TL;DR
- Exa is our default for an agent building account briefs from web pages. Its Search response can include text and highlights, and its company-oriented retrieval is useful for discovery.
- Tavily fits a compact search-and-extraction pipeline. Specify the search depth: Advanced uses two credits, while Basic, Fast and Ultra-fast use one.
- Brave fits conventional web research with country and language controls, particularly when the team already has a page-fetching layer.
- SerpAPI fits Google-specific and local-market research. Buy it for the result environment and structured fields, rather than expecting the lowest general-search cost.
- At 1,000 accounts and three searches each, fetching selected pages and processing their text can matter more than the difference between search-request prices.
Compare the output, not just the request
An account-research agent needs more than a relevant-looking list of links. It might need the official company domain, a current hiring signal and the paragraph supporting a useful sales observation. Search locates candidate sources. Retrieval supplies the text. The agent turns that text into a brief or proposed CRM property.
Those stages create different costs. A low-priced search result may need another page-fetching service. A richer response can reduce that work, but sending every retrieved page into a model increases processing time and token consumption. An answer endpoint performs more synthesis than a search endpoint and should have its own budget.
This comparison retains four distinct options: Tavily, Exa, Brave and SerpAPI. It focuses on account research rather than buying a contact database. None of the search prices below buys unrestricted verified emails or phone numbers. For the execution layer, see running GTM from a coding agent.
The four APIs compared
| API | Useful response | Current public price basis | Main controls | Best fit |
|---|---|---|---|---|
| Exa | URLs with requested text/highlights; company and other search categories | Instant $4/1,000 requests; Fast/Auto $7, up to 10 results | Domain filters, publication dates, content-cache age, search mode | Account discovery and briefs needing source passages |
| Tavily | Ranked results with relevant content; optional raw content and generated answer | $0.008/credit PAYG; one-credit Basic/Fast/Ultra-fast, two-credit Advanced | Domain restriction/preference, time filters, country boost, depth | Managed search plus selective extraction |
| Brave | Web results and snippets; separate LLM Context and Answers interfaces | Search $5/1,000 requests | Country, language, freshness and Goggles reranking | Conventional web research and localised queries |
| SerpAPI | Structured results from the selected engine, including Google-specific features | Best Effort Developer $75/month for 5,000 searches | Engine, location, country, language, date filters and cache choice | Google/local result fidelity and specialised search surfaces |
Prices were checked on 30 September 2026, before tax. Exa and Brave publish recurring free credits; Tavily includes 1,000 free monthly credits. The worked comparison uses gross rates so introductory or account-specific allowances do not determine the recommendation. Exa pricing, Tavily credits, Brave pricing, SerpAPI pricing.
Product capabilities and fit
1. Exa: best default for account discovery and source-backed briefs
Exa is a good fit when the research question describes a company or problem rather than reproducing a particular search-engine query. Its Search API has company and people categories, domain controls and requested text or highlights. That lets an agent discover candidate accounts and retain the passage that made a result relevant. Exa Search reference.
Use company discovery and event research as separate requests. The company and people categories do not support the publication-date and domain-exclusion filters available to general search. A search for companies fitting an operational description is different from a search for their announcements this month.
The live pricing table distinguishes $4/1,000 Instant requests from $7/1,000 Fast/Auto requests, with up to ten results. Additional results above ten cost more. Text and highlights are part of the Search offering; separately requested Contents and AI summaries have their own prices. Deep Search starts at a different $12–$15/1,000 basis. Exa pricing.
Non-fit: use SerpAPI if the requirement is Google's exact result structure or a specific local search surface. Exa's company discovery should not be treated as a complete account universe, and company categories should not inherit unsupported freshness filters.
Verdict: start with Fast/Auto for a recurring account brief. Use Instant when response speed matters more than search depth. Request a few useful passages rather than collecting every result's full text.
