inside a model call, not a generic SERP feed for a rank tracker. Tool calls have their own price in addition to model usage; check the current pricing page and count actual tool calls before estimating a bill.
Gemini Google Search grounding for a model-integrated answer. The official grounding guide describes search-backed model answers. Its billing differs by model family: Gemini 3 can bill each search query executed, potentially more than one for a single prompt, while Gemini 2.5 and older models use a per-prompt rule. If you need an answer, inspect citation support and executed-search counts, not merely prompt count.
These are not seven versions of one JSON shape. Brave and Serpent expose search data; Firecrawl emphasizes page retrieval; OpenAI and Gemini integrate search with generation. Tavily and Exa can span search and content depending on options. A lower discovery rate can be a poor deal if your team then spends heavily fetching pages you never needed. A bundled answer can be a good deal if it passes your citation test and saves integration time.
How to compare costs without pretending the units match
Define one workload before opening a price table. Say your agent asks 10,000 questions in a month, runs exactly one search for each, opens two pages for 20% of the questions, and generates an answer with your chosen model. That is 10,000 search operations and 4,000 page retrievals, before repeat attempts or extra model-triggered searches. The table prices documented pieces of that workflow, not an equivalent finished answer from every vendor.
| Documented unit, checked October 6 | 10,000 searches | What the figure excludes |
|---|---|---|
| Serpent published Default, one-page Web units at $0.60/1,000 | $6 of modeled spendable-credit usage | Page retrieval, model tokens, quality checks; lower tiers need upfront qualifying deposits |
| Brave Search at $5/1,000 requests | $50 before free monthly credit | Page retrieval and answer generation; Answers is priced separately |
| Tavily basic at one $0.008 credit per query | $80 PAYG credit cost | Advanced search doubles credits; extraction adds credits |
| Exa search starting at $4/1,000 | From $40 | Content options and any applicable plan or allowance |
| Firecrawl Search at two credits per ten results | Depends on results requested | Basic Scrape is a separate one-credit-per-page step |
| OpenAI / Gemini grounding | Calculate from actual model, tool and executed-search usage | Cannot infer a prompt-to-search ratio or token bill from 10,000 questions |
In the example, 4,000 basic Firecrawl page scrapes would use 4,000 additional credits if all are attempted. Its 5,000-credit Hobby allowance may fit that retrieval step, but any Firecrawl Search calls, other scrapes, or crawl work would consume the same allowance. Tavily documents one basic extraction credit per five successful URLs. Four thousand successful URLs fill 800 five-URL groups in pure arithmetic, but an agent opening two URLs at a time may pay differently because requests are grouped differently. Check rounding on the actual extraction calls; do not assume credits pool across questions. Add observed extraction credits to the 10,000 basic search credits before multiplying by $0.008.
For Serpent, $6 is usage at the Default new-credit rate for 10,000 one-page Web units, not the cash required to start every tier or the whole RAG pipeline cost. The lower $0.30, $0.10 and $0.03 rates for new credit require a single $100, $500 or $1,000 qualifying deposit respectively; that deposit is spendable credit but it is real upfront cash. A Deep request for five pages counts five units even if it is one HTTP call. The cost calculator separates call count, depth, deposit and usable results.
Most important, the table has no accuracy score. It cannot tell you whether two services found the same pages, returned full text, or supported the final answer. Do not market the cheapest search line as the cheapest grounded answer.
A practical pipeline for an agent
For a question-answering agent, start with a small result set. Save each candidate URL, title, snippet, engine, query and retrieval timestamp. Deduplicate by canonical URL and domain. Open only the sources needed to answer the question, and keep the retrieved passage or page snapshot beside the final citation. That lets you revisit a citation when a user challenges a sentence.
Ask the model to distinguish “the search snippet says this” from “the opened page supports this.” Require a citation for factual claims that depend on the web, then verify each citation against the page content. If a page cannot be retrieved, the agent should say the evidence is unavailable rather than fill the gap from the snippet. If your agent uses only search snippets, label that narrower evidence base in the product.
