After several rounds of revelations, the AI search track finally welcomed an "important player" - SearchGPT, an AI-driven search engine launched by OpenAI.
As early as April this year, an engineer revealed that OpenAI was evaluating its own AI search product SONIC-SNC (SearchGPT); in early May, the news that OpenAI was about to release a search product was widely circulated online. However, it was not until three months later that OpenAI's SearchGPT was officially unveiled.
The launch of SearchGPT is not surprising. What people are more concerned about is whether AI search is feasible and whether it has the opportunity to seize market share from traditional search engines.
OpenAI CEO Sam Altman seems very confident about this. He called SearchGPT a "new prototype" and said bluntly, "We think there is still room for improvement in today's search." He also said that compared with the "old-school" search engine, the new SearchGPT surprised him. While praising his new product, he also did not give face to competitors such as Google and Bing.
SearchGPT, Image source: OpenAI
At present, the competition in the AI search track is fierce. There are not only excellent overseas products such as New Bing, Perplexity, You.com, Lepton Search, Genspark, and Hebbia, but also rapidly developing domestic products such as MiTa AI Search, 360 AI Search, Tiangong AI Search, Bocha AI Search, Reportify, and Devv.
When talking about SearchGPT, 360 Group founder Zhou Hongyi told Jia Zi Guang Nian: "OpenAI originally worked on ChatGPT. They thought the chat interface could solve some problems, but now they have to make a parallel SearchGPT as a new user experience and entrance. We don't know whether Search or Chat is better. It is possible that they will be equally good, or one may be replaced by the other, but this at least shows that AI search is an important scene entrance that no one can miss."
Everyone says AI search is important, but one challenge facing AI search is how to establish a business model. "If the cost is very low, you can use advertising, but it is difficult to put ads now, and ads may not cover the cost of each word search." Zhou Hongyi said, "Not all users are willing to pay, this needs to be explored."
In this article, "Jia Zi Guang Nian" will disassemble SearchGPT and related AI search products, and analyze the barriers, breakthroughs, and future of AI search products.
1. AI search is not search
To understand the advantages of OpenAI's search products, we must first understand the underlying operating principles of AI search products.
Currently , there are three main categories of AI search products on the market, namely specialized AI search products, products with AI capabilities launched by traditional search engines, and products with search capabilities made by large model manufacturers .
The first category is represented by Perplexity, Mita AI Search, and Genspark, the second category is represented by New Bing and Google AI Overview, and the third category is represented by Kimi, Doubao, and Tencent Yuanbao. Although OpenAI is also a large model manufacturer, because they already have a large model product ChatGPT, they have redeveloped a dedicated AI search product, so SearchGPT also belongs to the first category.
The underlying principle of AI search can actually be summarized as "RAG (Retrieval-augmented Generation)," which involves two steps: Retrieval and Generation. Most of the "Retrieval" is done by the API of traditional search engines, and a small part is done in the form of self-built index libraries; the main thing AI search products do is "Generation" after getting the results, using AI instead of manual work, reading the search content, summarizing and giving users a direct answer.
Of course, the links behind this are more complicated, including some important links such as question rewriting (Intent Detection), search result reranking (Reranking), and obtaining detailed content (Read Content). After reading the search content, the big model also needs to pre-process the search results, display the search solution, etc. The entire AI search process involves multiple calls to the big model. For example, 360 calls the big model 9 times for each AI search.
AI search flow chart, image source: AI product Rena
360 multi-expert collaboration (CoE) architecture model workflow diagram, image source: "Jia Zi Guang Nian"
But in general, AI search products are still developed based on mature search products. The retrieval step can be solved by calling external APIs, and the context enhancement, summary and generation steps only need to call the capabilities of the underlying large model. In other words , compared with traditional search products, AI search products are not making innovations at the technical level, but more at the product level.
This is why Perplexity calls itself an " answer engine " - as a product that is directly "shelled" from the search API to the underlying large model, Perplexity does not provide direct search capabilities. Instead, it obtains the content retrieved by the search engine through the API, and then summarizes the answers through large models such as GPT-4 and Claude. It is finally organized into a fixed format and presented to users, saving users the time of viewing and summarizing page by page.