2. Tavily: best for an explicit search-and-extraction process
Tavily brings search content and extraction into a convenient agent-oriented API. Basic, Fast and Ultra-fast cost one credit per search; Advanced costs two. The modes also change the relevance/latency trade-off and the content returned. The current reference distinguishes multiple relevant snippets from Ultra-fast's single NLP summary. Tavily Search reference.
For official-company research, include_domains is a hard restriction by default. Its preference mode permits other domains as well. Country is a ranking boost for general search, rather than a guarantee that every result belongs to that market. These distinctions help a team avoid mistaking a preferred source for an enforced source rule.
PAYG costs $0.008 per credit. Project is $30/month for 4,000 credits and Bootstrap $100 for 15,000. Basic Extract charges one credit per five successful URLs; Advanced Extract charges two. Those units make batching selected pages useful. Tavily credits and extraction pricing.
Non-fit: do not choose Advanced for every domain lookup simply because it sounds better. Explicitly set the depth when a predictable budget matters; automatic parameter selection can upgrade a search to Advanced.
Verdict: choose Tavily when the team wants one provider for search and selective page extraction. Start ordinary company lookups at Basic or Fast and reserve Advanced for the more demanding research questions.
3. Brave: best for conventional web search with market controls
Brave provides search from its own web index. Its Search product includes URLs, snippets and additional LLM context, while Answers is a separately priced synthesis product. For an agent that already knows how to fetch and interpret pages, the search route is straightforward. Brave Search API.
Its documented country and language parameters suit territory-specific research. Freshness supports recent-day, week, month, year and custom-date windows. Goggles can rerank or filter results. These are useful when a UK sales team wants relevant company pages rather than accepting a default US-oriented result set. Brave web-search documentation.
Search costs $5/1,000 requests and includes $5 monthly free credits. Answers has a different request-plus-token basis, so it should not be budgeted as ordinary Search. The public Search capacity is 50 queries per second; that is throughput, not a latency guarantee. Brave pricing.
Non-fit: a conventional result snippet is not the full source page. If the agent needs the wording of a hiring brief or a detailed product limitation, include an extraction step or use the appropriate context interface. Do not assume Brave will reproduce Google's ranking or local-result fields.
Verdict: choose Brave when conventional results, geographic controls and a low request rate fit an existing retrieval stack. Its economics are especially attractive when page extraction is already paid for elsewhere.
4. SerpAPI: best for Google-specific and local research
SerpAPI is useful when the result environment is part of the job. Google's organic results, local information and related structured fields can matter for location-based prospecting or researching how a business appears to buyers in a particular city.
The Google endpoint supports location, gl for country, hl for language and advanced filters through tbs. Specify the location and country together for more consistent territory behaviour. The response also identifies search parameters and provides search metadata. Google Search API reference.
Best Effort Starter costs $25/month for 1,000 searches, Developer $75 for 5,000, and Production $150 for 15,000. Only successful searches count; cached and failed requests do not. The Google endpoint documents a one-hour cache for identical query parameters. SerpAPI pricing, cache behaviour.
Speed is a purchase choice. Ludicrous Speed doubles the Best Effort price, rather than merely changing a request parameter at the same cost. Developer therefore becomes a $150 monthly basis with that option. SerpAPI speed pricing.
Non-fit: it is an expensive default for a brief that only needs ordinary company pages and selected text. Its structured search response also does not remove the need to retrieve a destination page when the brief needs detailed evidence.
Verdict: buy SerpAPI for a particular engine, local result or structured surface. Choose the speed option around an actual interactive requirement; batch account preparation can often tolerate Best Effort.
