TL;DR
- Tavily: Search plus extraction with explicit depth and credit controls.
- Brave: Search and LLM-context options for teams that value market controls and can manage evidence retrieval.
- SearchApi: Google, Jobs, News and other named result surfaces, with a separately designed page-content step.
- SerpApi: An engine-specific interface and location controls; check throughput and monthly plan limits.
- Parallel: Search excerpts and URL extraction are separate tools; choose the right mode and price each request.
- Preserve source identity, event date, retrieval time and cache status separately. A search hit or plausible generated answer is not proof of a current buying signal.
If you are replacing Exa in an account-research agent, first identify the capability you are actually using. Finding companies from a descriptive query, returning passages with search results, refreshing a known URL and reproducing a Google Jobs result are different operations. A provider that matches one may require another service or a rewritten workflow for the rest.
Tavily is a strong starting point for an integrated search and extraction pipeline. Parallel is worth testing for search with excerpts and separate URL extraction. Brave Search API fits conventional web search and market controls; SearchApi and SerpApi are especially relevant when a particular search-engine result or structured surface matters. None is a guaranteed drop-in replacement for Exa's company discovery, ranking, content or freshness behaviour.
The decision should turn on supported account claims, not the number of URLs returned. Test each candidate against the same identity, hiring, company-news and product questions, then record what was retrieved, when it was observed and what the evidence supports.
What Exa capability must be replaced?
Exa can return ranked sources and requested content alongside Search results. Its separate Contents operation can retrieve known pages, with cache and live-fetch controls. If your application only uses the returned highlights, replacing Search and then fetching every result again adds cost and may change the evidence available to the model. If it depends on a company-specific discovery mode, a conventional search result for a rewritten query is a different candidate set.
Start by collecting representative requests: query, mode, requested result count, domain and date filters, content options, final URLs, output schema and the reason each accepted account was selected. Include difficult cases: a common company name, a subsidiary, an old news story updated this month and a job that has closed. Keep the original source passages and any permissible retention limits.
Exa's pricing lists Lite Search at $4 and Auto Search at $7 per 1,000 requests for up to ten results, with separately priced content types and additional results. Its pricing documentation identifies the older Instant and Fast search types as legacy; inventory the actual type used before comparing an existing integration with a new one. Its Contents documentation distinguishes content included in a search response from separate URL retrieval. Check exact options and current terms in your implementation. A published rate for one mode is not the cost of every Exa research workflow.
Our search API comparison covers the broader category. This guide focuses on replacing an Exa implementation without losing its evidence and freshness contract.
How do the five alternatives compare?
| Alternative | Principal input and output | When it is useful | What to preserve or verify |
|---|---|---|---|
| Tavily | Search query returns results and content; Extract accepts URLs | One-provider search and selected-page retrieval | Search depth, extraction credits and domain filters |
| Brave Search API | Web query returns results; LLM Context supplies relevant chunks | Conventional search with country and language controls | Context coverage, selected-page retrieval and date semantics |
| SearchApi | Search-engine query returns structured results for a chosen surface | Google Jobs, News or another named result environment | Engine parameters, cache policy and downstream source fetching |
| SerpApi | Query and engine parameters return structured results | Google-specific and location-aware research | Plan throughput, cached response and source retrieval |
| Parallel | Search keyword queries and optional objective return results/excerpts; Extract accepts URLs | Agent-oriented discovery with a separate extraction step | Search mode, excerpt scope, URL charges and freshness |
Forma Nôrden previously built a sponsored SearchApi project; that relationship does not establish how it will perform on your queries. The evaluation below specifies the job each product should be asked to perform.
1. Tavily
Tavily is a strong first evaluation when your existing workflow needs search plus selected-page extraction. Its Search API provides a depth setting and relevant results; Extract accepts URLs for the separate page-content step. A team can retain the distinction between discovering the official company announcement and reading the announcement itself.
Basic Search consumes one credit per request and Advanced consumes two. Basic Extract consumes one credit per five successful URL extractions; Advanced uses two per five. Its published pay-as-you-go rate is $0.008 per credit, while monthly bundles have their own allowances. Check current batching, rounding and automatic depth choices: a more expensive search mode may be worthwhile for difficult questions, but it should not silently become the cost assumed for every routine identity lookup.