Being able to use "shells" to achieve a valuation of 3 billion US dollars, it can be said that the victory of Perplexity, the "king of shells", is not a victory of technology, but a victory of product packaging above technology. It is a victory of taking a quick step after accurately understanding user needs.
Image source: Perplexity
Therefore, AI search is not a search, but a summary. The real competition among AI search products is not the underlying technical capabilities, but who can provide more accurate answers, faster response speed, and more intelligent user experience above the technology.
Among them, "accuracy" is the most critical point. The biggest barrier to AI search is data . To get accurate answers, the quality and quantity of the underlying data are crucial . Only when the underlying database is large enough, contains enough information, and the information is updated in a timely manner, can the large model be guaranteed to have "evidence to rely on" when acquiring content, so as to summarize and output more accurate and timely content. This is why Google has maintained a market share of more than 90% in the search engine field for many years - they have been doing indexing since the first day of their establishment in 1998, and have the world's largest and most complete index library, which can provide the most accurate and timely search results.
Therefore, if you want to make search results more accurate, building your own index library is an important solution.
At present, most AI search products are only connected to the API of traditional search engines, without re-building a set of underlying search systems. Only a few, such as Mita AI Search (podcast and library sections), 360 AI Search, and a few vertical AI search engines, have built index libraries. This is mainly because connecting to the API of traditional search engines can solve 95% of the problems, and the cost of building a self-built index library is very high, requiring a lot of manpower, financial resources and time. Therefore, if the self-built index library cannot provide better content than the Google and Bing APIs, there is no need to build a self-built index library.
AI search flow chart Image source: ThinkAny founder idoubi, AI product editor Huang Shu
How much does it cost to build your own index? Liang Zhihui, vice president of 360, once said in a podcast that the cost of crawling 50 million web pages is about 1 million to 2 million RMB, but 50 million web pages is a very small number for a search engine. Basically, to build a search engine, you need to crawl at least 100 billion web pages; if you want to index the world's web pages, you basically need 3,000 to 10,000 servers to provide support.
In other words, to make a simple search engine, you need at least 2-4 billion yuan in budget, not including the cost of PageRank servers, terminal manufacturers' protection fees and personnel costs. This is an insurmountable cost for any small or medium-sized startup.
This is also the reason why currently only a few large companies such as Google, Microsoft, and Baidu are making search engines - the cost of making search engines is too high, and only large companies have sufficient funds, talents, and equipment to do this.
In addition to the high cost, search technology and algorithms are also quite a barrier. Take Google's proud ranking algorithm as an example. It takes into account hundreds of different factors, including content quality, user experience, mobile friendliness, page loading speed, security, etc. It is not only complex in structure, but also updated in real time according to the external environment. It is reported that Google releases an average of 6 algorithm updates per day, up to 2,000 times per year; and the algorithm is highly confidential, and few people inside Google know the full picture of its search ranking algorithm.
It is conceivable that with such huge costs and extremely high technical barriers, it is as difficult for small and medium-sized search engines/AI search companies to build their own index database for the entire network as it is for Yugong to move mountains.
However, this is a problem for small and medium-sized companies, but it is not a problem for OpenAI.
As a company with a cumulative financing amount of more than 20 billion US dollars and known for its high talent density, OpenAI is naturally not short of "money" and "people". Therefore, they are fully capable of building their own index library.
So does SearchGPT have its own index library?
After asking some engineers, the general answer was: "With OpenAI's funding and talent, they are capable of building their own index library. However, since building an index library for the entire network is too time-consuming, labor-intensive and expensive, they should have built a partial index library, but they will not give up access to external search APIs and use crawlers to perform real-time searches to supplement answers."
Technology media TestingCatalog News dug out the source code of SearchGPT, confirming this speculation - they found that SearchGPT still accesses the API of the Bing index library, but unlike the general Bing search function currently provided by ChatGPT, SearchGPT is better at providing real-time information. This is mainly achieved by web crawlers.
Image source: TestingCatalog News
The statement on the OpenAI developer page further confirms this speculation. In the overview of the developer platform, they wrote: "OpenAI uses web crawlers and user agents to perform actions for its products, which are either automated or triggered by user requests."