Common sales-research queries and controls
A practical research queue should distinguish identity, current signals and market discovery. These example queries are a common input set, not results from a comparative API test.
| Query | Buying purpose | Useful controls and expected evidence |
|---|---|---|
Wise revenue operations careers |
Find a relevant hiring source | Prefer the company's official domain; retain the job page and its actual date |
Celonis customer announcements September 2026 |
Find a recent commercial event | Use an event-date window and official news sources; distinguish announcement date from page update date |
Monday work management official website |
Resolve an ambiguous name | Include the product context and inspect the final domain before attaching a CRM ID |
manufacturers using Salesforce field service UK |
Discover accounts fitting a problem | Broader semantic discovery and named case-study sources; do not infer the entire installed base |
industrial automation suppliers Manchester |
Find geographically relevant businesses | Country/location parameters and an appropriate local-result interface |
For the hiring query, Tavily can restrict domains and return relevant content. Exa can request passages from a domain-restricted general search. Brave can combine the query with market controls and a freshness window. SerpAPI can reproduce a Google search with a specified location and country. The goal is the same supporting job page, but the response structures differ.
For event queries, the date filter does not prove the event happened during the requested period. A page may have been updated this month while describing an old announcement. Tavily documents optional removal of undated results; Exa distinguishes publication-date filtering from content-cache freshness; Brave describes age based on the page's relevant reported date. Use the event date in the article or brief, rather than simply copying the retrieval date. Tavily date parameters, Exa freshness parameters, Brave freshness.
A compact citation record should contain the account domain, exact source URL, observation, event date where known, retrieval time and API mode. Keep a supporting passage alongside a consequential observation such as “hiring a revenue operations manager”. A search score or generated answer is not that passage.
Worked cost for 1,000 account briefs
Assume 1,000 accounts, three searches each and two selected source pages per account. That is 3,000 search requests and 2,000 selected pages. Keep each search to ten results or fewer. This is a pricing model, not a measured API run.
For the link-oriented routes, use Tavily Basic Extract as a concrete common extraction basis: batch five successful pages per request, producing 400 credits for 2,000 pages. At gross PAYG rates that is $3.20. Other retrieval services or an existing crawler can change that line.
| Route | Search calculation | Selected-page extraction | Gross provider basis |
|---|---|---|---|
| Exa Fast/Auto with text/highlights | 3,000 × $0.007 = $21 | Already returned by the selected Search design | $21 |
| Exa Instant with text/highlights | 3,000 × $0.004 = $12 | Already returned by the selected Search design | $12, with a different depth/speed trade-off |
| Tavily Basic/Fast plus Basic Extract | 3,000 search credits × $0.008 = $24 | 400 credits × $0.008 = $3.20 | $27.20 |
| Tavily Advanced plus Basic Extract | 6,000 search credits × $0.008 = $48 | $3.20 | $51.20 |
| Brave Search plus Basic Extract | 3,000 × $0.005 = $15 | $3.20 | $18.20 |
| SerpAPI Developer Best Effort plus Basic Extract | $75 subscription covers 3,000 searches | $3.20 | $78.20, with unused monthly search capacity |
The Tavily examples deliberately price separate extraction. Requested raw content can reduce that need where it supplies the required text. Exa's example uses its included Search content; separately refreshing 2,000 pages through Contents at $1/1,000 for one content type adds $2. Extra AI summaries, more results, deep research and synthesis-model usage sit outside these bases.
At this volume, Tavily Advanced plus extraction uses 6,400 credits. Project's 4,000-credit package plus 2,400 PAYG overage credits would cost $30 + $19.20 = $49.20 before any other allocation. PAYG can still be preferable after applying available free credits. Bootstrap's $100 commitment buys considerably more capacity than this example needs.
The table omits free allocations, taxes, downstream model inference and labour consistently. It therefore explains the purchased operation rather than pretending to be an all-in invoice. A free allowance is useful for a small start, but it does not make every retrieved fact free to process.
Text volume creates another material difference. Sending 2,000 pages at an assumed 2,000 tokens each uses four million input tokens. Passing 500-token supporting passages instead uses one million. At a declared illustrative model input rate of $2/million tokens, that is $8 versus $2. The $6 difference is a planning assumption, not a vendor model quote.