Fit: a compact agent pipeline that needs control over search depth and selective content. Non-fit: assuming a result is the same company or has the same discovery coverage as Exa just because both return a relevant snippet.
2. Brave Search API
Brave Search API is a good choice when the job resembles conventional web research with defined country and language settings. Its Search product returns results and offers LLM Context for relevant extracted material; Answers is a separate synthesis product. Use the interface whose output the application requires.
The published Search rate is $5 per 1,000 requests with a monthly free credit, while Answers has separate request and token pricing. A list of snippets does not necessarily contain the passage needed to support a detailed account claim. Budget a page-fetching path for those cases, and inspect how the selected context endpoint handles dates and source links.
Brave's Answers pricing documentation distinguishes an answer request from the searches it invokes: the listed formula charges $4 per 1,000 searches plus $5 per million input and output tokens. For a hypothetical response using two searches and 1,534 total tokens, the gross cost is $0.01567, before any free credit. Record the returned usage metadata rather than multiplying the number of answer requests by a single search charge. This applies to Answers, not the $5-per-1,000 Search example in the comparison table.
Fit: teams with an existing content-retrieval layer and a need for market-aware web search. Non-fit: treating Brave's results as a reproduction of Google's order, or treating a recently retrieved snippet as proof that an event happened recently.
3. SearchApi
SearchApi is useful when a specific result environment is part of the task. Its documented interfaces include Google Search, Google Jobs and Google News. A job-board result with a stated employer or a news result from a chosen market can feed a structured discovery queue, while the original source still needs checking.
Our prior Buying Window Signal Scorer uses SearchApi Jobs and News results to identify documents worth reviewing; a score is a research priority, not evidence that a company is ready to buy. Capture the engine, query, location, language, cache setting, response and original source link. Do not translate a missing structured field into a negative business claim.
The pricing page lists a Developer plan at $40 per month for 10,000 searches on its displayed standard-speed option. It describes pay-per-success billing and says that up to 20% of a plan's credits can be used each hour. A 10,000-credit workload therefore needs five hourly allocations at that stated ceiling; the actual elapsed time depends on how the provider defines its windows and other bottlenecks. Confirm the selected engine, speed, throughput and exact success/cache rules before purchase. Also price whichever separate page-fetching method your application uses. Its result contract should not be assumed to equal Exa's search-with-content contract.
Fit: a workflow that needs a named search surface and structured result fields. Non-fit: choosing it solely for a semantic company index without testing the query and downstream identity-matching work.
4. SerpApi
SerpApi offers structured results for named search engines and supports location, country and language parameters on its Google endpoint. That is useful when a sales team needs to inspect a specific search environment rather than discover any relevant page across the web.
Its pricing page lists a $150 monthly Production plan with 15,000 searches and 3,000 per-hour throughput in the displayed best-effort option. Its documentation describes cached and failed requests differently from successful charged searches; verify the policy and cache controls for the selected endpoint. A result with a title and snippet may still need a separate fetch before a consequential claim can be accepted.
Fit: engine-specific or local-market research with a need for structured fields. Non-fit: buying a monthly capacity plan without accounting for unused searches, throughput, cache and page retrieval. Faster service levels and other surfaces can have different terms.
5. Parallel
Parallel offers Search for source discovery and Extract for retrieving content from known URLs. Search returns result excerpts and allows mode selection; a team can use those excerpts to decide which pages merit a separate fetch. Its research-agent products are different purchases from those web-tool calls.
The public price documentation lists Fast or Turbo Search at $1 per 1,000 requests and Basic or Advanced at $5 per 1,000, with ten results included by default. Extract is listed at $1 per 1,000 URLs. Extra results, research runs and other products have their own charges. Specify the mode and result count in your application so that a change in settings does not silently invalidate a budget.
For a new integration, check Parallel's current API version and request schema as well as the price. Its Search reference requires at least one keyword query for v1; the natural-language objective is optional. A migration that only maps Exa's question to the objective would need an additional query-generation step. The same reference says omitted mode defaults to Advanced and the EU endpoint supports only Basic. Set the mode explicitly, check regional availability and test the request against your own examples before switching traffic.