Image source: OpenAI Platform
In summary, SearchGPT is likely to have adopted a technical approach that combines " self-built partial index library + access to Bing's API + real-time web crawler " to ensure the accuracy and timeliness of search results.
In addition to its own index library, another data barrier that OpenAI has built for its AI search products is content .
On the release page of SearchGPT, OpenAI mentioned that they are committed to building a thriving ecosystem of publishers and creators, helping users discover publisher websites and experiences, while bringing more options to search.
The media they cooperated with included The Atlantic Monthly, the Associated Press, and Axel Springer, the parent company of Business Insider, as well as the media giant News Corp, which owns The Wall Street Journal, The Times, and The Sun. OpenAI tried to exchange the rights to crawl these media content with recommended links, and at the same time stated that media and publishers can choose how to present content sources in SearchGPT.
Cooperation with professional information content providers such as institutional media is a powerful blow to OpenAI's competitors. Due to the existence of paywalls and anti-crawler mechanisms, many AI search products are unable to crawl the content provided by professional institutional media, which also causes incomplete results and poor user experience generated by AI search products. After establishing a cooperative relationship with the media, many exclusive reports of the media can be searched in SearchGPT, which undoubtedly prospers the content ecology of SearchGPT and guarantees its user experience.
In addition to improving the underlying data and content ecosystem, establishing a cooperative relationship with the media is also a deliberate move by OpenAI to ensure copyright compliance.
Since the launch of ChatGPT, copyright lawsuits surrounding generative AI products have been endless. For example, The New York Times has spent $1 million to sue OpenAI and Microsoft, and eight publication groups under hedge fund Alden Global Capital (including the New York Daily News and the Chicago Tribune) have also filed lawsuits against the two companies. Perplexity, also a generative AI product, also received a letter from Forbes because it used a report from Forbes in its search results without accurately indicating the source.
Therefore, establishing a healthy cooperation mechanism with the media and publishers can not only obtain more content sources, but also avoid many unnecessary copyright disputes. There may also be opportunities to make profits through advertising in the future. It is a good thing to kill three birds with one stone.
Whether it is building its own index library or obtaining exclusive content through cooperation with the media, it will help SearchGPT obtain high-quality data, establish data and content barriers, and thus gain certain latecomer advantages in the already smoke-filled AI search war.
2. Reconstructing user experience with low latency and multimodality
If self-built index library and cooperation with the media are the first data barrier that OpenAI has established for itself, then extremely low latency, multiple rounds of question-and-answer interactions, and multimodal result presentation are the second user experience barrier that OpenAI has established for itself.
As we mentioned earlier, AI search is not a search, but a summary. AI technology itself solves the problem of "information matching", and the purpose of users using AI search is not just to obtain information, but to "hire" products to meet their needs. Although large model technology will enhance the original search capabilities to a certain extent, in order to make a good AI search product, in addition to the information level, efforts must be made on the user side, and creative interaction with users must be carried out to make the information side and the interaction side "roll on both wheels" to give birth to a phenomenal AI search product.
What kind of interaction can bring a good user experience?
In his 10,000-word review of his AI search engine, idoubi, an independent developer and founder of ThinkAny, mentioned: “ The three most important points for a good AI search engine are accuracy, speed, and stability, which means that the results should be accurate, the response speed should be fast, and the service stability should be high .”
The self-built index library and cooperation with the media mentioned in the previous article ensure "accuracy", but how to ensure "speed" and "stability"?
It is crucial to reduce the delay in responses and be able to conduct multiple rounds of questions and answers.
SearchGPT also achieves low latency in response. From the demonstration video, we can see that SearchGPT's response speed is very fast. It only takes 1-2 seconds from inputting a question to giving an answer.
SearchGPT’s low latency response, video source: OpenAI
If the official video still has the suspicion of editing, then the experience video released by some netizens who have obtained the experience qualification has eliminated this suspicion. From the video, we found that SearchGPT's response speed is indeed very fast, and it only takes 3 seconds from asking a question to giving an answer.
SearchGPT actual test video, video source: X @Paul Covert
How is this ultra-high "real-time" performance achieved? There are two ways: the first is to crawl the data in advance and embed it into the vector database, and the second is to search the content on the Internet in real time. These two points are also the basis for building a self-built index library, which requires the use of vector search, real-time indexing, real-time data analysis and other technologies.
