DeepSeek launches AI-native search engine, API billed at 5 yuan per 1,000 requests

The siloed apps have blocked the flow of information; users no longer start searches from browsers, but instead look for answers to their questions on WeChat, Douyin, Weibo, and Xiaohongshu. This is the structural pain point Baidu faces in the mobile internet era. Although Baidu Search was merely uncomfortable in the mobile internet era, in the AI era it may no longer be able to survive.

Recently, DeepSeek announced that the V4.1 Flash model has been updated to the "wish-style programming, one-click come true" V4.1-Flash-1017 version. In addition to the new model, the DeepSeek API now also supports web search and crawling, billed by number of requests at 5 yuan per 1,000 successful requests.

Regarding the search API, DeepSeek introduces it in its official documentation as follows: "DeepSeek Search is an AI-native general-purpose search engine that indexes hundreds of billions of web pages worldwide, supports multilingual search, and returns high-quality results with high relevance, high authority, and high timeliness in real time. Against SEO cheating, GEO manipulation, and low-quality AIGC, we filter through strict identification mechanisms to ensure that authoritative and trustworthy information is presented to users and LLMs."

As soon as the news of DeepSeek opening up its search engine came out, Baidu's stock price fell by more than 5% at one point after the market opened on Monday, and "AI subverting search engines" once again became a hot topic.

In fact, this is not the first time AI has tried to subvert search engines. As early as September 2023, with OpenAI announcing that ChatGPT could browse the latest information online, and the AI search unicorn Perplexity‌ emerging in the spring of 2024, generative AI reshaping the information acquisition model had already begun to be widely discussed.

As a product born in the 1990s, search engines are indeed a bit old. What users get through keyword searches is an almost infinite number of results, but that is all. To further obtain the desired results, users need to complete the "filtering" themselves. In short, traditional search engines are only responsible for presenting information; whether the information is useful still requires users to judge for themselves. But unfortunately, the new generation of netizens has already been "marinated" by recommendation algorithms.

AI search, on the other hand, depicts a picture in which AI thinks on behalf of humans and directly gives filtered, refined, and integrated results. However, users at the time found in practice that AI search was not as excellent as Perplexity‌ claimed. After all, AI models from one or two years ago had two undeniable flaws, namely severe hallucinations and excessive sycophancy.

At that time, AI search would talk nonsense and even cite fabricated sources. Some netizens once complained about this: "What I am looking for is not just the correct answer, but the correct answer within a certain amount of time. This has to be faster than researching the answer myself, and also faster than verifying the answer the machine gives me." Although AI search can indeed think on behalf of users and summarize and refine content from a large number of candidate web pages, it cannot withstand the fact that AI will choose content containing a large number of factual errors.

When a product whose selling point is thinking and judging on behalf of users still requires users to spend their own time and energy verifying the accuracy of the answers, how could AI search not be "cast into the cold palace"? And the reason AI search gives low-quality or even erroneous results is that GEO (generative engine optimization) is actually the chief culprit.

The social trend of knowledge worship and science worship leads many users to subconsciously regard highly intelligent AI as an authority. Because AI comes with a "professional filter," it usually outputs content in a certain, structured, and clearly organized way, and humans are very sensitive to the fluency of content. Therefore, the easier information is to process, the more likely it is to be judged as credible. In other words, as long as it reads smoothly and looks professional, you will naturally trust it.

In addition, in today's era of information overload, what everyone needs is not necessarily an absolutely correct answer, but a reference that can help them sort out their thoughts. At this time, an always-online and patient AI, with its hidden emotional value, lowers human defenses. GEO precisely exploits humans' blind trust in AI, making the quality of AI search fall short of expectations.

Traditional search engines, having accumulated rich experience over the years in offense and defense against SEO (search engine optimization), instead have a stronger "immunity" to GEO. Therefore, AI search has failed to show an overwhelming advantage for quite a long time. At the same time, traditional search engines are also actively innovating. For example, Baidu upgraded its original search box into an "intelligent box," and Google AI Overviews also combines AI search with traditional search, thereby stabilizing its position.

But DeepSeek's self-developed AI-native general-purpose search engine is different. It explicitly filters SEO cheating, GEO manipulation, and low-quality AI content. According to actual testing, DeepSeek's search API gives explicit scores to search results based on relevance, authority, and timeliness. A score of ≥1.5 in the "rerank_score" item means high-quality content, and ≥3 in the "authority" item means an authoritative result.

It should be known that traditional search engines such as Google and Baidu also have web page quality scores, but they have always been a black box, and SEO practitioners can only guess. DeepSeek, by contrast, makes the scores public. With scores that can be relied on and a declaration against spam content, what DeepSeek wants to do is already coming into view: it wants to solve the trust problem of agents acquiring information, thereby building a search engine for agents.

In the past, when agents wanted to obtain information from the public web, developers had to design crawlers themselves. Obviously, these crawlers could not compare with Google's and Baidu's crawlers. As a result, there were cases where agents searched online but gave wrong answers with absolute certainty. Yet without enabling online search, an agent relying purely on the pre-training results of the base model could instead predict an answer with a higher accuracy rate based on its massive parameter scale.

DeepSeek has given developers an extremely low-cost option, allowing agents to avoid using Baidu, and this is the key to the latter being bearish. In the past, Baidu used crawlers to collect web pages from the public web into its own index database, and you could use it only if it let you. Now DeepSeek has made index databases no longer scarce, turning search from a permission only Baidu had into a service that can be purchased per use.

[Images in this article are from the internet]

This article comes from the WeChat public account "SanYi Life" (ID: IT-3eLife), author: SanYi Jun, published with authorization by 36Kr.