Tavily vs Exa vs Seltz: Choosing an AI Search API
Compare Tavily, Exa and Seltz for AI account research: web search, URL extraction, people discovery, pricing units and a practical evaluation checklist.
On this page
Choose an AI search API by the evidence your application needs and the inputs it already has. Use search to discover sources, extraction to read known URLs, and separate identity checks when researching people. Tavily, Exa, and Seltz offer search and page-content retrieval, with different endpoints, settings, and billing units.
For account research, the useful comparison is whether a request produces evidence you can verify at an acceptable cost. A low request price means little if your agent needs repeated searches or receives the wrong person's profile.
This article compares official documentation and public pricing retrieved on November 4, 2026. It reports no new performance tests. Prices and endpoint options can change, so check the linked documentation before implementation.
Start with the research task
An API is an interface that lets software request data from another service. An endpoint is the particular operation it calls. One vendor can offer several endpoints that solve different problems.
Suppose you need to understand a prospect's expansion plans. Finding its announcements requires search. Reading an announcement at a URL you already possess requires extraction. Finding the executive responsible requires people research and an identity check.
| Task | Tavily | Exa | Seltz |
|---|---|---|---|
| Discover web sources | Search with relevant snippets and optional page content | Search with optional text and highlights | Search returns ranked documents and content |
| Read known URLs | Extract | Contents | Fetch |
| Find companies or people | General queries and domain filters | Company and people categories | Companies and people scopes |
| Generate an answer | Optional generated answer in Search | Answer and research modes | Answer and Agent |
The table describes documented capabilities, not equivalent output quality. See the references for Tavily Search, Exa Search, and Seltz Search. Seltz documents its answer and research operations separately.
Where Tavily, Exa, and Seltz differ
Tavily: search depth and URL extraction
Tavily's Search endpoint offers basic, fast, ultra-fast, and advanced search depths. It can return source snippets, cleaned page content, and an optional generated answer. Its documented topics include general, news, and finance.
For account research, domain filters let you restrict a request to company websites or selected publications. The max_age_hours setting controls how old a cached page copy can be. Requesting fresh content can increase waiting time. Tavily Search reference
When you already know the URLs, Tavily Extract returns their content as Markdown or text. Basic and Advanced extraction use different credit amounts, so record the extraction depth in any comparison.
Exa: search categories and content controls
Exa supports company and people categories alongside general web search. Its documentation describes account research, stakeholder discovery, and company-list research. These are retrieval capabilities, not guarantees that every returned field is correct. Exa companies and people
The distinction between Exa's lite and auto modes affects more than price. Lite serves page content from Exa's cache and does not fetch pages live. Auto supports additional controls, including generated structured output. Exa labels fast and instant as legacy modes. Exa Search guide
For known URLs, Exa Contents can return text, highlights, or summaries. Its maxAgeHours setting lets callers control cache age, including requesting a fresh fetch.
Seltz: explicit scopes and ranking tiers
Seltz separates the data you search from the ranking tier you request. Its scopes include people, companies, news, wikipedia, financial-reports, and web, which the documentation labels beta. Base and Pro select different ranking configurations. Pro is the default. Seltz Search reference
That distinction matters when comparing a broad web query with a people lookup. Record both scope and tier. A successful people search says nothing by itself about coverage of company announcements.
Seltz Fetch reads up to 20 supplied URLs per request. The API reference currently specifies Pro, so the pricing table below uses that tier.
AI search API pricing: compare the billing unit
These are public list prices from the research snapshot. Tavily figures use its pay-as-you-go rate of $0.008 per credit. Subscription discounts, downstream model calls, and application costs are outside this table.
| Operation | Price | Billing unit and condition |
|---|---|---|
| Tavily Basic, Fast, or Ultra-fast Search | $8 | Per 1,000 requests, at one credit each |
| Tavily Advanced Search | $16 | Per 1,000 requests, at two credits each |
| Tavily Basic / Advanced Extract | $1.60 / $3.20 | Per 1,000 successful URL extractions |
| Exa Lite / Auto Search | $4 / $7 | Per 1,000 requests with up to ten results each |
| Exa Contents | $1 | Per 1,000 pages, per content type |
| Seltz Base / Pro Search | $3 / $5 | Per 1,000 searches |
| Seltz Pro Fetch | $10 | Per 1,000 successful URLs |
Sources: Tavily credits, Exa pricing, and Seltz pricing.
