Skip to content
CiteScore

Guide

How to get cited by DeepSeek

What DeepSeek’s Search and Fetch guide publishes: an in-house search API, a Markdown fetch, authority and snippet fields, and no robots.txt token. A search result is not a chat citation.

20 October 2026 · CiteScore Team · 11 min read

DeepSeek publishes a Search and Fetch guide for two APIs that return live web content. Search finds pages. Fetch reads one URL and returns Markdown. The guide does not publish a citation rate, a required schema type, a word count, or a robots.txt token. CiteScore does not log into DeepSeek, chat.deepseek.com, or the DeepSeek API. The claims below are from that guide and the Search API reference, read on 20 October 2026.

The guide says the Search API is powered by DeepSeek’s in-house search engine and is for agents and RAG applications. It does not say the consumer chat product uses that engine, those authority tiers, or those fetch errors. Until DeepSeek publishes that link, the chat product is undescribed on this point. The same kind of split — a published control, and the files the operator does not require — is written up for ChatGPT, Perplexity, Claude, Gemini, Microsoft Copilot, Grok, Meta AI, Mistral, Brave Search, and You.com, and DuckDuckGo, and Perplexity Comet. The short map is how to get cited by AI answer engines.

A search result is not a citation

Search returns a results array. An empty array is a normal response when nothing matched, and that response is not billed. Each item is a page the engine chose to return for that query. The guide does not say the model then quotes it.

FieldWhat the guide says it is
title, urlThe title and the URL of the result.
snippetA short summary related to the query, typically 100 to 200 characters, intended for display in a UI.
summaryA longer summary related to the query, typically 600 to 800 characters, intended to be read by a model.
contentThe page body. Empty when the caller did not ask for it, or when the content is unavailable.
dateThe publication time in UTC, or null when unknown. The guide does not say which HTML field that time is read from.
authorityAn integer from 0 to 4. -1 means unknown. 3 or above means a highly authoritative source. The guide does not say how a page receives the tier.
rerank_scoreA score from 0 to 3 that combines signals such as relevance, authority, and freshness. 1.5 or above is called high quality. The formula is not published.

site_name can be null. The guide does not say how the name is chosen. A caller sets count from 1 to 50. A larger count does not change the price, and it does not put your URL into a query that did not match it.

The caller decides which sites are in the query

The query is forwarded with its search syntax. A caller can restrict results with site:, exclude a host with -site:, require a phrase, drop a phrase, limit language with lang:, and limit dates with after: and before:. Those operators are on the request. A publisher cannot add a host to someone else’s query, and cannot remove it from an exclusion. robots.txt is not the mechanism the guide documents for that.

ori_question is optional. The guide says providing the user’s original question can improve query understanding. That is a field on the API call, not a tag on your page. localization is an optional IP address used to localize results. It is also the caller’s, and this page does not guess what a given IP changes.

Fetch reads one page, and it can fail

Fetch takes one http(s) URL and returns a title and Markdown. The guide says fetching is subject to the target site’s availability and access policies. HTTP 404 on that API includes an unavailable URL, a URL blocked by site policy, an error from the target site, an unsupported content type, or a page with no extractable Markdown. HTTP 403 means DeepSeek’s security policy rejected the fetch. The guide does not publish the rules of that policy.

A fact that exists only in an image, or only after JavaScript runs, is a fact that can miss a Markdown extract. The guide does not say Fetch executes JavaScript. Text in the HTML the server sends is the part a Markdown conversion can keep. A login wall is an access policy. A page the public cannot open is a page Fetch is allowed to fail on.

There is no published robots.txt token

OpenAI, Anthropic, Perplexity, Google, and Meta each publish a token and say what disallowing it does. DeepSeek’s Search and Fetch guide does not. It does not name a user-agent for Search or for Fetch. Public crawler directories sometimes name DeepSeekBot. That string is not in the guide or the API reference. This page does not treat a directory entry as a robots.txt token.

A group in robots.txt matters for a client that reads the file and identifies itself with that token. DeepSeek has not said its Search or Fetch clients do either. Pasting an unpublished name into robots.txt does not recreate the control OAI-SearchBot or PerplexityBot actually is. If your logs show a client calling itself DeepSeekBot, that string can be spoofed. Confirm it the way you would confirm any other user agent before you write a rule for it.

Training data is a separate question, and it is also unpublished. Issue 53 on the DeepSeek-LLM repository, opened in January 2025 and still open, quotes a line about self-collected data respecting robots.txt and asks for the crawler name. The thread does not give one. A sentence that says a crawl respects robots.txt, without a token, is not a control you can set. If DeepSeek later names a token and says what disallowing it does, that page becomes the control. The tokens operators have published are on the robots.txt guide and the AI crawler user-agent reference.

What you can actually change on the page

Put the name of the business, what it does, and the sentence that answers the question in the title and the first paragraph of the HTML. The documented snippet is typically 100 to 200 characters, and the summary a model is meant to read is typically 600 to 800. Those lengths describe the API fields. They are not a promise about how a chat answer quotes you. A fact that starts after a long preamble is easier for a short extract to miss. DeepSeek does not publish a word count, and this page does not invent one.

