DeepSeek SEO: the technical guide to getting cited by DeepSeek

How DeepSeek's Smart Search finds and cites pages#

DeepSeek added web search to chat.deepseek.com on December 10, 2024, telling users to toggle Internet Search “for real-time answers” (release note), and shipped it in the mobile apps on January 15, 2025. Today it’s the Smart Search toggle beside Deep thinking in the composer (some builds label them Search and DeepThink). The model behind it has been DeepSeek-V4.1-Flash since September 10, 2026, and the app’s September 11 update unified the Instant, Expert and Vision modes into one.

Which index it searches is not documented. DeepSeek’s privacy policy, last updated February 10, 2026, says only: “We integrate third-party APIs to provide search services, and we will share your input keywords to provide these services.” It names no provider. The one public claim came from outside: in March 2025 Weng Rouying, CTO of the Hangzhou search-API startup Bocha (博查), told Daily Economic News that DeepSeek had integrated Bocha’s Search API before its January 2025 surge. DeepSeek hasn’t confirmed it, and nothing public says whether it still holds in 2026. DeepSeek’s own V3.1 model card adds one clue: its search-agent benchmarks run on “a commercial search API + webpage filter + 128K context window.”

1. 1

   Search is on — but the model still decides

   Search runs only when the user turns on Smart Search, and even then the model chooses whether to call it; users report answers that skipped the search despite the toggle (GitHub, May 2026). Everything else comes from training: V4 was pre-trained on public web data plus licensed datasets.

   Your lever: Be worth remembering: consistent, widely published facts about your brand reach the answers that never search.

2. 2

   An agent searches, reasons and searches again

   V3.1 (August 2025) brought “Stronger multi-step reasoning for complex search tasks,” per its release note. In the app each round shows as a step like Found 18 web pages, with the favicons of the hits. The official prompt passes the date alongside the results: “Today is {cur_date}.”

   Your lever: Use the words your buyers type — product names, category terms, places — in titles and headings.

3. 3

   A third-party search API returns candidates

   The provider is undisclosed. On DeepSeek’s API-side search — the one Claude Code uses — each result carries a URL, a title and a page_age, per DeepSeek’s own harness docs.

   Your lever: Be indexed widely, and show a real publish date an index can read.

4. 4

   The agent opens a handful of pages

   A Read 4 pages step lists the pages it opened as linked titles. Who fetches them is undocumented — no user agent, no IP ranges. In a June 2026 test DeepSeek executed JavaScript on a page a user pasted in.

   Your lever: Keep the page public, fast and free of bot challenges, with the answer in the HTML.

5. 5

   The model filters, combines and cites

   The official template tells it to filter irrelevant results, cite with [citation:X] at the end of the relevant sentence, “synthesize information from multiple relevant webpages,” and reply in the user’s language. The UI renders each marker as a small numbered badge.

   Your lever: Write sentences that carry one fact on their own, so they can be cited mid-paragraph.

Crawlers and access: what DeepSeek documents (almost nothing)#

DeepSeek publishes no crawler documentation: no user agent, no IP ranges, no robots.txt guidance for search or for training. Cloudflare’s AI Crawl Control bot reference lists no DeepSeek bot, and none of the 705 entries in Cloudflare Radar’s bots directory — in a mirror synced on September 21, 2026 — belongs to DeepSeek.

Some bot directories do list a DeepSeekBot token. Known Agents gives the string Mozilla/5.0 (compatible; DeepSeekBot/1.0; +https://www.deepseek.com/bot) and notes that DeepSeek publishes no way to verify it. The deepseek.com/bot page in that string returned a 404 when we checked on September 21, 2026. Treat DeepSeekBot as unconfirmed.

On training, DeepSeek describes collection but offers no opt-out. The 2023 DeepSeek-LLM README says its corpus included “self-collected data respecting robots.txt.” The V4 model card (April 2026) says only “We use publicly available information on the internet” and “technical methods to acquire and filter these freely accessible data,” alongside licensed third-party datasets. Neither names a crawler.