How did OpenAI acquire these technical capabilities? This has to do with their recent "big move".
In June this year, OpenAI acquired Rockset, a real-time analytical database company, for $500 million, which is the largest deal in OpenAI's history. Rockset is a real-time fully indexed database known for its real-time indexing and querying capabilities. Its services are mainly aimed at application scenarios that require real-time processing and querying of large amounts of data, such as real-time dashboards, search indexes, streaming data analysis, etc.
Rockset official website, image source: Rockset
Rockset has a key-value storage system called RocksDB , which was developed by two co-founders Venkat Venkataramani and Dhruba Borthakur who previously worked at Facebook. It is particularly good at fast reading and writing, and therefore allows users to query and analyze large-scale data sets in milliseconds .
Through technologies such as streaming data ingestion, automatic indexing, memory optimization, and high-concurrency queries, Rockset can greatly improve SearchGPT's response speed to user inquiries, ensuring the "speed" of AI search products.
In addition to being "fast", Rockset can also ensure the "stability" of AI search products, namely multiple rounds of interactive questions and answers.
As a real-time fully indexed database, Rockset can provide more comprehensive and complete data and also supports multi-dimensional queries, which means that users can filter results based on different attributes and conditions. In addition, through contextual indexing, SQL query and other technologies, Rockset can enable AI search products to better store and retrieve contextual states, maintain the continuity of conversations, and enhance the multi-round interactive question-and-answer experience of AI search products.
SearchGPT’s follow-up function. Video source: OpenAI
Before OpenAI acquired Rockset, the common solution in the industry to solve the data indexing, querying and storage problems of large models was to add a "plug-in" to the large model, which is the so-called "vector database". In the past year, vector databases have become popular all over the country, and almost every vector database manufacturer is marketing with "memory of large models" as a selling point. Rockset can not only do vector indexing, but also index different forms of data such as keyword indexing and metadata indexing, which can provide higher quality search results and meet diverse query needs.
This is another core technology of Rockset besides RocksDB - Hybrid Search. This is a multifaceted search method (integrating vector search, text search, metadata filtering, etc.), which is highly flexible and allows indexing and using multiple types of data, including real-time data. Therefore, it is often used in scenarios such as context-related product recommendations and personalized content aggregation. This technology plays an important role in Microsoft's New Bing and Google's AI Overview.
Thanks to this "hybrid search" technical capability, SearchGPT has a third highlight - multimodality. It not only provides text-based answers, but also adds a variety of content dimensions to the result display, including data, lists, pictures, videos, etc., which improves the richness and comprehensibility of information.
For example, a netizen who was qualified for the test asked it about the weather in London, and SearchGPT directly gave a weather forecast for the next seven days, including small components such as weather icons.
Image credit: X @Kesku
Netizens also asked SearchGPT what the best time and place for a picnic in London was. While it gave several park options, it also listed some photos of London in different weather conditions on the left.
Image credit: X @Kesku
On the mobile side of SearchGPT, the multimodal performance of its answers is also excellent: for example, when querying Nvidia's stock, a stock chart will be given; for another example, when searching for a song, SearchGPT will directly give a Youtube video, and users can play it without clicking on the web page.
Image credit: X @Kesku
Why is multimodal presentation so important? This is mainly because most AI search products currently still give answers in text. Although text replies can solve most of the user's problems, in a competitive environment with an increasing number of players and homogeneous internal competition, whoever can provide more dimensional and richer answers may be able to hold the key to the next stage of competition in AI search.
Multimodal retrieval not only helps users understand search content more intuitively and improve their search experience, but is also of great benefit to product iteration - AI search products equipped with multimodal models can simultaneously process different types of data such as text, images, sound, video, etc., and obtain more comprehensive knowledge and context from diversified information, which is crucial for the understanding and execution of complex tasks; through cross-modal learning, AI search products and the models behind them can also be better generalized to unseen situations, improving the accuracy, adaptability and practicality of search results.
