DeepSeek

1. What DeepSeek is

DeepSeek is a conversational AI product with an optional web search mode. When search is enabled, the hosted product can retrieve current web material, combine it with a model-generated answer, and display source information. This makes the product a retrieval-augmented generative engine within generative engine optimization, rather than a conventional search-results page with ranked links.

The name also refers to a model family, an API, and open weights. These surfaces share technology and branding, but they do not share a proven source index or ranking system. DeepSeek’s V4 announcement made V4 Pro and V4 Flash available through the official web product, app, API, and open-weight distribution. It did not say that self-hosted or third-party deployments inherit the official product’s search system.

Surface	Who controls retrieval	What a citation test measures
Official DeepSeek web and app	DeepSeek controls the hosted search and answer experience	Visibility in the consumer product at a recorded date, mode, language, and region
Official DeepSeek API	The request selects an available tool and model; DeepSeek runs documented server tools	Behavior of that API configuration, not consumer ranking parity
Open-weight or self-hosted model	The deployer supplies the index, search provider, fetcher, prompts, and interface	The deployer’s retrieval system
Third-party product using DeepSeek	The third party combines a DeepSeek model with its own data and tools	The third party’s product behavior

This distinction is especially important when comparing DeepSeek with other ecosystems that combine models and products, such as Tongyi and Doubao. A model benchmark can describe reasoning or tool use, but it cannot reveal which web pages the official consumer product retrieves or credits.

2. How DeepSeek web search works

DeepSeek launched Internet Search on its website on December 10, 2024. The launch notice said the product could extract several keywords and search them in parallel. The app followed in January 2025, adding web search, DeepThink, file upload, and synchronized chat history.

That public description supports a broad answer loop: interpret the question, create search queries, retrieve candidate pages, select evidence, synthesize an answer, and present sources. DeepSeek’s current privacy policy says it uses third-party APIs to provide search. It also says DeepSeek shares users’ input keywords with those providers. The policy does not name the provider or disclose the ranking model, reranking features, or rules that decide which retrieved page receives visible credit.

Stage	Public evidence	What remains unknown
Search activation	The consumer product exposes a web search mode	Whether every enabled turn invokes retrieval
Query generation	The 2024 launch described several extracted keywords and parallel searches	Current V4 query count, rewriting rules, and language routing
Candidate retrieval	DeepSeek says third-party APIs provide search; independent tests observe current web information and sources	Provider identity, corpus boundaries, and candidate ranking
Evidence selection	Answers use some retrieved material and omit other candidates	Reranking weights and source-quality thresholds
Attribution	Consumer tests record source links or source labels	The rule that converts retrieved evidence into a visible citation

Search mode, model mode, and conversation state should be recorded separately. A May 2026 consumer-interface study enabled both Deep Think and Web Search for DeepSeek V4 Flash, but that setup does not prove that every combination behaves the same. File uploads and prior turns can also place material in the context without open-web discovery.

Language is another independent variable. Chinese and English prompts may generate different queries, retrieve different domains, and use different entity names. A multilingual GEO test should keep the intent constant while recording query language, source language, location, and the exact product surface. Comparisons with Baidu AI Search, Quark AI Search, and Kimi require the same controls.

3. Crawlers and user-agents

DeepSeek has not publicly named a crawler or user-agent for consumer search in the official materials reviewed on August 29, 2026. The Tow Center’s March 2025 audit also reported that DeepSeek’s crawler was not publicly known. Stanford’s December 2025 transparency assessment reached the same conclusion after examining DeepSeek’s disclosures.

The absence of a disclosed identity does not mean that access never occurs. Search systems can discover a page through background indexing, a user-triggered fetch, a third-party search provider, a syndicated copy, or another public page that quotes the original. Each path creates a different access-control problem.

