DeepSeek SEO: how to get cited by the search engine DeepSeek built

01

DeepSeek now searches its own index

On October 17, 2026, DeepSeek’s API changelog put on record what had never been public: “DeepSeek’s web and API search capabilities are powered by our in-house DeepSeek search engine.” It describes “an AI-native general-purpose search engine that indexes hundreds of billions of web pages, supports multilingual search,” handling “hundreds of millions of search requests from around the world every day.” The same day DeepSeek opened that engine to developers as the Search and Fetch APIs.

That replaces what site owners worked from in 2025. In March 2025 Weng Rouying, CTO of the Hangzhou search-API company Bocha, told Daily Economic News that DeepSeek had connected to Bocha’s Search API before its January 2025 surge; DeepSeek never confirmed it. DeepSeek’s privacy policy, last updated February 10, 2026, still says “We integrate third-party APIs to provide search services.” Nothing public dates the switch, so treat DeepSeek citation data from before October 2026 as possibly describing the old setup.

The engine also travels beyond DeepSeek’s chat. Its Claude Code guide says that when the model decides a question needs the web, Claude Code connected to DeepSeek will “perform the search through the API provided by DeepSeek.” The Search & Fetch guide says the same server-side calls happen in Codex and DeepSeek Harness, and become billable on November 1, 2026 (Beijing time).

02

Search on, search off: two answers to one question

DeepSeek’s composer has two toggles, DeepThink and Search, as the deepseek.com homepage shows them (the Chinese interface says 深度思考 and 智能搜索, “smart search”). Web search arrived on chat.deepseek.com on December 10, 2024 — “toggle ‘Internet Search’ for real-time answers,” said the release note. Since October 17, 2026 the web and app answer with DeepSeek-V4.1-Flash-1017 (release note); the app’s September 11 update had already merged its Instant, Expert and Vision modes into one (App Store).

With Search off, the answer can only come from what the model learned in training, and it carries no sources. With Search on, DeepSeek runs a search agent. Its V4 encoding notes document the special tokens of “the internal search agent pipeline” — written like <｜action｜>, shortened to action below — and they read like a map of the steps:

1. 01

   A router picks Search or Answer

   An action step “determines whether the user prompt requires a web search or can be answered directly,” returning a decision such as “Search” or “Answer.” A question the model thinks it knows may be answered from memory even with the toggle on.

   Your lever: Be in the training data too: consistent facts about your brand on public, crawlable pages.

2. 02

   It writes queries and rates how authoritative sources must be

   query generates the search queries. authority “classifies the user prompt’s demand for source authoritativeness,” and domain identifies the topic area. The Search API also takes the user’s original question (ori_question) and an IP address to localize results.

   Your lever: Use the words buyers type — product, category and place names — in titles and headings, in their language.

3. 03

   DeepSeek's index returns scored results

   Each result carries a title, URL, a 100–200 character snippet for display, a 600–800 character summary “intended to be read by a model,” a publication date, an authority tier and a rerank_score.

   Your lever: Front-load a summary worth reading, and give the page a date a machine can parse.

4. 04

   It decides which pages to open

   extracted_url and read_url decide “whether each URL in the user prompt should be fetched and read.” DeepSeek’s Fetch returns a page as Markdown and is “subject to the target site’s availability and access policies.”

   Your lever: Keep public pages free of logins, bot challenges and JavaScript-only content.

5. 05

   The answer cites with numbered markers

   DeepSeek’s published search prompt (R1 repository) wraps each result as [webpage X begin]...[webpage X end] and tells the model to cite with [citation:X] at the end of the sentence, filter irrelevant results, “avoid repeatedly citing the same webpage,” and answer in the user’s language.

   Your lever: Write sentences that carry one fact each, so a marker can sit right after them.

03

The scorecard behind every DeepSeek result

DeepSeek is the only major assistant that publishes the ranking fields its search returns. They’re in the Search & Fetch guide, for an API the changelog says runs on the same engine as DeepSeek’s web search. Users never see these numbers; the model does.

• Official

  Freshness is scored, and dates are filterable

  Freshness is one of the three named inputs to rerank_score, and queries accept after: and before: publication-date operators. A date the engine can read is part of how it ranks you.