A rank tracker has a different job. It may need organic positions, ads, local blocks or AI Overview fields from the result page; whole-page markdown and generated prose are not substitutes. For that case, compare structured SERP output and the fields your report actually reads. The SERP API for AI agents guide covers a search-first integration, while the real-time RAG tutorial focuses on retrieval and citation flow.
Run this small acceptance test before choosing
Build a dated set of 50 questions from your actual product: recent facts, niche technical questions, local intent, ambiguous names and questions with no trustworthy answer. Freeze the prompt, country, language, device and time window where each API permits it. Save the raw outputs; a screenshot of the final answer alone is not enough to debug a bad citation.
- For discovery, count returned URLs that are reachable, relevant and distinct. Record whether a result card has the title and snippet your application needs. Do not award a pass just because the HTTP response is 200.
- For retrieval, check whether the returned page text contains the needed passage, whether the page is fresh enough, and whether extraction truncated a table, code block or key paragraph.
- For answers, score each factual claim against its cited passage. Mark unsupported, stale and missing citations separately; one citation at the end of a paragraph may not support every sentence.
- For cost and speed, log each request, search query executed by a grounding tool, opened URL, credit, token use, latency, repeat attempt and billable failure. Divide total cash or credit use by accepted answers, not by prompts alone.
Repeat a few questions on another day. Search results and live pages change; a single favorable query can make a weak integration look strong. If you see “no result,” check a comparable browser search and a second location before deciding that the document is absent from the web. Your acceptance sheet becomes the evidence behind the vendor decision.
Which route should you choose?
Choose search results first when you already have a retrieval layer, need engine-specific SERP fields, or want control over which sources the agent opens. Serpent and Brave Search belong in that trial. Choose a content-oriented service when page extraction is the bottleneck; evaluate Exa, Firecrawl and Tavily's extraction option on your real domains. Choose model-integrated grounding when time to a cited answer matters most; evaluate OpenAI web search, Gemini grounding and Brave Answers by citation support and full tool/token cost.
You can mix the layers. A cheap search call can identify candidates, and a content service can open only the few pages that matter. But mixing providers adds latency, handling for failed page fetches and a more complicated bill. If a single service meets your acceptance rule, simpler may be worth a higher nominal call price.
Whichever route you choose, keep the vendor behind a small adapter that returns source_url, title, snippet, retrieved_text, retrieved_at and citation_span as separate optional fields. Do not silently turn a search snippet into full page evidence. That boundary makes later provider changes and citation audits much easier.
Start with the search contract
If structured Web results are the missing layer, inspect Serpent's SERP API, its documented response and per-unit pricing. If you need retrieved pages or a finished cited answer, budget and test those steps too.
Model search usage and depthFAQ
Is an LLM web search API the same as a SERP API?
No. A SERP API returns structured search results. An LLM search tool may also retrieve pages or have a model write a cited answer. Compare the output fields your application needs, not the label.
Can an agent ground an answer on snippets alone?
It can use snippets as a narrow source, but a snippet may omit the qualifying context on the page. For high-stakes factual claims, open the source page and verify the cited passage.
Does one prompt always equal one search charge?
No. Some tools can execute more than one search query for a prompt, and others charge separately for extraction or generation. Gemini 3 grounding, for example, can bill each executed search query. Log the actual units.
Which option gives full page text?
Firecrawl is built around scraping pages to content, while Exa and Tavily offer content or extraction options. A structured SERP result usually includes links and snippets, not the complete linked page.
What should I measure in a trial?
Use a dated set of real questions. Score relevant and reachable URLs, complete page passages, claim-level citation support, latency and total cost per accepted answer. Keep raw outputs for audit.
Can I combine a SERP API with another retrieval tool?
Yes. Search for candidate URLs, open only selected pages, then generate an answer from the retrieved passages. Budget the search, page and model steps separately and preserve the link between each citation and source passage.