Access role	Public DeepSeek identifier	Publisher control that can be verified today
Model-training collection	None documented	Apply a general data-use policy and review logs; do not guess a DeepSeek-specific UA
Search indexing	None documented	Maintain ordinary search accessibility and inspect verified requests
User-triggered page fetch	None documented	Test the public product with controlled pages and compare server logs
Third-party search or syndication	No provider disclosed	Audit copies, canonical links, and the source that the answer actually credits

For an access audit, capture the complete user-agent string, source IP, requested URL, response status, rendered content, timestamp, and request pattern. Verify IP ownership or published ranges when available. A string containing DeepSeek is not enough, because user-agent text is easy to spoof. The general AI crawler audit method still applies even when the vendor has not supplied a named bot.

An invented robots.txt directive can create false confidence while leaving the real retrieval path untouched. Baidu AI Search should be evaluated separately because a documented search crawler or existing index relationship on one Chinese platform does not establish DeepSeek’s path.

4. Citation preferences

DeepSeek does not publish a consumer source-ranking formula. Its official product announcements confirm web search, but they do not state universal preferences for first-party sites, recent pages, Chinese domains, scholarly sources, or a particular content format. Those hypotheses need query-level evidence.

Visible outcomes should be classified before any preference is inferred:

Outcome	Observable evidence	What it means
Citation or source link	A page or domain is linked as support for an answer	The interface gave that source visible credit
Named source without a link	A publisher, author, or brand is named but no page is linked	A mention occurred; the underlying page may be unclear
Uncredited answer	No source is shown	The answer may rely on model memory, retrieval, supplied context, or a mixture
Misattribution	The answer links or names the wrong origin for supplied material	Visible credit is inaccurate even if the prose is plausible

The difference between citation, mention, and link matters because a brand can appear without its page being retrieved, and a retrieved page can influence an answer without receiving a visible link. Search presence, source selection, and attribution should therefore be measured as separate events.

The clearest DeepSeek-specific accuracy evidence comes from a failure audit. In 2025, the Tow Center supplied DeepSeek Search with excerpts from 200 news articles and asked for the original headline, publisher, date, and URL. DeepSeek misattributed the source in 115 cases. The result shows that visible source credit can be wrong. It does not identify a ranking factor, and it predates the current V4 product.

A broader 2026 Chinese-language generative search study observed an average of 9.2 citations per DeepSeek answer on both the web and app surfaces. Among 580 paired queries, the web and app results had 40.8% exact-URL overlap and 50.7% domain overlap. This is evidence of surface-level variation, not a disclosed preference or a stable platform-wide rate. The study is a preprint based on a June and July 2026 snapshot, and its public dataset should be used with those boundaries intact.

Practical experiments can test source type, freshness, language, and passage structure, but they should not turn correlation into a platform rule. Direct, self-contained evidence remains easier to inspect and quote after retrieval, which is the editorial logic behind writing for AI citation. It is not a disclosed DeepSeek ranking weight.

5. API and integration

DeepSeek’s API now has two different tool paths. Ordinary function calling lets the model request a developer-defined function; the developer then executes it and returns the result. The current Responses API also supports a server-side web_search tool executed by DeepSeek.

A minimal tool declaration looks like this:

{
  "tools": [{ "type": "web_search" }],
  "tool_choice": { "type": "web_search" }
}

The documented response can include web_search_call output items. Their actions can describe searching, opening a page, or finding text within a page. The current reference does not define a separate citation or source-URL object contract, so a collector should not assume every search action yields a machine-readable citation. The API is stateless, so a multi-turn test must send the relevant history again. Model support and fields can change; the request, returned model, tool choice, search actions, answer text, and any displayed or returned URLs should be retained together.

Surface	Search capability	Measurement use	Main limitation
Consumer web or app	User-facing Web Search	Measures the actual hosted experience	Automation and result extraction may be limited
Responses API	DeepSeek-hosted web_search server tool	Repeated prompts can record whether search was called and what the answer returned	No documented consumer-ranking equivalence or dedicated citation-object contract
Chat Completions with function calling	Developer supplies the search function and results	Tests a controlled retrieval pipeline	Measures the developer’s search stack
Open-weight deployment	Deployer chooses every retrieval component	Supports reproducible custom experiments	Does not reproduce official DeepSeek Search

The API became materially closer to a search-measurement surface in 2026, but it remains a proxy for the consumer product. AI citation tracking should store consumer and API observations in separate series. A model name, location, tool configuration, or account context can change the candidate set before citation selection begins.