The multimodal presentation of search results mainly comes from the support of model capabilities. Basically, manufacturers who are capable of long large models are also capable of making multimodal AI search products. This is also the reason why a considerable number of multimodal AI search products come from large manufacturers - for example, the 360AI search launched by 360 Group supports users to ask questions by taking photos, and also supports document, image, audio, and video search; Kunlun Wanwei's Tiangong AI search not only supports AI image recognition and AI image generation, but also has multimodal capabilities of receiving, writing, reading, chatting, speaking, talking, listening, and singing, and also generates mind maps and outlines; Genspark, an AI search product made by Jing Kun, the former CEO of Xiaodu, can also generate visual charts and pictures based on user questions.
360 AI search results, Image source: 360 AI search
Tiangong AI search results, image source: Tiangong AI
Genspark search results, images stitched together by "Jia Zi Guang Nian"
As one of the most powerful large-model companies in the world, OpenAI's models are naturally capable of multimodal processing.
In the demo released by SearchGPT this time, a feature called "Visual Answers" was demonstrated, but OpenAI did not explain in detail how it works.
Image source: OpenAI
As mentioned above, TestingCatalog News has uncovered the source code of SearchGPT, which not only reveals the Bing interface, but also shows that the search results are powered by a multimodal model. Although we cannot tell what the specific model is and its processing flow, the multimodal model should have the ability to automatically understand images.
Image source: X @TestingCatalog News
Zhu Jie, chief product architect of Baidu's database department, believes that if SearchGPT wants to provide a different experience from Perplexity and other products and achieve a curve overtaking as a latecomer, the biggest attraction is "multimodality". Rockset has already provided the ability to retrieve multimodal data. If it can rely on the multimodal large model to play more "fancy" on the interactive end, then greater user growth is just around the corner.
In short, Rockset's hybrid search capabilities, coupled with OpenAI's multimodal large model, allow SearchGPT to provide a better interactive experience, which is also the original intention of OpenAI's acquisition of Rockset.
In the announcement of the acquisition of Rockset, OpenAI wrote: "AI has the opportunity to transform how people and organizations leverage their own data... We will integrate Rockset's technology to power our retrieval infrastructure across products, and members of Rockset's world-class team will join OpenAI." (AI has the opportunity to change the way individuals and organizations use their own data... We will integrate Rockset's technology to power our cross-product retrieval infrastructure, and members of Rockset's world-class team will join OpenAI.)
OpenAI acquires Rockset announcement, image source: OpenAI
OpenAI said it hopes to acquire Rockset to support its own cross-product retrieval infrastructure. From this, it can be seen that access to and processing of real-time data has become an important part of the current AI arms race. Some industry insiders also pointed out that this acquisition essentially shows that vector databases cannot really solve the problem of "artificial intelligence memory (memory)", but general real-time databases can do it . OpenAI, aware of this, is also trying to establish a solid "general data base" for its internal large models, thereby reducing the illusion of large models and improving the accuracy, timeliness and contextual relevance of AI-generated content.
The acquisition of Rockset was an extremely important step before OpenAI launched SearchGPT. It made up for OpenAI's data shortcomings, improved the efficiency and speed of OpenAI's data extraction, processing and analysis, and enabled OpenAI to continuously extract and index data from various sources. It can be said that Rockset not only helped OpenAI reduce the latency of AI search products and allowed SearchGPT to give answers "faster", but also played a huge role in building an efficient, complete, and real-time updated index library, further improving the timeliness and accuracy of AI search product answers.
3. Search GPT, why now?
Why is SearchGPT launched now? This may be related to the financial difficulties faced by OpenAI.
A recent report in The Information mentioned that as of March this year, OpenAI's inference costs (the cost of renting Microsoft servers) reached US$4 billion, and training costs may soar to US$3 billion this year, plus US$1.5 billion in labor costs, so OpenAI's corporate operating costs this year are around US$8.5 billion.
From the revenue perspective, although OpenAI's recent monthly total revenue is US$283 million, its annual revenue may reach US$3.5 billion to US$4.5 billion (according to FutureSearch's estimates, OpenAI's annual recurring revenue in 2024 will be approximately US$3.4 billion, which is not much different), but whether it can reach this figure depends on sales in the second half of this year.
That is to say, even if OpenAI tries its best to achieve annual sales of US$3.5 billion, it will still show a loss of about US$5 billion on its books.