Use one name for the site in the title, the first paragraph, and the footer. site_name can come back null, and the guide does not say it reads your logo or your schema. A visible publication date gives a parser a date to find. The result’s date is null when unknown. The guide does not say that date is read from datePublished. Write the date where a person can see it. Schema can repeat it. It cannot supply a date the page hides.

Relevance, authority, and freshness are the signals the guide names inside rerank_score. It does not say which sentence raises relevance, which link raises authority, or which date format raises freshness. A page that answers the query in plain text is the part you control. A page that stuffs the query is not a method this guide recommends, and DeepSeek has not published a spam rule on these two pages for this page to quote.

Nothing on the Search and Fetch guide asks for a schema type or an llms.txt file. Schema that matches the visible page is for parsers that read JSON-LD. An event page can use the Event schema generator. A page that shows reviews of a product or an app uses the Review schema generator. A service page uses the Service schema generator. A course page uses the Course schema generator. A job page uses the Job posting schema generator. A recipe page uses the Recipe schema generator. A forum or support page where people can submit answers uses the QAPage schema generator. A page that names a short passage for text-to-speech uses the Speakable schema generator. An image with a visible credit, caption, or license uses the ImageObject schema generator. A book page uses the Book schema generator. A show page or an episode page uses the Podcast schema generator. A film page uses the Movie schema generator. A dataset page uses the Dataset schema generator. A vehicle page uses the Vehicle schema generator. A hotel page uses the Hotel schema generator. A restaurant page uses the Restaurant schema generator. A clinic page uses the Medical clinic schema generator. A law office page uses the Attorney schema generator. A dental office page uses the Dentist schema generator. An agent office page uses the Real estate agent schema generator. A plumbing business page uses the Plumber schema generator. An electrical contractor page uses the Electrician schema generator. The homepage can use the WebSite schema generator. None of those scripts is a DeepSeek switch. A mention on someone else’s site is a different URL. The guide does not say that a directory listing, a Wikipedia article, or a review causes your page to be returned.

Free AI Answer Snapshot, then the Fix Pack

The free AI Answer Snapshot asks logged-out ChatGPT and Perplexity who they name for a local business and sends one page. AI answers vary run to run. No payment. The Fix Pack, $49 is the done-for-you work after you have seen who was named.

The free check reports GPTBot, PerplexityBot, and Google-Extended. It does not look for a DeepSeek token, because DeepSeek has not published one for it to read. A clear row on that check is not a DeepSeek citation. The GEO audit scores one public URL from the HTML we fetch and returns about ten fixes. It does not call the DeepSeek API or open the chat product.

Run the free check on the host. It reads the homepage, robots.txt for GPTBot, PerplexityBot, and Google-Extended, /llms.txt, a sitemap, and JSON-LD on the homepage. It does not log into an answer engine. Paid files are listed at checkout.

Questions about DeepSeek citations

Does DeepSeek publish a robots.txt token for citations?

Not on the Search and Fetch guide read for this page on 20 October 2026. That guide describes a Search API and a Fetch API. It does not name a user-agent, an IP range, or a robots.txt group. Third-party crawler lists sometimes use the string DeepSeekBot. This page does not treat that string as a control DeepSeek has documented. A January 2025 issue on the DeepSeek-LLM repository asks for the crawler name behind a line about self-collected data respecting robots.txt. The thread is still open and does not supply a name.

Does a hit in the Search API mean DeepSeek’s chat product will cite the page?

No. The Search API returns a ranked list for a query: a title, a URL, a short snippet, a longer summary, an optional page body, a date that can be null, and an authority tier. The guide says that API is for agents and RAG applications, powered by DeepSeek’s in-house search engine. It does not say the consumer chat product uses the same index, the same tier, or the same ranking, and it does not say that a returned URL is quoted in an answer. A caller can also limit or exclude a site. Being returned is not a citation.

Does Event schema, or any JSON-LD, make DeepSeek cite a page?

No. The Search and Fetch guide does not list a schema type, an llms.txt file, or a word count. JSON-LD repeats facts a person can already read. It is not a switch on that guide. The free check reads GPTBot, PerplexityBot, and Google-Extended. It does not read a DeepSeek token, and it does not log into DeepSeek.

Can a site raise the authority tier from the documentation?

The guide does not say how. authority is an integer from 0 to 4, and -1 means unknown. Values of 3 or above are described as highly authoritative sources. rerank_score runs from 0 to 3 and combines signals such as relevance, authority, and freshness, with 1.5 or above called high quality. The guide and the API reference do not publish the inputs to either number. Treat the fields as labels on a result, not as a checklist you can complete.

Sources

Request a free AI Answer Snapshot

Same request as the free snapshot page. AI answers vary run to run. This is a snapshot, not a ranking, and not a promise that any assistant will name the business.

All fields are required.

We email the snapshot to this address. We do not ask for passwords, cookies, or content behind a login. We do not sell emails. Privacy.

Email this request instead
citescore@mail.grokbot.com

No mail app opened? Copy this and email it to citescore@mail.grokbot.com from any inbox.

If that send fails, email this request instead, or use the copy buttons.

Sends the business name, city, service, website, and email to citescore@mail.grokbot.com through FlowForm. This page path is included. No account and no paid checkout.

Prepared by the CiteScore Team.