• Our read

  An undated page may drop out of date-filtered searches

  When DeepSeek can’t find a publication date, date is null. A page with no date can’t satisfy “published after 2026-01-01,” so it’s likely left out whenever the agent filters by recency.

• Our read

  Authority is a tier, and the query sets the bar

  Pages get a tier from 0 to 4, and the agent separately rates how much authority a question demands. Health, money and legal questions plausibly demand higher tiers than a product comparison.

• Official

  Language is a filter

  The engine “supports multilingual search,” queries accept a lang: operator such as lang:zh,en, and DeepSeek’s search prompt answers in the user’s language.

• Official

  Results can depend on location

  The Search API’s localization field takes an IP address “used to localize search results.” Expect a buyer in Moscow and one in Shanghai to see different sources for the same question.

• Our read

  The model may read a summary, not your page

  Full page content is returned only on request, while the summary is built for the model. If your answer sits below a long preamble, the summary may miss it.

04

Check your pages against DeepSeek's index

Because the engine answers API calls, you can see roughly what DeepSeek’s agent sees. Search and Fetch each cost $0.75 per 1,000 successful requests; failed and empty searches aren’t billed. You need an API key from platform.deepseek.com. Run these five checks in order:

1. 1Are you indexed? Search site:yourdomain.com with a topic word. No results for pages that rank in Google or Bing means DeepSeek’s index hasn’t picked them up.
2. 2What date does it hold? Compare each result’s date with the page’s real publication date. null means it found none.
3. 3What tier are you? Note your authority against competitors’ on the same queries. It’s a relative signal; the scale isn’t explained.
4. 4Who wins your buyers’ questions? Run the questions they ask, in their language, with lang: and with localization set to an IP address in your market.
5. 5What does it extract? Fetch your key URLs and read the Markdown. The prices, specs and dates should be there — not a cookie banner or an empty app shell.

# 1–3. Is the page in DeepSeek's index, with what date and tier?curl -s https://api.deepseek.com/v1/tools/search \  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \  -H "Content-Type: application/json" \  -d '{"query": "free plan site:yourbrand.com", "count": 10}' \  | jq '.results[] | {url, date, authority, rerank_score}' # 4. A buyer question, in Chinese, published this year#    -d '{"query": "免费 发票 软件 推荐 lang:zh after:2026-01-01", "count": 20}' # 5. What does DeepSeek extract from the page?curl -s https://api.deepseek.com/v1/tools/fetch \  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \  -H "Content-Type: application/json" \  -d '{"url": "https://yourbrand.com/pricing"}' \  | jq -r '.content' | head -40

05

DeepseekBot, Common Crawl and the crawler nobody has named

DeepSeek now names its training crawler. Its EU public summary of training content for DeepSeek-V4.1, last updated September 24, 2026, lists it as DeepseekBot, used “to collect publicly available online content for training” and to “verify existing data by cross-referencing multiple sources and fact-checking information.” It says the crawler “is designed to respect robots.txt instructions and other standard web protocols, where applicable” and is “not designed to circumvent captchas, paywalls, or password-protected websites.” The V4 and V3.1/V3.2 summaries say the same.

What DeepSeek doesn’t publish is a user-agent string, IP ranges or a way to verify a request. Bot directories fill the gap with their own observations: Known Agents lists Mozilla/5.0 (compatible; DeepSeekBot/1.0; +https://www.deepseek.com/), measures 79.8% robots.txt compliance, and warns that any client can claim the name. Cloudflare’s AI Crawl Control list has no DeepSeek entry.

Two more paths matter. The same summary lists Common Crawl among V4.1’s training data, and the V4.1 tech report says DeepSeek “re-bootstrapped” its own crawling system from Common Crawl — so CCBot feeds DeepSeek models as well as many others. And nothing at all is named for the search index, or as Fetch’s user agent.

• DeepseekBotYour call

  Model training

  DeepSeek’s own crawler, named in its EU training summaries: collects public content for model training and cross-checks facts.

  Honors robots.txt. “Where applicable,” per DeepSeek; no user-agent string or IPs to verify it.