# DeepSeek documents no crawler. This token comes from third-party
# bot directories and is unconfirmed: harmless, but unverifiable.
User-agent: DeepSeekBot
Disallow: /
 
# DeepSeek's search visibility depends on not blocking the unknown:
# keep public pages open to well-behaved agents.
User-agent: *
Allow: /
Disallow: /account/

The only way to learn what DeepSeek’s fetcher looks like on your stack is to catch it. Borrow Search Engine World’s method: publish a page at an unguessable, unlinked URL, paste it into DeepSeek with a request to summarize it, and log the full headers of every hit that follows.

# Every request for the canary page: time, IP and user agent
grep "/dsk-canary-7f3k2" access.log \
  | awk -F'"' '{print $1, $6}'
 
# Then check whether that IP or user agent reads your real pages
grep "203.0.113.42" access.log | awk '{print $7}' | sort | uniq -c | sort -rn

Technical requirements#

None of DeepSeek’s documents say what a page needs. The list below is built from what it does publish — the search template, the API’s result fields — and from independent tests, with the confidence marked on each row.

Public, unchallenged pagesRequired

The agent can only cite what search returns and what it can open. A login, paywall, CAPTCHA or bot challenge in front of the facts means it reads nothing — or someone else’s page.

Facts in server-rendered HTMLHelps

DeepSeek ran JavaScript in one June 2026 test, but that was one fetch on one site, the search provider’s crawler is unknown, and ChatGPT, Claude and Gemini read raw HTML in the same test. Server-render anyway.

Visible, accurate datesHelps

Results carry page_age, the official prompt injects today’s date, and DeepSeek showed one of the strongest recency responses of four engines in the July 2026 study.

Self-contained sentencesHelps

Citations land at the end of individual sentences, and the model must combine several pages. A sentence that states its fact without the paragraph around it is easier to cite.

The user's languageHelps

The template answers in the user’s language. Our read: queries are searched in it too, so an English-only page competes badly for questions asked in Chinese or Russian.

Structured dataUnconfirmed

No statement from DeepSeek. Results expose a URL, a title and a date; keep every fact in visible copy as well.

robots.txt rulesUnconfirmed

No documented token for search or training. A DeepSeekBot group may never match a real request.

Sitemaps and IndexNowUnconfirmed

They help the engines you can name. Whether DeepSeek’s undisclosed provider reads either isn’t known.

llms.txtUnconfirmed

DeepSeek’s docs never mention it, and nothing suggests its search reads it.

Two quick checks cover what you can control: the answer is in the raw HTML, and a plain client isn’t refused. If either fails, an unidentified fetcher is likely to fail too:

# 1. Is the answer in the raw HTML?
curl -s https://yoursite.com/pricing | grep -c "Plans start at"
 
# 2. Does your CDN or WAF refuse a non-browser client?
curl -s -o /dev/null -w "%{http_code}\n" https://yoursite.com/pricing
 
# 0 matches, or a 403, 429 or challenge page = fix before anything else

What DeepSeek cites#

We found no DeepSeek citation study from the big Western SEO platforms. The best evidence is an arXiv study (July 2026) that sent 614 controlled Chinese-language queries — including dining, hotels, beauty and health services — 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 behavior; it may not transfer to English B2B questions.

• Many sources per answer

  Observed

  DeepSeek averaged 9.2 citations per answer on both web and app. With that many slots, a page doesn’t need to be the best source — just one of the useful ones.

• A long tail of domains

  Observed

  Its most-cited domain took only about 5% of citations, and its cited pages had the lowest average 5118-Baidu quality score of the four engines. Domain authority isn’t the gate.

• Official sources

  Observed

  In the study’s analysis of business contact details, DeepSeek relied most heavily on official sources, particularly government and corporate official sites. Your own site is a citable asset.

• Freshness on time-sensitive queries

  Observed

  Pages DeepSeek Web cited for time-sensitive queries were about 55 days old, against 181 days for evergreen ones — among the strongest recency responses measured.

