--- title: What DeepSeek’s search API reveals about authority in AI search. description: A first look inside DeepSeek’s newly opened search engine, using its own authority and rerank scores across 2,636 results from 53 real queries. url: https://tanujsharma.xyz/research/deepseek-search-api-authority author: Tanuj Sharma status: Ongoing dataset: 2,636 DeepSeek search results across 53 queries + 30 fetched pages confidence: Directional published: 2026-10-19 modified: 2026-10-19 --- # What DeepSeek’s search API reveals about authority in AI search. A first look inside DeepSeek’s newly opened search engine, using its own authority and rerank scores across 2,636 results from 53 real queries. ## 01 Research question When an AI search engine decides which sources to retrieve, how much does relevance matter, and how much does authority? Until now, that question could only be answered from the outside. You asked an assistant something, looked at what it cited, and guessed at the reasons. On 17 October 2026, DeepSeek opened its in-house search engine to developers through two endpoints: Search and Fetch. [1][2] Search does not just return links. Every result comes with the scores DeepSeek uses to judge it. [1] > **FACT:** Each result includes an authority tier from 0 to 4, where 3 or above means highly authoritative, and a rerank score from 0 to 3 that combines relevance, authority and freshness, where 1.5 or above counts as high quality. > Source: DeepSeek API documentation [1] That makes it possible, for the first time, to look at what an AI search engine retrieves alongside how it scored each source. ## 02 Working finding In this dataset, authority behaves less like one ranking signal among many and more like a gate. Tier 3 and 4 sources took 83% of all first-page positions. Fifty of the 53 number-one results came from tier 3 or 4. ``` Authority tier of every top-10 result 530 results · 53 queries Tier 4 █████ 10.0% Tier 3 ████████████████████████████████████ 73.4% Tier 2 ███████ 13.2% Tier 1 ██ 3.4% Tier 0 0.0% ``` The sources that cleared that gate were rarely company websites describing their own products. They were developer platforms, app stores, documentation hubs, news sites and communities. > **HYPOTHESIS:** In AI search, trust is earned mostly off-site. A page can be relevant and readable and still rarely reach the first page if the engine does not rate its source as authoritative. ## 03 Methodology I ran real search demand through DeepSeek’s public API and analysed everything it returned. ### 1. Queries - 55 queries taken from Google Search Console data in the CAD and 3D printing space, not invented - 2 were site-specific checks (a brand query and a `site:` query) and are excluded from the analysis - The remaining 53 cover converter searches, file viewers, “how to” questions, CAD model searches and a few German, Spanish and French queries ### 2. Search API - 50 results requested per query, the maximum the API allows [1] - 2,636 results returned across 853 unique domains - Fields stored for every result: rank, URL, domain, title, authority, rerank score and publish date ### 3. Fetch API - 30 pages from one website fetched to see what DeepSeek’s fetcher actually receives - Stored: HTTP status, page title, content length and a content preview ### 4. Cost - 85 API calls in total - About $0.06 at $0.75 per 1,000 successful requests [1] ``` Search Console queries (55) │ ├── 2 site-specific checks ──► excluded │ ▼ DeepSeek Search API (53 queries × up to 50 results) │ ▼ 2,636 results · 853 domains │ ├── authority tier ├── rerank score ├── publish date └── rank position DeepSeek Fetch API (30 pages) │ ▼ status · title · content length ``` ## 04 Results These are observations from one snapshot of one engine, in one niche. ### 1. Authority dominates the first page > **OBSERVATION:** 83% of top-10 results came from tier 3 or 4 sources. Tier 1 made up 3.4%. No top-10 result was tier 0. Of the 53 number-one positions, 50 were tier 3 or 4. Average rerank score also rises with every authority tier. Because rerank is the score that orders results, authority is built directly into the ranking, not applied after it. ``` Average rerank score by authority tier (0–3 scale) All 2,636 results Tier 1 ████████████████ 0.96 Tier 2 ██████████████████ 1.10 Tier 3 ████████████████████████ 1.47 Tier 4 ████████████████████████████ 1.70 ▲ 1.5 = DeepSeek’s “high quality” line ``` ### 2. Platforms beat company websites > **OBSERVATION:** GitHub was the most frequent domain on the first page. Developer platforms (GitHub, PyPI, npm, Stack Overflow, dev.to) took 19% of top-10 positions. App stores and marketplaces took another 9%. For a simple query like “stl to step”, the top results were open-source packages and apps, not the many websites offering that exact conversion online. ### 3. Guides beat tools for questions > **OBSERVATION:** For “how to” queries, documentation and step-by-step articles from established