On Fast/Auto, Exa's $21 search-content basis is $2.80 above the Brave-plus-extraction example. At $50/hour, just 3.36 minutes of operator time equals that difference. Choose the source quality and retrieval design that reduce work; the cheapest search request is not necessarily the cheapest completed brief.
Latency evidence and pipeline design
There is no measured four-provider latency ranking in this article. Published evidence supports choices of mode, not an invented universal winner.
| Provider | Available latency evidence | Practical interpretation |
|---|---|---|
| Exa | Vendor pricing page describes configurable Search latency of 180 ms to one second | Instant is a speed-oriented option; fetching and model synthesis still add time |
| Tavily | Modes explicitly trade latency against relevance; API returns response_time |
Keep the chosen mode with the timing record; a sample response is not a benchmark |
| Brave | Search advertises 50 queries/second capacity | Useful for concurrency planning; it does not establish milliseconds per completed brief |
| SerpAPI | Vendor reports Ludicrous Speed averaging 2.2× faster than its own baseline | A relative vendor result; it does not rank it against the other three APIs |
The supporting sources are the Exa price/mode description, Tavily response reference, Brave capacity and SerpAPI speed page.
For an interactive seller assistant, issue independent identity and signal searches concurrently, then fetch only the selected pages. For a nightly account queue, higher relevance and fewer manual exceptions can be more useful than shaving a fraction of a second from a search call.
A simple timing model makes that distinction clear. Assume three search calls at 0.8 seconds each, two selected-page fetches at 1.2 seconds each and one second of synthesis. Sequential execution takes 5.8 seconds. Running each independent group concurrently takes approximately three seconds, before queueing and orchestration overhead. Those timings are architectural assumptions, not observed provider performance.
Store client elapsed time as well as any provider timing field. Include retries and extraction in the completed-brief measurement. A fast cached response should not be compared directly with another provider's fresh crawl without recording the difference.
For current product or pricing questions, fetch the directly applicable official page. For broader account discovery, let the index find candidates first. Our signal-based outbound playbook connects those retrieval choices to a commercially useful observation.
Which API to choose
Choose Exa Fast/Auto for a new account-research agent whose main output is a brief with supporting passages. Choose Tavily when search depth, selective extraction and one-provider implementation are the main priorities. Choose Brave when conventional results and market controls fit an existing retrieval layer. Choose SerpAPI when Google or a specialised local surface is part of the requirement.
One primary provider plus a narrow fallback is usually enough. For example, Exa can handle descriptive company discovery while SerpAPI handles a specific local-result request. Running all four providers for every account increases cost and duplicate sources without necessarily improving the brief.
FAQ
Which is the best default for a sales-research agent?
Exa is our default for source-backed account briefs because requested text and highlights fit the job well. Tavily is a strong alternative for a managed search-and-extraction process. Brave and SerpAPI have clearer advantages when conventional market controls or a particular search surface determine the work.
Is a generated answer the same as a citation?
No. A useful citation links to the page and identifies the passage supporting the observation. A generated answer can help summarise research, but the agent should preserve the source evidence for consequential account claims rather than citing only its own synthesis.
Do date filters guarantee a current buying signal?
They reduce the candidate set. They do not establish when an event occurred or whether a job is still open. Use the event's wording and date from the supporting page, and distinguish an old event on a recently updated page from a new signal.
When is SerpAPI worth the higher subscription cost?
When Google-specific structure, a named location or a specialised result type affects the commercial decision. A local supplier search can justify that interface. A general account brief that only needs official URLs and paragraphs often has a cheaper route.
How many searches should we budget per account?
Three is a useful starting design for identity, a current signal and one qualification fact. Reuse the known domain and cached stable company information where appropriate. Give a higher-value account a targeted additional query rather than allowing every missing detail to start an unlimited research loop.