Fit: an agent-oriented search and extraction workflow that can preserve source URLs and evaluate excerpts. Non-fit: equating a fast response with the same semantic company coverage or freshness guarantees as Exa. Evaluate the actual query set and source policy.
How do you test 20 account questions?
Use a fixed, dated fixture before moving traffic. The examples below are synthetic questions, not completed tests or factual claims about real companies. Replace Aster, Birch and the example domains with your own authorised test accounts and sources. For each question, log the search query, provider and mode, request time, returned URL, source date, observed event date and whether the content was cached, freshly fetched or of unknown freshness.
| Questions | Two cases to test | Acceptance rule |
|---|---|---|
| 1–2 | Aster's official website; Birch's official website when its name is also a common word | Match the legal or trading identity and final domain; reject a namesake. |
| 3–4 | Aster's parent company; whether a similarly named subsidiary is the intended account | Keep parent and subsidiary separate unless source evidence establishes the relationship. |
| 5–6 | Whether Aster offers warehouse software; whether Birch offers a product discontinued last year | Quote a current supporting page; do not turn an old product page into a current offer. |
| 7–8 | Aster's current UK office; whether Birch has an office in a city with an old directory listing | Require an attributable source and note any unverified or stale location. |
| 9–10 | Aster's open operations role; whether Birch's previously posted role is still open | Confirm employer, source and status; an indexed closed listing is not an open role. |
| 11–12 | Aster's expansion announced in the past 90 days; an older expansion story updated this week | Use the actual event date. A recent update is not a new expansion. |
| 13–14 | Aster's new sales leader; Birch's replacement for a leader who departed | Verify the person's role and dates; do not infer a replacement from an unrelated profile. |
| 15–16 | Companies providing a defined service in a country; a niche supplier with an uncommon description | Record discovery candidates and evidence separately; a result set is not a complete market census. |
| 17–18 | Aster's current partnership; a claim repeated across several news syndications | Trace to the primary announcement where possible; duplicated copies are not independent corroboration. |
| 19–20 | Aster's pricing page as of a specified date; a page blocked during refresh | Preserve the stated price and date only if supported; report the blocked page as unavailable. |
Score identity, claim support, date/status, and traceability independently. A result passes only if the source refers to the right entity, the passage supports the requested claim, the date is adequate for the question and a reviewer can find the evidence. For an unavailable or uncertain case, a correctly abstained answer is better than a confident invention. Report each category separately, including the count of questions with no acceptable source.
Preserve three clocks: the source's event or publication date, when the provider obtained its page copy if disclosed, and when your application received the response. A fresh request can return older cached content; an old announcement can be accurately fetched today. If cache age is unknown, label it unknown rather than silently calling the evidence live. A job or product page may also change after capture, so set a freshness window appropriate to the decision.
What do 10,000 searches and 20,000 page fetches cost?
Assume 10,000 searches, at most ten results per search, and 20,000 separately requested successful page extractions. For Tavily, assume the URLs are grouped so that each five-successful-extraction credit is fully used; smaller batches or different rounding could change the bill. This deliberately prices a workload that needs a second fetch even if a search response already included usable excerpts. It is a model, not a provider trial. Costs below exclude taxes, free allowances, retries, synthesis-model tokens, storage and labour; check current rates and contract terms before buying.
| Route | Search basis | Separate 20,000-page basis | Modeled retrieval total |
|---|---|---|---|
| Exa Auto + Contents | 10 × $7 = $70 | 20 × $1 = $20 for one content type | $90 |
| Tavily Basic + Basic Extract, gross PAYG | 10,000 credits × $0.008 = $80 | 4,000 credits × $0.008 = $32 | $112 |
| Tavily Advanced + Basic Extract, gross PAYG | 20,000 credits × $0.008 = $160 | $32 | $192 |
| Brave Search + Parallel Extract | 10 × $5 = $50 | 20 × $1 = $20 | $70 |
| SerpApi Production + Parallel Extract | $150 monthly, covering 10,000 within 15,000 searches | $20 | $170 |
| Parallel Fast + Extract | 10 × $1 = $10 | 20 × $1 = $20 | $30 |
| Parallel Advanced + Extract | 10 × $5 = $50 | $20 | $70 |
| SearchApi Developer + Parallel Extract | $40 monthly for 10,000 searches on the displayed option | $20 on the stated Parallel basis | $60 |
The table does not establish equal quality, access or freshness. Exa can include text and highlights with eligible Search results: if those passages satisfy the task, an extra Contents request is unnecessary and its $70 Search basis is the more relevant starting point. Conversely, two distinct requested content types or additional results can raise the bill. Tavily Search can also return content; reserve separate Extract calls for pages that actually need them. At 14,000 credits, Tavily's published 15,000-credit Bootstrap plan is $100/month, a lower bundle base than the $112 gross PAYG example if the plan and selected features fit.
