DeepSeek's Search API Scores Sites 0 to 4: What SEOs Can Test

DeepSeek opened its in-house search engine to developers on 17 October 2026 through a Search API and a Fetch API, and every search result carries an authority tier from 0 to 4 and a relevance score. For SEOs, it is the first time one of the large AI labs exposes how it grades a site, although early tests suggest the grading is rough.

Key Takeaways

• DeepSeek's update log dated 17 October 2026 says its in-house search engine now powers web and API search. Each API costs $0.75 per 1,000 successful requests, per the Search guide.
• Results include rerank_score (0 to 3, with 1.5 or higher treated as high quality) and authority (0 to 4, or -1 if unknown). The docs do not explain how either is computed.
• DeepSeek's changelog says the engine filters SEO spam, "GEO manipulation" and low-quality AI content. It does not define GEO manipulation, according to Search Engine Journal.
• One independent tester, Suganthan Mohanadasan, ran 110 calls and found spam PDFs on government and university hosts at authority 4 while his own site scored 1. That is a small, preliminary test.
• The same tester found the Fetch API sends a Firefox user agent with no DeepSeek identifier and returned content from three URLs that robots.txt disallows. The three-URL sample is too small to generalize.
• No source I opened names the crawler behind the index, so you cannot block or allow it by name today.
• Sources disagree on whether authority is page-level or host-level. The tester's evidence points to host-level.

What DeepSeek released

The 17 October update note also upgraded the V4.1 Flash model to a version called DeepSeek-V4.1-Flash-1017. Crypto Briefing reports vendor-stated coding results of 77.0% against 74.2% for the previous version on DeepSWE v1.1, and says independent checks were not yet reported. The search launch matters more to marketers than the model update.

Some secondary articles list a 5 requests-per-second limit instead. I use the figures in DeepSeek's own guide. Per the guide, the Fetch endpoint takes one URL, and each Search result also includes a short snippet, a longer model-readable summary and a date that can be null.

What the scores mean, and what nobody can say

Search Engine Journal notes that neither the guide nor the API reference explains how a page gets its authority tier or rerank score, and that no documented method exists to diagnose why a page did not appear. It describes the tier as page-level. Mohanadasan's testing disagrees: in 16 site: probes across 12 hosts, every URL on one host shared the same authority value. Examples from his post: his own domain scored 1, medium.com and reddit.com 2, ahrefs.com, moz.com and linkedin.com 3, and zenodo.org 4. He also found that subdomains of one site could differ.

His sample of 1,000 results from 20 SEO and GEO queries put 60% at tier 3, 32% at tier 2, 6% at tier 4 and 2% at tier 1, with a median rerank score of 1.63. That describes the sample he chose, not DeepSeek's whole index.

The spam filter claim

DeepSeek says the engine strips SEO spam and "GEO manipulation." His examples cut against that: "Free Robux Generator" PDFs on an inspector general site and a CBD gummies upload on a business school site, both at authority 4, and casino content at authority 3. He himself says the results "don't tell us how much spam DeepSeek removed," which is the right reading: a filter can work on average and still miss examples. My view is that the claim is unproven rather than false.

Who supplies the results?

Search Engine Journal reports that a Chinese company, Bocha, says in a GitHub README that it is DeepSeek's official web search supplier, and that DeepSeek's launch pages do not mention Bocha. DeepSeek's privacy policy says it shares keywords with third-party search APIs, but does not clarify whether that covers these queries. I treat the supplier relationship as one company's claim.

Crawling and robots.txt

Mohanadasan's write-up reports that Fetch requests used a Linux Firefox 149 user agent with a zh-CN language header and came from five IPs in a Huawei Cloud network in Hong Kong. The fetcher did not run JavaScript or load images. It returned full content from three URLs disallowed for all agents, including a Google search results page, and he could not confirm whether it ever requested robots.txt. Search Engine Journal adds that DeepSeek's EU training-data summary says its crawler respects robots.txt but does not name the crawler, so rules aimed at a "DeepSeekBot" may not govern search index inclusion.