The differences in units change the calculation. Exa charges separately for each content type on Contents. Requesting text and highlights for one page therefore counts as two content types. Search includes text and highlights, while generated page summaries cost extra. Exa billing details
Exa also charges for search results above the first ten. Seltz prices Search per search regardless of the returned document count. Tavily charges Extract for successful extractions, with one Basic credit covering five URLs.
Keep research modes out of a basic retrieval price comparison. A request that performs several searches and generates a cited answer does more work than one that returns documents.
Test an account-research workflow
Consider a hypothetical consulting team researching a manufacturer before a meeting. It needs the company's expansion announcement, the announced location, and the name of the executive responsible.
Prepare a small evaluation set with the reference sources already checked by a person. Include companies with similar names, older announcements that remain searchable, and cases where public sources don't name an executive.
Then test each stage separately:
- Discover the source. Ask for the expansion announcement using the same company name and date window across providers.
- Read the page. Give each extraction endpoint the same URL. Check whether it returns the relevant passage and preserves the announcement date.
- Check the person. Confirm the executive's identity and current role from supporting sources. Preserve missing information as unknown.
- Assess the answer. Check whether every reported fact has evidence. Measure the complete workflow's cost and elapsed time.
Use comparable result counts and content limits. Record every endpoint, mode, query, and collection date alongside the response. Otherwise, a later configuration change can make two apparently similar runs incomparable.
Repeat difficult cases. One successful request won't tell you how reliably the application handles ambiguous company names or unavailable pages.
Measure cost per usable answer
Define a usable answer before calculating its cost. For the manufacturer example, it might require the correct company, a dated expansion source, and a supported location. Executive ownership can remain unknown when the evidence doesn't establish it.
The basic calculation is total workflow cost divided by accepted answers. Include unsuccessful searches and retries because they consume budget too.
For example, imagine 100 research tasks cost $2 in API and model charges, and 80 meet your acceptance criteria. The cost per accepted answer is $0.025. Those are illustrative numbers, not measurements of any vendor.
Track review time separately. A cheap response that regularly needs manual identity correction may cost more to use than its API bill suggests.
Retrieved content also becomes input to the language model. Long pages can increase that cost even when search pricing stays constant. Our LLM cost optimization article explains how repeated context and model calls affect complete research tasks.
People discovery still needs identity checks
A people-search result identifies a candidate. It doesn't establish a verified email address, a working phone number, or complete employment history.
Compare identifiers, employer details, and dates against your reference record. Names alone can produce plausible mismatches. If your application needs contact information, evaluate that separately: Exa, for example, prices Agent email and phone enrichment apart from its other usage. Exa pricing
Our LinkedIn data enrichment benchmark examined a specific set of profiles and configurations. It distinguishes retrieving known profile URLs from searching with a name and company. Those results do not establish which provider performs best across general web research.
The same evidence requirement applies when research becomes a list. GridIQ is alkemy's research workspace for that workflow. A useful row needs a source that supports the requested field, rather than merely mentioning the company.
Frequently asked questions
Which AI search API is best for account research?
Choose based on source coverage, identity accuracy, freshness, and cost per accepted answer in your own cases. This documentation comparison supports an evaluation shortlist, but it doesn't establish a performance winner among Tavily, Exa, and Seltz.
Is a web search API the same as a scraping API?
Web search discovers relevant sources from a query. Extraction reads content from URLs you already have. These vendors offer both kinds of operations, so compare the endpoint that matches your input.
Can search results become verified contact records automatically?
Search results can supply evidence for a contact record, but each required field still needs a check. A matching profile doesn't prove that an email works or that two people know each other.
Choose the research task, define the evidence it requires, and compare equivalent requests. To see how researched company context fits into a consultant's workflow, explore the alkemy demo.