• CCBotYour call

  Model training

  Common Crawl’s crawler. DeepSeek lists Common Crawl in V4.1’s training data, as do many other model makers.

  Honors robots.txt. Verifiable: Common Crawl publishes IP ranges and reverse DNS.

# Opt out of DeepSeek model training (its own crawler).# Matching is case-insensitive (RFC 9309), so this also covers# the "DeepSeekBot" spelling that bot directories report.User-agent: DeepseekBotDisallow: / # Optional: Common Crawl, which DeepSeek also trains on.User-agent: CCBotDisallow: / # No DeepSeek search crawler exists to allow by name,# so keep public pages open to everyone else.User-agent: *Allow: /

On JavaScript the evidence is one test. In June 2026 Search Engine World pasted unlinked URLs into a set of assistants: DeepSeek executed the page’s script and reported the rendered value, while ChatGPT, Claude and Gemini read the raw HTML. That covers the user-pasted path only, not the index. Server-render anyway, then confirm with a Fetch call.

Fetch also documents how it fails: HTTP 403 when DeepSeek’s own “security policy” rejects a request, 404 for “a URL blocked by site policy” among other causes, and a per-domain rate limit. Our read: a CDN rule that challenges unknown clients turns your pages into 404s for DeepSeek’s agent.

06

What DeepSeek says it filters out

The changelog is blunt: the engine “filters SEO spam, GEO manipulation, and low-quality AIGC through strict detection mechanisms, ensuring authoritative and trustworthy information for users and LLMs.” DeepSeek is the first assistant maker to name GEO manipulation as something it removes. It hasn’t defined the term or described the detection, so the table below is our read of what’s most exposed.

The positive version is visible in DeepSeek’s citations: in the July 2026 study, DeepSeek relied most heavily on official sources — government and corporate sites — for business details. A first-party page that states the facts plainly is both hard to classify as spam and what DeepSeek already leans on.

07

What DeepSeek's answers cite: the 2026 evidence

The largest look at DeepSeek’s citations is an arXiv study (July 2026) that sent 614 controlled Chinese-language queries — dining, hotels, beauty, health, legal and finance among them — to the web and app interfaces of DeepSeek, Doubao, Tencent Yuanbao and Qwen, three times each, in June and July 2026, and analyzed 160,860 citations. Read it as Chinese local search measured before DeepSeek’s October announcement; it may not transfer to English B2B questions.

• Observed

  Many slots per answer

  With about nine citations per answer, a page doesn’t need to be the best source — one of the useful ones is enough.

• Observed

  A long tail of domains

  DeepSeek’s cited pages had the lowest average 5118-Baidu quality score of the four engines. Big domains don’t own the slots.

• Observed

  Official sources for business facts

  For contact details and similar facts, DeepSeek leaned most on government and corporate official sites.

• Observed

  Strong recency on time-sensitive questions

  Doubao and DeepSeek showed the strongest recency response; Tencent Yuanbao cited the oldest pages.

• Observed

  Listed isn't cited

  On DeepSeek and Yuanbao, answers listed about 10.5 sources but put about 6.9 into the text; 39.3% of listed sources never appeared in the answer body.

• Observed

  Brands without a source

  16.9% of brand mentions in DeepSeek Web answers matched no page it cited — the highest rate of the four. What the wider web says about you still speaks for you.

Web and app don’t cite the same pages. DeepSeek was the most consistent of the four engines, and still its web and app answers shared only about half their domains (a Jaccard similarity of 0.51) and 48% of their top three sources. A page cited in the browser may be missing in the app, so test both.

08

Where DeepSeek answers reach buyers

DeepSeek’s reach is large and lopsided. Microsoft’s AI Economy Institute (January 2026) put its market share at 89% in China, 56% in Belarus, 49% in Cuba and 43% in Russia, estimated usage across Africa at 2 to 4 times other regions, and found that “Adoption remained low in North America and Europe.”

Language decides which pages compete

The engine filters by language and DeepSeek answers in the user’s language. Our read: a question asked in Chinese or Russian is matched mostly against pages in that language, so an English-only site competes poorly for it. If those buyers matter, publish real localized pages — a machine-translated stub is exactly what a low-quality AIGC filter is built to catch.

