DeepSeek opens its search engine to everyone and gives every page an authority score: what changes for a brand
Since October 17, 2026, DeepSeek’s search engine can be bought per call: 5 yuan per 1,000 searches, a declared filter against “GEO manipulation,” and, for every result, an authority score from 0 to 4. It is the first time an AI search engine has put in writing how it judges a page. What the technical documents say, what Zhipu, Kimi, Tencent, and ByteDance do, and how much DeepSeek weighs on Italian pages.
In brief
- On October 17, 2026, DeepSeek opened to everyone the search engine behind its chat: 5 yuan per 1,000 searches, an index of “hundreds of billions of pages,” and a statement written into the technical documentation: it filters “SEO cheating, GEO manipulation, and low-quality AI content.”
- Every result comes with an authority score from 0 to 4 and a quality score from 0 to 3. It is the first time an AI search engine has put in writing the grade it gives a page. OpenAI, by comparison, says only “multiple factors, placement not guaranteed.”
- In China, search is already a service bought per call, from Zhipu, Kimi, Tencent, and ByteDance. In Italy, DeepSeek’s crawler still counts for little: 981 visits in 120 days on the site we run, against 117,907 for OpenAI’s.
For a year and a half, companies have tried to understand how a model chooses its sources, and the search engines have always answered in generic terms. On October 17, 2026, DeepSeek did something different: it opened its search engine to anyone who wants to use it, wrote into the technical documentation which content it filters out, and attached an authority score to every result. We read the changelog and the developer guide, and compared them with what OpenAI and the other Chinese engines declare.
Index
What DeepSeek announced on October 17
DeepSeek’s search engine already existed: it is what the chat uses when it searches the web. Since October 17 it is a service in its own right, which any developer can use through two functions: Search searches the web and returns structured results; Fetch reads a single page and turns it into text a model can read. The price is 5 yuan per 1,000 successful requests, a little under 70 cents, for both functions.
In the changelog DeepSeek describes the engine as “AI-native,” with an index of “hundreds of billions of pages” and “hundreds of millions of searches a day.” Then comes the sentence that matters to anyone with a brand: “Against SEO cheating, GEO manipulation, and low-quality AI-generated content, we apply a strict recognition mechanism to filter them out, so that users and LLMs are presented with authoritative, trustworthy information.” No Western engine, so far, has used the word GEO in a technical document.
The authority score: 0 to 4 for every page
The technical guide is more interesting than the announcement. Every result returned by Search carries two numbers. The first is authority, “the authority of the page, from 0 to 4,” with a declared threshold: from 3 up, the result is “extremely authoritative.” The second is rerank_score, a score from 0 to 3 that “considers relevance, authority, and timeliness together,” with another threshold: from 1.5 up, the result is “high quality.” Each result also comes with its publication date, a short summary for people looking at a screen, and a long one, 600 to 800 characters, written specifically to be read by the model.
For anyone who works on a brand’s presence in AI answers, this changes the job. Until now, an engine’s judgment of a page could only be inferred; here it is a number that anyone, with an API key and 5 yuan, can read for their own pages and for their competitors’. OpenAI states “multiple factors, placement not guaranteed” and says nothing more. Google publishes guidelines hundreds of pages long, but no score. DeepSeek does not explain how it calculates the grade, but it shows it.
What “GEO manipulation” means
DeepSeek does not define the term. The vocabulary, though, has existed in China since August: the industry standards on GEO give names to three practices: corpus poisoning, meaning satellite sites and mass publishing; manufactured consensus, meaning the same content repeated across many platforms to look like a widely held view; and answer dominance, meaning taking up an abnormal share of the answers on a topic. Until now these were rules written by associations and aimed at agencies. With DeepSeek, for the first time, it is a search engine that declares it applies them upstream, when it decides which pages reach the model.
How well the filter actually works cannot be measured from the outside, because DeepSeek publishes neither the method nor the results. What can be seen is a choice: defending against content built to fool the models has become part of the product, written into the document developers read before they pay.
In China, search is bought per call
DeepSeek is not alone. Zhipu, the company behind the GLM models, sells from its developer portal a search “built specifically for large models”, with an engine of its own and, optionally, those of Sogou and Quark, from 0.01 to 0.05 yuan per call. Kimi, by Moonshot, charges 0.03 yuan per call for its built-in search. Tencent offers on Tencent Cloud the search that powers its Yuanbao assistant, in a lite and a standard version. ByteDance, at the end of July, opened Doubao’s search to developers, with content from its own platforms and a ranking of sources by authority.
The result is a market in which search becomes a component, like space on a server: whoever builds an assistant picks the supplier on price, coverage, and quality, and can combine more than one. Footing the bill is the engine that for twenty years sold search as the final destination: in its August quarterly report Baidu writes that “our online marketing business remains under pressure”, while its AI cloud grows. The end user will not know which engine answered: they will only see the assistant.
How much it matters in Italy, today
On quotidianomotori.com, the online motoring daily we run, we log the visits of every crawler. Between June 18 and October 16, 2026, DeepSeek’s crawler made 981 visits. Over the same period OpenAI’s OAI-SearchBot made 117,907, the crawler behind Meta AI’s search 231,945, Perplexity’s 45,741. On Italian pages, DeepSeek reads little today.
Two things could change the picture. The first is the opening itself: from now on anyone building an assistant, in Europe too, can pick DeepSeek’s engine, and from November 1 its search becomes a paid feature in the coding tools that already use it. The second is the low price, which makes querying several engines at once the normal thing to do. A brand that today thinks only about Google, ChatGPT, and Perplexity will find itself inside answers built from indexes it does not know, each with its own criteria.
What changes for a brand
The first thing is that a page has to hold up to a grade, not just show up. DeepSeek says it weighs relevance, authority, and timeliness, and discards what looks made for the engine. Pages that answer a precise question, with data, dates, and a recognizable source, are the same ones that work on ChatGPT and in AI Overviews; here, on top, they get a number.
The second is that the grade can be read. With an API key you query the engine on the questions that matter in your industry and see which pages come out, with what authority and what date: a direct way to learn how a source is judged, instead of inferring it from the answers. It works for your own pages and for your competitors’.
The third is the usual one: nobody can promise the grade. DeepSeek does not publish the method, can change it whenever it wants, and says it is hunting precisely for content built to please engines. Whoever sells a “guaranteed 3” on DeepSeek is selling something they do not control. The only thing you control is your own page.
Frequently asked questions
It is the in-house engine DeepSeek uses in its chat and its API to search the web. Since October 17, 2026 it is open to developers through two functions, Search and Fetch, at 5 yuan per 1,000 successful requests. DeepSeek claims an index of hundreds of billions of pages and hundreds of millions of searches a day.
Every DeepSeek search result carries two numbers: authority, the page’s authority from 0 to 4, where 3 and above is defined as “extremely authoritative,” and rerank_score, a score from 0 to 3 that combines relevance, authority, and timeliness, where 1.5 and above is “high quality.” DeepSeek does not explain how it calculates them.
DeepSeek states in its changelog that it filters “SEO cheating, GEO manipulation, and low-quality AI content” with “a strict recognition mechanism,” but it does not define what it means by GEO manipulation and publishes neither the method nor the results. The filter targets bad practices, such as satellite sites and duplicated content, not the work of making a page clear and verifiable.