The 20,000 fetches are separate operations, not 20,000 unique pages guaranteed by the 10,000 searches. In a real workflow, deduplicate repeated URLs where permitted, record the retrieval time and choose a refresh interval appropriate to the claim. A cached announcement may suffice for a historical funding event while a current pricing claim needs a fresh source. Include any cache or storage charges and source-access restrictions in that decision.
At a $0.008-per-credit PAYG basis, 1,000 additional Basic searches add $8; 1,000 Advanced searches add $16. Five thousand additional successful Basic extracted pages add 1,000 credits, or $8, assuming the documented grouping and no other fees. Failed extraction and search billing differ by provider; use actual logs rather than assuming all failures are free or all retries are charged.
Now include research quality. Suppose $90 of retrieval and $30 of model use produce 800 accepted claims from 1,000 accounts. The direct provider cost is $0.15 per accepted claim. If 200 exceptions take three minutes each, at an assumed $60/hour, review adds $600 and total cost becomes $0.90 per accepted claim. These yields and labour rates are assumptions, not a provider comparison. A vendor's low request price matters less if it causes enough unsupported or wrongly matched claims to increase review time.
How should you migrate the research contract?
First, create a provider-neutral result envelope: account ID, resolved company domain, query, provider and mode, source URL, source type, retrieval time, source or event date, passage, cache/freshness status and acceptance decision. Preserve the original response where the provider's and source's terms permit. Where content retention is limited, retain the allowed evidence and a link to its origin rather than assuming an API licence grants unlimited republication.
Map Exa-specific features explicitly. A category or filter may have no direct analogue; mark that as a changed requirement and compare its impact on the candidate set. Map unknown publication dates to unknown, rather than a default timestamp. Search ranking, URLs and passages are not guaranteed to remain identical. Keep the existing provider as a fallback during a scoped, read-only evaluation, with a fixed request and cost cap.
Run the 20 questions against each candidate and review difficult cases blindly where practical. Then run a limited production shadow on representative accounts without updating CRM records. Record differences in accepted evidence, missing accounts, duplicated sources, retrieval time and cost. Select a provider only when the new contract meets your minimum acceptable quality, and monitor it after the change. The SERP data quality guide explains why engine, location, device and collection time must remain visible when result position is part of the job.
For a hiring or news signal, use the buying-signals workflow to turn the source into a research question rather than a declaration of purchase intent. The GTM engineering playbook can hold the input, ownership and exception-handling decisions that connect the research pipeline to the sales team.
FAQ
Is an Exa alternative a direct replacement for company search?
Not automatically. A provider may retrieve strong results from the open web without using the same company index, ranking or filters. Compare accepted companies and difficult exclusions on a common query set before changing a discovery workflow. Retain the old result set for the trial, and record missed companies and incorrectly matched subsidiaries separately; a relevant page about the wrong entity is not a successful replacement.
Does a recent search result prove a recent event?
No. The search result, page update and event can have different dates. Read the source, record the event date where available and say when it cannot be established. For a role, also verify whether the listing is still open before treating it as a current hiring signal. A newly fetched copy of a closed vacancy is still a closed vacancy.
Should you fetch every search result again?
Only when the existing response lacks the necessary evidence or your freshness policy requires a new fetch. Search responses with useful source passages can avoid duplicate retrieval, while a claim about a changing job or price may need the current page. Define the acceptable age by claim type, retain the provider's cache information when available and mark it unknown when it is not.
What happens when an original page is blocked?
Record the retrieval failure and retain any permitted search-result evidence with its limits. Do not treat a snippet as a complete page or generate the missing detail. Retry only within a defined budget and escalate when the evidence is essential. If a current source is required for an outreach claim, leave the claim unverified until an authorised source supports it.





