When an AI agent researches a purchase, it usually doesn’t open Google. It calls a search API built for agents, and the two best known are Tavily and Exa. Tavily is widely described as LangChain’s default search tool. Exa runs its own index of more than 100 billion pages. If your brand isn’t in their results, agents built on them never see you.
We analysed 1,685 Tavily and Exa result lists from RankBits scans, checked our own server logs for their crawlers, and tested whether our latest posts show up. In short:
- Tavily mirrors Google. When a brand ranked in Google’s top 3 for a question, its site appeared in Tavily’s top 8 results 89% of the time. When it wasn’t in Google’s top 100, the figure was 3%. About 40% of Tavily results are the exact URLs from Google’s top 10.
- Exa has its own view of the web. Only 18% of its results match Google’s top 10. It crawled each of our new posts within 2–5 hours, and six days later ranked three of them #1, while Google showed none of them in its top 100.
- The two rarely agree with each other: their results share about 13% of domains.
What Tavily and Exa are
| Tavily | Exa | |
|---|---|---|
| What it is | Search API for AI agents; returns ranked results and an optional AI answer | Search API with its own "AI-native" web index; returns ranked results and highlights |
| Where results come from | Not documented. Tavily "aggregates up to 20 websites per API call" and ranks them by relevance | Its own crawler, ExaSearchBot; an index of "more than 100 billion webpages", refreshed continuously |
| Crawler you can see | None identified in our logs over 14 days | ExaSearchBot/1.0 (follows robots.txt) and ExaUser (user-triggered fetches) |
| Country targeting | country boost | userLocation |
| Recent news | Acquired by Nebius in February 2026 (reported at about $275M) | Published its own crawler and index details in October 2026 |
What the data shows
Method and limits
- 1,173 Tavily and 512 Exa result lists from RankBits scans, 30 brands, June–October 2026, for prompts that don't name the brand. Aggregated and anonymised.
- Tavily ran with
search_depth: advanced, Exa withtype: auto; both returned 8 results. "Ranked" means the brand's own site was among those 8. - Google (top 100) and Bing (top 50) results for the same prompt text came from the same scans.
- Crawler data comes from our own rankbits.com server logs, 26 October – 9 November 2026: one site, so treat it as a case study.
- The brands are self-selected, mostly software companies. That likely inflates GitHub's share of Exa results.
1. Ranking on Google predicts ranking on both
- Tavily is close to all-or-nothing. Outside Google’s top 100, only 3.4% of brands made Tavily’s results.
- Exa is more forgiving. Brands at Google positions 11–100 made its results 40% of the time, twice Tavily’s rate, but that group is small (20 questions).
- When a brand does appear, it’s usually near the top. About three quarters of appearances were in the top 3 results on both engines.
2. Tavily overlaps with Google; Exa mostly doesn't
We didn’t see a Tavily crawler in our logs at all, while Exa’s crawler visited about 700 times in two weeks. Combined with the overlap, the most likely reading is that Tavily draws heavily on existing search indexes, while Exa ranks from its own crawl. Tavily doesn’t document its sources, so treat that as an inference.
3. They return different kinds of pages
| Share of results | Tavily | Exa |
|---|---|---|
| Homepages | 3.4% | 16.6% |
| Blog posts and articles | 29.2% | 14.3% |
| Comparison and alternatives pages | 5.8% | 5.6% |
| YouTube | 6.4% | 0.2% |
| 5.7% | 0% | |
| GitHub | 1.0% | 10.5% |
| Brands' own sites | 2.4% | 4.7% |
| Review sites (G2, Capterra…) | 0.4% | 0.3% |
- Exa likes homepages and repositories. Ask it for “tools that do X” and it often returns the companies themselves.
- Tavily likes articles, YouTube and Reddit, much like Google’s own results.
- Review sites barely register on either.