6. History and timeline

DeepSeek’s GEO-relevant history is a sequence of product and tool changes rather than a simple model-release list.

Date	Change	Effect on web visibility
July 25, 2024	The API added function calling	Developers could connect DeepSeek models to an external crawler or search function, but the tool remained developer-supplied
December 10, 2024	Internet Search launched on the official website	The consumer product gained live retrieval; the announcement explicitly said the API did not yet support search
January 15, 2025	The official mobile app launched with web search and DeepThink	Search became available across the hosted app and web experience
September 22, 2025	DeepSeek-V3.1-Terminus reported that Search Agent performance had improved	Search-agent behavior changed with a model update, although the announcement did not disclose ranking details
April 24, 2026	V4 Preview reached the app, web, API, and open-weight channels	Surface and model labels needed to be recorded separately as the same model family spread across different retrieval systems
August 13, 2026	V4 Pro GA reached app, web, and API; native Responses API support launched	The official API gained a documented server-side web-search path suitable for separately labeled tests

The API change log is the best current record for model and interface changes. Search experiments should preserve the collection date because a result gathered before V4 Pro GA does not describe the same product snapshot as one gathered after it.

7. Measured citation behavior

Public evidence now includes one large Chinese-language observational dataset, but each study answers a different question. No current public benchmark establishes a universal DeepSeek citation rate, source preference, or ranking formula across topics, languages, dates, and surfaces.

Evidence	Surface and method	Result that can be reused	Boundary
Tow Center, March 2025	DeepSeek Search received one excerpt from each of 200 news articles and had to identify the original source	DeepSeek misattributed 115 of the 200 excerpts	Source-identification stress test, not ordinary user queries; predates V4
Stanford FMTI, December 2025	Disclosure audit of DeepSeek documentation and related materials	No crawler name or crawler opt-out protocol was disclosed	Transparency result, not a retrieval or citation-performance test
Zheng et al., August 2026	Official consumer interface, the model displayed as DeepSeek V4 Flash, Deep Think and Web Search enabled; 52 mpox questions, one run each in Wenzhou during May 2026	Confirms a recent, explicitly labeled search-enabled consumer setup and reports response-quality measures	One health topic, one region, no repeated runs, and no general citation-rate metric
Chinese-language generative search study, July 2026	614 Chinese queries across the web and app interfaces of four platforms, three replications; 214,119 raw records and 160,860 cleaned citation records	DeepSeek averaged 9.2 citations per answer on both surfaces; among 580 paired queries, exact-URL overlap was 40.8% and domain overlap was 50.7%	Preprint and observational June-July 2026 snapshot; no rejected candidate pool or ranking mechanism

The Frontiers study is useful because it labels the surface, mode, location, dates, account type, and query procedure. It also warns that the displayed model name does not independently verify the hosted backend. The larger Chinese-language citation study adds repeated, cross-surface observations and makes its cleaned citation data public. Neither study reveals the rejected candidate pool or a causal ranking mechanism.

A defensible DeepSeek test should record at least five outcomes: whether search ran, which sources were displayed, whether the answer cited the correct origin, whether the target brand was mentioned, and whether the result recurred across runs. Search mode on and off should be stored separately. Chinese and English prompts should also be separate cohorts rather than pooled into one percentage.

The API can support repeated collection, but its results cannot be merged with consumer observations. Likewise, a result from Kimi, Metaso, or Baidu AI Search cannot fill a missing DeepSeek measurement. The same query and metric are necessary, but they are not sufficient unless the surface, date, language, region, and retrieval setting also match.

8. Optimizing for DeepSeek

DeepSeek-specific optimization begins with what can be tested. The company has not published a ranking formula, so every content recommendation needs a stated evidence level and an observable result.