Restrictions to factor in

• Italy: the Garante ordered an immediate limitation on processing Italian users’ data on January 30, 2025.
• US federal agencies: a Homeland Security order published October 20, 2026 requires civilian agencies to remove DeepSeek products, including third-party products built with DeepSeek technology (TechCrunch). It doesn’t cover consumers or companies.
• Government devices elsewhere: Australia, Canada, South Korea and the Czech state administration, among others, restricted it on official systems in 2025 (Wikipedia summary).
• Data location: DeepSeek’s privacy policy says it collects, processes and stores personal data in the People’s Republic of China — the point most regulators cite.

Hosted DeepSeek models search elsewhere

DeepSeek publishes its model weights, so other companies run DeepSeek models inside their own assistants. Our read: those products bring their own retrieval, so a citation there is won in that product’s search, not DeepSeek’s. Test them separately if your market uses them.

09

Finding DeepSeek in your analytics

DeepSeek documents nothing about how its links open or what referrer they send. Google’s channel definitions name it: GA4’s AI Assistant channel covers “sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok.” But Google hasn’t published the referrer list behind the channel, which arrived on May 14, 2026.

There may be little to catch. SE Ranking, using GA4 data from 101,574 sites between January 2025 and April 2026, found DeepSeek held 0.37% of AI referral traffic — the fourth-largest source — and that “from September 2025 onward, DeepSeek’s referral traffic dropped to essentially zero across all regions.” SE Ranking’s explanation is that users were still trying many AI tools in 2025. Our read: app-to-browser clicks often arrive with no referrer and land in Direct, so check there before concluding DeepSeek sends you nothing.

# Session source: DeepSeek, when a referrer arrives(^|\.)deepseek\.com$ # A custom channel for the major assistants, placed above Referralchatgpt\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|deepseek\.com

Server logs are harder. A DeepseekBot match is a training crawler and can be spoofed, and Fetch’s user agent isn’t published. You can learn what DeepSeek’s fetcher looks like on your own stack: create an unlinked page, request it through the Fetch API, and log every hit on that path.

# Ask DeepSeek to fetch a page nobody else knows aboutcurl -s https://api.deepseek.com/v1/tools/fetch \  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \  -H "Content-Type: application/json" \  -d '{"url": "https://yoursite.com/dsk-canary-7f3k2"}' > /dev/null # Then read who came for it: time, IP and user agentgrep "/dsk-canary-7f3k2" access.log | awk -F'"' '{print $1, $6}'

For citations, run a prompt panel the way the July 2026 study did: a fixed set of buyer questions, Search on, three runs each, on web and app. Record whether you’re cited in the text, listed without a citation, or named with no source at all.

10

Six DeepSeek assumptions that are now out of date

1. Assumed

   DeepSeek searches Bing, Google or Bocha.

   Now

   DeepSeek says its web and API search run on its own engine. Bocha’s CTO claimed the 2025 integration; DeepSeek never confirmed it, and its privacy policy still mentions third-party search APIs.

2. Assumed

   DeepSeek documents no crawler.

   Now

   Its EU training summaries name DeepseekBot. What’s still missing is a user-agent string, IP ranges and any crawler for the search index.

3. Assumed

   Blocking DeepseekBot takes you out of DeepSeek's answers.

   Now

   DeepSeek describes it as a training and fact-checking crawler. Nothing says it feeds search; check with a site: query after you change robots.txt.

4. Assumed

   There's no way to see how DeepSeek ranks pages.

   Now

   The Search API returns each result’s authority tier, rerank_score and date. The weighting is unpublished, but you can see where you stand.

5. Assumed

   GA4's AI Assistant channel will show your DeepSeek traffic.

   Now

   Google names DeepSeek in the channel, but SE Ranking measured DeepSeek referrals at essentially zero from September 2025. Check Direct, and track citations in the answers.

6. Assumed

   Anything running a DeepSeek model searches the way DeepSeek does.

   Now

   Products that host DeepSeek’s open weights bring their own retrieval. Only DeepSeek’s web, app and API use DeepSeek’s index.