4. Exa finds new pages within hours; Tavily waits for Google
We published five posts on 3 November and searched for each by its title on 9 November:
| Post | Exa crawled it after publishing | Exa, 9 Nov | Tavily, 9 Nov | Google top 100, 9 Nov |
|---|---|---|---|---|
| RankBits vs ZipTie | 3 h | #1 | Not in top 8 (older RankBits pages at #1 and #3) | No |
| RankBits vs Ahrefs Brand Radar | 2 h | #1 | Not in top 8 (older pages at #1 and #2) | No |
| Which AI engine is hardest… | 4 h | #1 | Not in top 8 | No |
| How to get recommended by ChatGPT | 4.5 h | Not in top 8 | Not in top 8 | No |
| Best AI visibility tools for agencies | 5 h | Not in top 8 | Not in top 8 | No |
For new content, Exa can be the fastest route into AI agents’ results. Tavily showed our older pages, the ones Google already ranks, instead of the new ones. ExaSearchBot also re-crawled our blog roughly twice a day.
5. Agent search visibility goes with AI answer visibility
How often each AI engine named the brand, depending on whether the brand’s site was in Tavily’s or Exa’s results for the same question:
| AI engine names the brand… | In Tavily results | Not in Tavily | In Exa results | Not in Exa |
|---|---|---|---|---|
| ChatGPT | 24.9% | 8.8% | 50.0% | 12.6% |
| Claude | 44.8% | 13.3% | 69.9% | 24.2% |
| Perplexity | 55.2% | 12.4% | 61.5% | 18.6% |
| Google AI Overviews | 55.1% | 17.4% | 60.6% | 23.8% |
These are correlations, not causes: brands with a strong web presence do well everywhere. But they show that agent search and AI answers reward the same thing.
How to rank on Tavily and Exa
- Let ExaSearchBot in. Its robots.txt token is
ExaSearchBot. Check that your CDN or firewall doesn’t block it; its requests are signed with Web Bot Auth, which Cloudflare and Akamai can verify. Look forExaSearchBotin your logs to confirm it’s crawling. - For Tavily, rank on Google and Bing for the question. It’s the strongest lever in our data: 89% inclusion from Google’s top 3, against 3% outside the top 100. The two engines rank very different pages; see Google vs Bing for AI Visibility.
- Make your homepage and product pages say plainly what you do. Exa returns homepages for 17% of results, five times Tavily’s rate. A clear “X is a [category] for [audience]” helps it match you to “tools that…” queries.
- Publish articles that answer the question. Nearly a third of Tavily’s results are articles.
- Put a visible published or updated date on pages. Both engines estimate publish dates (Exa says it parses them from your HTML). Agents that filter by date with Tavily’s
filter_by_published_datedrop pages with no detectable date. - Be where each engine looks: GitHub and READMEs for developer tools on Exa, and YouTube and Reddit on Tavily.
- Serve country pages if you sell locally. Tavily boosts results by
country; Exa localises withuserLocation. - Test it. Query both APIs with your buyers’ questions, or track them in RankBits (Tavily and Exa are included from the Starter plan).
See whether Tavily, Exa and 15 other AI engines and search sources surface your brand. Start with the free AI visibility checker or see plans.
Sources
- Tavily: Search API reference (search depth, results “ranked by relevancy”, country, cached copies, published-date filters); What is the Tavily Search API? (“aggregates up to 20 websites per API call”).
- Tavily acquisition: Morgan Lewis, Nebius acquisition of Tavily (February 2026); deal value as reported by SiliconANGLE.
- LangChain default: described as the default search tool in several third-party guides, for example KodeKloud; LangChain also offers an Exa integration.
- Exa: Exa Search Crawler (ExaSearchBot, ExaUser, robots.txt token, noindex, Web Bot Auth); What is web indexing? (own crawler, more than 100 billion pages, continuous refresh; 6 October 2026); Search API reference (search types, userLocation, published-date estimates).
- RankBits data: 1,685 Tavily and Exa result lists from scans (aggregated and anonymised); rankbits.com server logs 26 October – 9 November 2026; title searches on 9 November 2026.
- Related: Which AI Engine Is Hardest to Get Recommended By?
