Developers can now query DeepSeek’s own web index through a Search API that went live Oct. 17. According to the company’s changelog, the system suppresses spam aimed at search rankings, “GEO manipulation,” and weak AI-written pages. Search Engine Journal covered the launch on Oct. 19, and every capability described below is DeepSeek’s own account.

The changelog says the engine already powers search on DeepSeek’s website and inside its API. It puts the index at hundreds of billions of pages and daily traffic at hundreds of millions of searches worldwide. Those figures come from DeepSeek alone.

GEO, generative engine optimization, is the practice of shaping content so AI answer engines cite it. DeepSeek credits strict detection for its filtering but has not defined GEO manipulation or explained how detection works. Google stated on May 15 that its spam policies extend to generative AI responses in Search, so the idea is not new, but a vague label gives publishers nothing to test against.

Every result carries an authority field scored on a zero-to-four scale, and -1 means the tier is unknown. DeepSeek’s API reference calls tier 3 and above “highly authoritative.” A second field, the rerank score, runs from 0 to 3 and blends relevance, authority, and freshness; 1.5 or higher counts as high quality. Neither the guide nor the reference describes how a page earns either value.

Results also include a short display snippet (usually 100 to 200 characters) and a longer machine-oriented summary, usually 600 to 800 characters. Full page text can be requested too, where the engine has it. Developers can restrict queries with site:, and filter by language and publication date.

Both Search and Fetch are priced at $0.75 per thousand successful calls. Searches with no results are not billed, and asking for 50 results is priced like asking for one. Server-side searches and fetches made through Claude Code, Codex, or DeepSeek Harness on that API are free for now. Billing for those calls starts Oct. 31 at 16

UTC, which is noon ET.

Fetch launched alongside Search and returns a single page as Markdown. The guide says results hinge on whether the target site is reachable and what access rules it sets. DeepSeek’s EU summary of V4 training data says its crawler is designed to honor robots.txt where applicable, but it does not say which crawler it means. None of the launch pages (guide, API references, changelog, announcement) names which crawler feeds the search index or the user agent Fetch uses.

That gap matters for access control. The Known Agents directory carries an entry for DeepSeekBot and labels it an AI training data crawler. A robots.txt rule aimed at that bot may not touch the search index, and the launch material does not settle the question.

Privacy is another open item. DeepSeek’s privacy policy, last revised Feb. 10, says DeepSeek relies on outside APIs for search services and shares users’ keywords with them. Bocha, a Chinese search API company, says in a GitHub README that it is DeepSeek’s official web search supplier. DeepSeek’s launch pages do not mention Bocha, and neither the changelog nor the policy says whether any queries still go to an outside service. Dan Petrovic of DEJAN raised both points in his post on the launch.

For SEO and GEO teams, the practical value is cheap visibility into one index. A site: search on your own domain costs under one cent with a funded account and can surface as many as 50 pages, each with a tier and a score. A missing page could mean it sits outside the index, was caught by a filter, or was dropped for another reason, and the documentation offers no way to tell which.

Treat the tiers as an unexplained diagnostic, not a benchmark. Run site: queries on your priority domains now, log the results, and hold off on robots.txt changes until DeepSeek defines GEO manipulation, explains how tiers are set, and names its index crawler.

Based on reporting by Matt G. Southern in Search Engine Journal, published October 19, 2026.