Tavily vs Exa: sourced company research still needs publication dates and page versions
Try Tavily first for control over news searches and passage extraction; try Exa for company categories, path filtering, and content returned within search. The key is to assess publication dates, content freshness, and missing pages separately, then budget for deliverable research.
For a sales or investment research assistant, a search API should deliver company facts that can be checked, rather than merely a description with links. Tavily and Exa both search the web and retrieve page content. If your team orchestrates news searches, page reads, and field verification itself, try Tavily Search/Extract first. If company-category search, path-level source filtering, and returning text or relevant passages directly with search results matter more, try Exa first. Both require separate checks for when an article was published and when this copy of its content was retrieved. Interface capabilities alone cannot establish which service produces more accurate facts.
Ask separately about company pages and recent news
Suppose you need a company's products, current leader, and latest funding. Its homepage can establish the products, its leadership page can identify the leader, and a funding announcement needs an event date and amount. Restricting sources to official domains first and searching news separately helps avoid treating a leader mentioned in an old funding announcement as the current one. This is a selection scenario, not a completed performance test.
According to the Tavily Search documentation, you can restrict domains and set a date range, but results without a detectable publication date are not removed by default. Enabling filter_by_published_date also excludes those results. A strict date rule useful for news can exclude an undated company homepage that is regularly updated. Company background information and recent events should therefore not share the same filters.
Exa Search supports domain and path-prefix filtering and offers a company category. That category does not support startPublishedDate or endPublishedDate; unsupported combinations return 400. It suits finding company pages first, then using general web or news searches for recent changes. If your product only needs one fixed query, this split adds calls and orchestration. Category search should not be treated as an already verified, continuously updated company database.
Once you have the text, check whether it is the version you need
Tavily Extract accepts up to 20 URLs per request and can use query to rank passages, so it does not merely hand the whole page to a model. In this mode, raw_content may contain excerpts only; the field name does not prove that you have the full text. Even with HTTP 200, check both results and failed_results.
Exa Contents accepts 100 URLs per request, lets you select text, highlights, or summary, and returns per-URL states in statuses. When you already have search results, request content within Search to avoid reading them again. Generated summaries can organize information, but to verify qualifications on a funding amount, retain the supporting passage and URL.
Caching is the easiest detail to misread. Tavily Search's max_age_hours: 0 requests a live fetch; a positive value limits the age of the cached copy. Pages that cannot satisfy that condition are omitted, leaving fewer results. Live fetching on non-paying plans, including the free plan, is best-effort and may return 434 under heavy load. For a task that must cover every target company, missing pages should enter a follow-up queue, not be taken to mean the company has no updates.
Exa's freshness guide says that with a positive maxAgeHours, a failed fetch may return the latest cached copy. A value of 0 returns content retrieved within the last 60 seconds; repeated requests for the same page within those 60 seconds may return the same copy. The old livecrawl parameter is deprecated. Its Search documentation also says lite reads the index without fetching pages at query time and does not support freshness options requiring live retrieval. To check a price or leader updated today, use a mode supporting that control and inspect the returned content. Do not treat low-cost Lite as an equivalent deliverable.
Company research should therefore keep, for each field, the fact, supporting passage, source URL, page date, request time, and fetch or cache status. Leave unknown page dates blank. The request time only tells you when you called the API, not when the website changed. A successful response from either service does not replace this judgment. If the version supporting an important field cannot be confirmed, mark it for verification rather than letting the language model merge two versions into one confident answer.
Budget for usable research, not a single request
The Tavily pricing lists Basic Search at 1 credit per call and Advanced at 2. Basic Extract costs 1 credit per 5 successfully extracted URLs; failed extractions are not charged. Pay-as-you-go costs US$0.008 per credit, with 1,000 free credits each month. Assuming no free-credit deduction, 100 companies with 2 Basic Searches each and 5 successfully extracted URLs each total 300 credits, or US$2.40. This example excludes generation models, retries, and human verification.
The Exa pricing checked the same day lists auto at US$7 per 1,000 calls. The base price includes up to 10 results, with additional results charged separately. Standalone Contents costs US$1 per 1,000 pages per content type. With the same 200 Auto Searches, followed by text only for 500 URLs, the cost is 1.40 + 0.50 = US$1.90, before free credits and extras such as summaries. If search already returns the content you need, do not automatically count standalone extraction again. The free balance resets to US$10 each month, and the first team can earn a one-time US$10 bonus by completing onboarding. This cannot be compared directly with Tavily's credit count, and neither example establishes equal result quality.
The budget must also track failures, follow-up searches, and reruns. Tavily explicitly waives charges for failed URL extractions; that does not establish that every failed Search is free. The Exa pricing checked here is also insufficient to claim that all failure cases are uncharged. Reconcile actual usage records and confirm how missing pages and timeouts are billed during a trial, rather than estimating production capacity from free allowances.
When to delegate research and when to retain control
If you want the service to repeat searches, read pages, and organize a report itself, compare Tavily Research with Exa Agent. The former supports an output schema and citation format; the latter can asynchronously return structured results and field-level grounding. This is a different deliverable from a single Search/Extract call. Do not compare one provider's complete research task with the other's search unit price. With a fixed company list, fields, and evidence requirements, controlling search and retrieval yourself makes item-by-item acceptance easier. Research agents are more worth trying when the question is open-ended and needs several rounds of exploration.
Procurement concerns about Tavily should also refer to current terms. Its official statement dated November 2, 2026 says the Nebius acquisition did not change existing agreements, API keys, pricing, or data processing. The acquisition alone is not grounds to assume migration is necessary, nor is this statement a guarantee that future terms will never change.
The choice can be concrete: prioritize control over general search and passage extraction, and start with Tavily; prioritize company categories, source-path filtering, and content returned within search, and start with Exa. Trial both on the same companies, fields, and freshness requirement. Record the completion rate for evidence-backed fields, misuse of old versions, the cost of following up failures, and human review time. Include undated homepages and recently updated pages in particular. Lower API costs matter only once those results meet your requirements.
Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.
Original source:docs.tavily.com