Action	Why it is reasonable	How to verify it
Publish the answer in accessible, rendered HTML	Any live retrieval path needs readable source text, even though the crawler identity is unknown	Compare server responses, logs, and consumer answers for a controlled page
Keep facts, dates, and original evidence together	Search answers can misattribute sources; explicit provenance makes the correct origin easier to verify	Check whether the cited URL is the original page and supports the nearby claim
Write self-contained passages under literal headings	A complete passage is easier to extract and attribute after retrieval	Track whether the passage or its distinctive facts appear in cited answers
Maintain stable Chinese and English entity descriptions	Cross-language queries may retrieve different domains or aliases	Test paired prompts and record source language and entity naming
Keep developer documentation versioned and current	Technical queries can change when models, modes, and API fields change	Test version-specific questions after every relevant product release
Repeat a fixed prompt set	One answer cannot distinguish a stable pattern from generation variance	Store several runs with the same surface, search setting, region, and date window

Do not add a guessed DeepSeek user-agent to robots.txt. Use the AI crawler audit process to verify requests, and preserve normal access paths that your publishing policy permits. A vendor-specific allowlist becomes defensible only after the vendor publishes verifiable identifiers or your own logs establish a controlled pattern.

Multilingual GEO matters more than literal translation. Product names, company names, technical terms, dates, and version numbers should resolve to the same entity across languages. Writing for AI citation can improve passage clarity, while AI citation tracking determines whether the change affected search, citation, mention, or nothing observable.

The test protocol should compare Chinese products without importing assumptions from one to another. Tongyi and Doubao combine models and consumer assistants; Quark AI Search and Metaso present different search contexts. Their crawler disclosures, source pools, and citation interfaces require their own evidence.

9. Why DeepSeek matters for GEO

DeepSeek combines three forms of distribution: a hosted consumer search product, a programmable API with server-side web search, and open weights used in independent deployments. That breadth creates opportunities for visibility, but it also creates a measurement trap. A source cited by a self-hosted DeepSeek model with a third-party search provider has not demonstrated visibility in DeepSeek’s official consumer product.

A cross-platform comparison should use one measurement contract:

Field	Required record
Product surface	Official consumer app, official API, self-hosted deployment, or third-party product
Model and mode	The displayed model name, thinking mode, and search setting
Query context	Exact prompt, language, region, account state, and prior conversation
Retrieval result	Whether search ran and which source pages were displayed or returned
Visibility result	Citation, link, mention, answer use without credit, or no appearance
Reliability	Source accuracy, support for the cited claim, variation across repeated runs, and web-app overlap

Apply that contract to Baidu AI Search, Doubao, Kimi, Tongyi, Quark AI Search, and Metaso before drawing a China-wide conclusion. A common spreadsheet does not make the platforms identical; it makes their differences interpretable.

DeepSeek is consequential for GEO because the same model family can appear behind many products while the surrounding retrieval and interface system assigns credit to web sources. Publishers gain clearer evidence when they label the exact surface, verify source accuracy, and measure citations separately from mentions. Those habits remain useful as DeepSeek changes models, search tools, and product modes.

References

Official DeepSeek sources:

• DeepSeek V2.5: The Grand Finale (Internet Search launch, December 10, 2024)
• Introducing DeepSeek App (January 15, 2025)
• DeepSeek-V3.1-Terminus (September 22, 2025)
• DeepSeek-V4 Preview (April 24, 2026)
• DeepSeek-V4-Pro GA Release (August 13, 2026)
• Responses API reference, Responses API guide, and API change log
• DeepSeek Privacy Policy (updated February 10, 2026)

Independent evidence:

• Tow Center for Digital Journalism, AI Search Has a Citation Problem (March 6, 2025)
• Stanford Center for Research on Foundation Models, DeepSeek Transparency Report (December 2025)
• Zheng et al., Evaluating search-enabled large language model interfaces for mpox public health consultation (August 26, 2026)
• What Do Chinese-Language Generative Search Engines Cite and Surface? A Large-Scale Empirical Study and the accompanying CN-GEO Citation Dataset (July 2026)