This forced OpenAI to try its best to find ways to make money. In OpenAI's revenue structure, the bulk of revenue is still contributed by the C-end - 55% of revenue comes from paid subscribers who purchase ChatGPT Plus (about 7.7 million), and the remaining 45% of revenue comes from ChatGPT Enterprise Edition (21%), API revenue (15%) and ChatGPT Team Edition (8%). Therefore, starting with a strong C-end base is the key to OpenAI's revenue breakthrough.
Image source: FutureSearch
In the C-end scenario, search is one of the most valuable business models. As a natural scenario with a large number of users, search was once the first traffic entrance on the PC side. Baidu has achieved its status in the industry through search, and Google has also become one of the Internet companies with the highest market value in the world by relying on search. Of the $370.394 billion in revenue of its parent company Alphabet in 2023, $175.033 billion was contributed by Google's search business (accounting for about 47.26% of the total revenue).
Image source: Alphabet 2023 Annual Report
Although ChatGPT's daily active users remain around 50 million and it is still the only killer application in the field of generative AI, it is still some distance away from Facebook, Youtube, Instagram and other platforms in the Internet era.
Li Guangmi, CEO of Shixiang Technology, said in an interview in April this year: "If it were me, I would first increase ChatGPT's DAU from 50 million to 300 million... Because ChatGPT currently has less than 10 million paying users. If we can achieve 30 million paying users, that would be $6 billion in subscription revenue per year, which would be able to support AGI's annual investment in a relatively healthy way."
Search is a $200 billion market. The current market size of so-called "AI native search products" is only a few billion dollars, and there is still a lot of room to be filled. Once the search function is introduced into ChatGPT, ChatGPT may derive more business models such as advertising and recommendations in addition to the subscription system, expanding more revenue sources.
In addition to using search to boost traffic and revenue, seizing the time window brought about by the reduction in reasoning costs may be the second reason why OpenAI chose to launch SearchGPT at this time.
The revenue of search products does not come directly from C-end users, but from B-end users brought in after a certain amount of C-end users have accumulated. Therefore, before the cash flow of search products becomes positive, there is not only a long user acquisition and operation cycle, but also considerable product and technology construction and maintenance costs.
Take traditional search engines as an example: According to a report by Morgan Stanley, Google will have about 3.3 trillion searches in 2022, with an average cost of about 0.2 cents per search. In terms of revenue, according to data revealed by Kunlun Wanwei Chairman Fang Han in a podcast, Google's advertising revenue for searches in 2023 will be about $160 billion. Based on this calculation, the revenue generated by a search is about 5 cents.
So how much does AI search cost? As we know from the previous article, AI search adds the ability of large models to traditional search engines. Therefore, the cost of AI search is 0.2 cents. In addition to the computing power, electricity consumption and hardware investment generated by calling large models, the cost of a single AI search is about 10 times that of traditional search engines, which is 2 cents.
Since May this year, with the improvement of hardware performance, optimization of algorithms and innovation of technical architecture, the inference cost of large models has been declining: for example, the Deepseek-V2 version of the model released by Magic Cube Quantitative has reduced the memory usage to 5%-13% of the past through the innovative MLA architecture, and its original DeepSeekMoESparse architecture has reduced the amount of calculation to the extreme. All of this has contributed to the reduction of its model inference cost; for example, OpenAI itself released GPT-4o mini on July 19. By improving the model structure and optimizing training data and processes, GPT-4o mini has achieved performance close to GPT-4, but the cost is more than 60% cheaper than GPT-3.5 Turbo, and compared with the text-davinci-003 version of GPT-3 two years ago, the cost has been reduced by 99%.
As the cost of large model inference continues to decrease, the launch of SearchGPT at this time not only has advantages in terms of model call costs, but can also further spread the cost of a single search with the help of the growing number of users.
4. Don’t just search.
If OpenAI really regards SearchGPT as an important step to attract traffic and break through commercialization, then how can SearchGPT make breakthroughs in its products to have a better chance of becoming OpenAI's "traffic breakthrough point", attracting more daily active users and achieving better conversions?
"Jia Zi Guang Nian" believes that there are two directions:
The first is UGC+AIGC , and the second is automated workflow .