• Spread, don't repeat

  Official

  The official template tells the model to synthesize several pages, “avoid repeatedly citing the same webpage,” and cap list answers at 10 points while pointing users to the sources for the full list.

• Brands without a source

  Observed

  16.9% of brand mentions in DeepSeek Web answers couldn’t be matched to any page it cited — most plausibly they came from training. What the web says about you consistently matters.

The same study found DeepSeek’s web and app cite overlapping but different sources: a domain-level Jaccard similarity of 0.51, the most consistent of the four engines and still only about half. A page cited in the browser may not be cited in the app, so test both.

Who actually sees DeepSeek answers#

DeepSeek’s reach is large but lopsided. Microsoft’s AI Economy Institute (January 2026) put DeepSeek’s market share at 89% in China, 56% in Belarus, 49% in Cuba and 43% in Russia, estimated usage in Africa at 2 to 4× other regions, and found that “Adoption remained low in North America and Europe.” Similarweb’s August 2026 data puts China at 51.43% of deepseek.com’s traffic, Russia at 8.27% and the US at 5.86%.

Where it's restricted

• Italy: the Garante ordered an immediate limitation on processing Italian users’ data on January 30, 2025. MIAI reported the ban still in force in January 2026.
• South Korea: DeepSeek pulled its app from local stores on February 15, 2025, and resumed service on April 28, 2025, after notifying the privacy regulator of its compliance.
• Germany: Berlin’s data protection commissioner asked Apple and Google to delist the app on June 27, 2025. The notices were non-binding, and per MIAI the app stayed available.
• Government devices: the US Navy, NASA, the House of Representatives and several US states, plus government bodies in Australia, Japan, Canada, Belgium, the Netherlands, the Czech Republic and Denmark, restrict it on official systems.
• 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.

Third-party hosts use their own search

DeepSeek’s weights are MIT-licensed, so many products run DeepSeek models on someone else’s retrieval. Tencent’s Yuanbao added DeepSeek-R1 in February 2025 with web search that draws on WeChat official accounts (CLS), and Tencent Cloud’s DeepSeek API connected Sogou’s search API (IT Home). A citation there is won in Tencent’s search, not DeepSeek’s. The reverse holds too: DeepSeek’s own search now travels beyond its app. Its API docs say Claude Code pointed at DeepSeek will “perform the search through the API provided by DeepSeek.”

Measuring DeepSeek#

DeepSeek documents nothing about how its links open or what referrer they send. Google’s channel definitions do name it: GA4’s AI Assistant channel, added in May 2026, covers “sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok.” But Google hasn’t published the referrer list behind it, and there may be little to catch. SE Ranking found DeepSeek held 0.37% of AI referral traffic in early 2025 — the fourth-largest source — and that “from September 2025 onward, DeepSeek’s referral traffic dropped to essentially zero across all regions.”

SE Ranking offers no explanation. Our read: either DeepSeek users rarely click out, or their clicks arrive without a referrer and land in Direct, as app-to-browser clicks often do. Check both before concluding DeepSeek sends you nothing — and if deepseek.com sessions still sit in Referral, give them a custom channel above it:

# Session source — DeepSeek, when a referrer arrives
(^|\.)deepseek\.com$
 
# All major AI assistants, DeepSeek included
chatgpt\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|deepseek\.com

Server logs can’t be filtered by a DeepSeek user agent, because none is published. Use the canary test in the crawlers section to learn what its reads look like on your stack, then watch for that pattern on real pages. Remember that a page can be found without being read: only pages listed under Read were opened.

For prompt tracking, copy the July 2026 study’s design: a fixed set of buyer questions, Smart Search on, three runs each, on both web and app. Record whether your pages appear in the answer’s sources pill (it reads like 10 web pages) and in the inline badges, and whether competitors are named with no source at all.

Myths worth dropping#

The action checklist#

Everything above, in the order we’d do it. Tick items off as you go — your progress is saved in this browser.

DeepSeek SEO: frequently asked questions#