In plain terms: if you need to keep something out of this system, robots.txt is not a dependable control based on current evidence. This is a single tester's observation. I have not run these tests myself.

What this means for marketers and SEOs

1. Check your own authority values. If you have a developer or a small budget, run site:yourdomain.com queries and record authority and rerank_score. At $0.75 per 1,000 calls, Mohanadasan's 110 test calls cost roughly $0.08.
2. Do not chase a 0 to 4 number. The score is opaque and, per the one test, may be assigned per host. If your site scores 1, the fix is not a content tweak. It may mean third-party mentions on hosts DeepSeek already rates higher, which matches the off-site approach in Off-Page AI Search Visibility Audit and 90-Day Plan.
3. Put clear dates on pages. His test found date filters removed every undated result, and about 11% of sampled results had no date. Visible publish and update dates with machine-readable markup cost little. Whether DeepSeek recognizes your markup is something to verify, not assume.
4. Do not rely on JavaScript for key content. The fetcher did not render it, consistent with what I found for other bots in Do AI Crawlers Render JavaScript?.
5. Watch your logs, carefully. The Firefox 149, zh-CN and Huawei Cloud pattern may help you spot this fetcher, but verify matches first. Blocking an entire network could affect other tenants, as the tester warns. Compare with what I explain about bot identity in OAI-SearchBot vs GPTBot.
6. Add DeepSeek to your prompt panel, not your roadmap. Track a few queries in the app with the setup in Prompt Tracking: A 50-Prompt Panel, and read DeepSeek Citations: What We Genuinely Don't Know for the limits of citation claims. For a broader comparison, see Multi-Model AEO.
7. Avoid manipulation tactics. Whether or not the filter works, building pages purely to game an AI index is the pattern DeepSeek says it targets, and it is the pattern I would not bet a brand on.

What we do not know yet

• How authority and rerank are calculated, and whether authority is page-level or host-level.
• Which crawler builds the index, and whether it honors robots.txt.
• How much of DeepSeek's consumer app answer quality depends on this index. The update note says the engine powers web search, but I have no data on citation share.
• Whether results from Bocha or other suppliers feed the index.
• How representative the 110-call test is. It was one person, one day, one location (Dubai).

As of 13 November 2026, DeepSeek's update log shows no November entries, so nothing newer has changed this picture.

FAQ

What is the DeepSeek Search API?

It is a developer API, opened on 17 October 2026, that exposes DeepSeek's own search engine. Each result includes a title, URL, snippet, summary, date, a 0 to 3 rerank_score and a 0 to 4 authority tier.

How much does the DeepSeek Search API cost?

$0.75 per 1,000 successful requests for each of the Search and Fetch APIs, according to DeepSeek's guide. Empty results and failures are not billed. A free window for some tool-based calls ended on 1 November 2026.

What does the DeepSeek authority score mean?

DeepSeek says 3 or higher indicates highly authoritative sources, and -1 means unknown. It has not explained how tiers are set, and one independent test found values shared across all URLs on a host.

Does DeepSeek's Fetch API respect robots.txt?

Unclear. One tester got content from three disallowed URLs, but could not tell whether robots.txt was requested. DeepSeek's EU summary says its crawler respects robots.txt without naming it.

Does DeepSeek filter SEO spam and GEO manipulation?

DeepSeek's changelog says so, but it does not define GEO manipulation or describe detection. A small independent test found spam on high-authority hosts, so the claim is unproven.

Can I rank in DeepSeek search by improving my authority score?

No one knows how to move it. Build third-party mentions, clear dates and crawlable content, and measure results on a fixed set of queries rather than aiming at a number.

Is DeepSeek traffic worth tracking?

Add it to your referral and prompt tracking, but I would not shift budget on this evidence. I found no data on DeepSeek referral volume, so start by seeing whether it appears at all in your analytics.

Want help with this?

I'm Younus Fardeen, an organic growth, SEO and AEO strategist with 4+ years of marketing experience across edtech and startup teams. If you want help auditing how your site appears across AI search engines such as DeepSeek, see my work and reach me through the contact form at younusfardeen.com.