Let's talk about the first one first. As we mentioned earlier, the biggest barrier to AI search is data, that is, content. China's Internet content has been divided into various camps such as Baidu, WeChat, Zhihu, Xiaohongshu, Taobao, etc., and there is a high degree of separation between them. Compared with China, the ecology of the Internet abroad is more open, and data flow and sharing between various content camps occur more frequently.
As a result, some overseas AI search products seized this opportunity and began to explore the AIGC+UGC content generation method, trying to enrich their own content ecology. Typical representatives are Perplexity and Genspark.
Perplexity has launched the Pages feature, which allows users to generate content directly on the Perplexity platform. Users can not only convert the answers originally obtained from the search into "Pages" in the form of an article, but also directly enter the query word, and Perplexity will search for public articles or reports to directly generate a Page. Not only that, Perplexity can also rewrite, format or delete Pages according to user needs, or insert relevant pictures and videos.
Image source: Perplexity
Genspark, founded by Jing Kun, the former CEO of Xiaodu Technology, is no exception.
On ProductHunt, Genspark describes itself as: "An AI agent engine where specialized AI agents conduct research and generate custom pages called Sparkpages . Sparkpages is unbiased and SEO-driven content that synthesizes credible information to provide more valuable results and save time for users."
Sparkpages is the core of Genspark: after users search for questions, Genspark will not only give an answer first like other AI search products, but also generate a theme webpage by integrating search results. This theme webpage is Sparkpages. It will display text, pictures, videos, etc. related to the search topic in a structured way and present them to users in a waterfall flow.
Image source: Genspark
Sparkpage not only allows users to access information more quickly and efficiently through the form of "information aggregation webpage", but also allows users to copy and edit the generated pages: there is a button " Copy and Create My Plan " in the upper right corner of each Sparkpage. Clicking it will open an editing page with the content of the previous Sparkpage copied. Users can ask contextual questions in the Copilot column on the right for the Context Pool on the left, call external searches to get answers, and can also copy the content in the Copilot to the Sparkpage on the left with one click. Like Perplexity Pages, Sparkpage can also be indexed by search engines' SEO.
Liu Jia (pseudonym), Jing Kun's former subordinate and AI practitioner, told "Jia Zi Guang Nian" that when Jing Kun was still working at Xiaodu, he mentioned many times that he wanted to "reconstruct the content ecology behind search", so it is not surprising that Genspark adopted this "AIGC+UGC" model.
"The biggest barrier to AI search currently lies in the quality of the information source behind it. The advantage of Genspark is that in the short term, it can meet search needs through high-quality information sources + AI understanding, and generate high-quality structured summary pages Sparkpage after meeting the needs. These contents are again included in the search, which is equivalent to the 'self-production and self-sales' of AI search product content. This method can form a content data flywheel. Assuming that this flywheel can turn, then in theory, the more users use it, the more high-quality content there will be, and the content barrier behind the search can be slowly built up." Liu Jia said.
Liu Jia's comments on Genspark on social media, image source: Jike
Although SearchGPT has done a good job in AIGC content, with a simple interface, comfortable user experience, and high-quality information sources, it still lacks in UGC. I wonder if this feature can be seen in subsequent official versions.
The realization of "automated workflow" of AI search products mainly depends on Agent, especially the ability of multi-agent. The biggest difference between AI search engines and traditional search engines is that they can not only summarize and classify search results, but also take further actions, such as generating mind maps, generating PPTs, and planning work and life. The ability behind this depends on Agent.
In fact, most AI search products have added AI Agent capabilities. Those "fancy" presentations on top of data organization are basically implemented by Agent plug-ins, such as 360 AI search document analysis Q&A, web content analysis, audio and video analysis and mind map generation functions, Mita AI search's mind map, sorting out various related people and events, searching for academic papers and podcasts, Tiangong AI search's document + audio and video analysis, AI music, AI PPT and other functions;
The B2B AI search engine recently launched by Alibaba International seems to have the capabilities of an AI Agent behind it. It can actively understand the natural language of buyers and convert it into professional purchase requests. Furthermore, it can predict demand and provide suggestions based on global market data to achieve more accurate matching. As long as users ask in "plain language" in the new AI search engine, it will understand the user's needs more accurately through dialogue with the user, recommend multiple options for the user, and actively compare and summarize the pros and cons, and finally help the user complete the transaction and fulfill the delivery.