software and manufacturing companies filled the top positions. Pages that perform the task directly rarely appeared. ### 4. Language is loosely matched > **OBSERVATION:** For the English query “how to open step file”, 6 of the top 10 results were in other languages. ``` "how to open step file" (English query) · top 10 #1 English #2 English #3 Italian #4 Japanese #5 Korean #6 Spanish #7 English #8 German #9 German #10 French ``` Translated versions of the same documentation were treated as separate results. News coverage behaved the same way: for “stl to step”, one 3D printing news article about a free converter ranked 8th in English, 10th in Chinese and 12th in Russian. ### 5. Top results almost always carry a date > **OBSERVATION:** 79% of top-10 results had a publish date DeepSeek could detect. Freshness is one of the three inputs to the rerank score. [1] A page without a machine-readable date may be giving up part of that score. ### 6. The fetcher reads what the server sends > **OBSERVATION:** 29 of 30 fetched pages came back as clean Markdown with full content. The one exception crashed in the browser with a client-side JavaScript error, and DeepSeek received only the error message. **In this dataset, relevance gets a page considered. Authority decides whether it reaches the first page.** ## 05 Interpretation The useful distinction here is between being relevant and being trusted. Most on-page SEO work improves relevance: clearer titles, better answers, cleaner structure. That still matters. It is what gets a page into the candidate set. But DeepSeek’s first page is dominated by sources whose authority comes from outside the page itself: platforms where others publish and maintain things, stores with reviews, documentation from established vendors, and independent coverage. That suggests a different priority for AI search: - Earn presence on platforms the engine already trusts - Answer questions with genuine guides, not just tool or product pages - Make publish and update dates machine-readable - Make sure the server sends complete, working content - Get listed and covered by directories, communities and the press > **INTERPRETATION:** You can’t optimise your way to trust on your own website. In AI search, you mostly earn it in the places the engine already trusts. That remains an interpretation until the same pattern holds across more niches, more engines and more time. ## 06 Limitations This study covers one engine, one snapshot and one niche: CAD and 3D printing. DeepSeek does not document how authority tiers are calculated. The tiers are DeepSeek’s judgement, not a universal measure of authority. DeepSeek states that its engine filters SEO spam, GEO manipulation and low-quality AI-generated content. [2] I have not tested that claim yet. DeepSeek is also a small share of AI search usage. Comscore data put it at around 2% of US AI prompt volume in June 2026. [3] What holds for DeepSeek may not hold for ChatGPT, Gemini or Perplexity. The Fetch results come from 30 pages on a single website and show how the fetcher behaves, not how it treats the web at large. Queries and fetched URLs are processed by DeepSeek’s hosted service, and its privacy policy says data is stored in the People’s Republic of China. [4] Only public keywords and public URLs were used. ## 07 What I am testing next The next phase is to repeat the same test monthly with the same queries, and widen it. The primary questions are: - Do authority tiers and rankings stay stable from month to month? - Does adding machine-readable dates change a page’s rerank score? - Does new presence on trusted platforms move a site’s authority tier? - Do DeepSeek’s authority tiers agree with third-party authority metrics? - Does the same gate appear in other niches? - Do DeepSeek’s spam filters hold up under direct testing? I will update this page when the second run is complete. ## Sources 1. DeepSeek API Docs: Search & Fetch. https://api-docs.deepseek.com/guides/search 2. AI TL;DR: DeepSeek Search and Fetch API release. https://ai-tldr.dev/releases/deepseek-search-fetch-api/ 3. PPC Land: DeepSeek prices web search for AI agents at $0.75 per 1,000 calls. https://ppc.land/deepseek-prices-web-search-for-ai-agents-at-0-75-per-1-000-calls/ 4. Startup Fortune: DeepSeek opens its Search and Fetch APIs at $0.75 per 1,000 calls. https://startupfortune.com/deepseek-opens-its-search-and-fetch-apis-at-075-per-1000-calls/ ## Changelog - 19 Oct 2026: First snapshot, methodology and working finding published. ## Related - Research: [How AI referral traffic behaves differently from traditional search.](https://tanujsharma.xyz/research/ai-referral-traffic-vs-search) - Essay: [SEO Is Not Dead. Search Has Become Bigger Than Google.](https://tanujsharma.xyz/writing/seo-is-not-dead) ## Research into practice I publish what I learn here. At MarketingPal, I use what holds up to run search and marketing for lean teams. https://marketingpal.co