Alibaba International announced the launch of the world's first AI-driven B2B search engine in Paris, France. Image source: Alibaba International
However, in terms of multi-agent integration, there are relatively few AI search products that can do this. Only products such as Bocha AI Search and Genspark are trying.
For example, Bocha AI Search pioneered the multi-agent search (Agent Search) + multi-modal search (Media Search) model, introduced AI agents and high-quality UGC content, can identify user search needs, and match corresponding agents to answer, and also supports searching web pages, short dramas, videos, pictures, etc.
The overall architecture of Bocha AI search is developed based on the Coze developer platform, and has cooperated with Coze to open up the intelligent agent publishing channel. In the future, the intelligent agents created by developers can be directly published to Bocha, which can not only provide everyone with richer search content, but also provide targeted traffic for developers' intelligent agents.
Genspark also uses a multi-agent framework, where each agent can provide professional services for specific types of queries to ensure the accuracy and relevance of search results. A blogger disassembled Genspark's search process and found that during each search, Genspark calls at least 6 AI agents and 4-14 large models.
Image source: AI product Rena
The above-mentioned agents are more about using agents to realize automated workflows, and are more focused on general search. In terms of vertical search, Genspark also uses three agents to process users' search intentions for travel, products, and pictures.
Image source: Genspark
Users can not only click on the corresponding label to enter a specific Agent page, but also click on the input box, select the corresponding Agent, and enter a targeted Agent search.
Image source: Genspark
During the 12th Internet Security Conference, Zhou Hongyi told "Jia Zi Guang Nian" that the most important thing about Agent is that it solves the problem that large models cannot "think slowly", and Agent is indispensable in AI search products.
"The human brain has two ways of thinking: fast thinking and slow thinking. Big models are often fast thinkers . When asked what 2+2 equals, they can answer it fluently. But when it comes to complex problems, such as writing a paper or doing analysis, humans need to think slowly , and mobilize planning, reflection, and logical reasoning abilities. The speed is slow, but the accuracy is high. Big models currently only have the ability to think fast, not slow. We propose to use the Agent framework to create a slow-thinking system, and enhance the planning of big models through knowledge and tools, so as to build the slow-thinking ability of big models." Zhou Hongyi said.
However, perhaps because it is still in the product prototype stage, we have not seen the "multi-agent" capabilities in SearchGPT. But since SearchGPT is integrated in ChatGPT, in the future OpenAI may connect it with the agents (GPTs) in the GPT Store and work together to provide services to users.
In addition, Zhou Hongyi also emphasized that when large-scale model companies make AI products, they must pay attention to "combining with scenarios" to achieve the best user experience. The launch of OpenAI's search product also means that it is paying more and more attention to the importance of "scenarios".
As Marc Andresson, founder of a16z, advised Aravind Srinivas, founder of Perplexity, “Don’t do search alone at all costs” - whether it is UGC or multi-agent, they are all the “arts” of search products; only by finding the right user usage scenarios can you “combine the art” and achieve growth.
*References:
"In-depth Analysis of AI Search Products - Analysis of Search Principles and Business Models", AI Product Rena
"Punch Google, kick Perplexity, Genspark wants to be a new species where beauties cook meals for you and feed them to you!", AI Product Uncle Huang
I made an AI search engine, Aidoubi
"Why OpenAI Could Lose $5 Billion This Yer", The Information
"OpenAI Revenue Report", FutureSearch
"Zhang Peng talks with Fu Sheng and Fang Han: Where are the opportunities in AI search and why didn't OpenAI join in?" AI Insider | AGI Insider Podcast
"What is 360AI Search, the fastest growing search engine in China, doing? | Interview with 360VP Liang Zhihui", AI Product Manager Podcast
"AI Search Revealed: How difficult is Perplexity, the King of Shells? Is there a chance for independent development?", Hard Ground Hacker Podcast
(Cover image source: "Jia Zi Guang Nian" generated using AI tools)
This article comes from the WeChat public account "Jia Zi Guang Nian" , author: Wang Yi, and is authorized to be published by 36Kr.
