The "new search" business behind large models, how deep is the water?

"We have created so much brand-related content, how can it be showcased in the responses of large models?"

Recently, this question has become a KPI set by many business owners for their marketing departments, which is "how to do GEO."

More than a decade ago, business owners were most concerned about "how to do SEO." This subtle shift reflects a phenomenon: the "power center" of search engines is gradually migrating from traditional web indexing to generative AI models.

Similarweb shows that in July 2025, chatgpt.com will have a monthly visit volume of approximately 5.7 billion, a month-on-month increase of 6%, with its global website ranking rising to 5. OpenAI disclosed on August 7 in its official blog that "nearly 700 million people use ChatGPT weekly"; processing more than 2.5 billion prompts daily.

The usage data of the AI search newcomer Perplexity is also accelerating, with Perplexity's CEO stating on June 5 that its query volume in May was 780 million, with a monthly growth rate of over 20%. On July 24, Fortune further reported that Perplexity has "approached 1 billion queries per month."

At the same time, traditional search's traffic delivery to websites continues to plummet: Authoritas's research indicates that when AI Overview appears on Google results pages, news sites' click-through rates can be diluted by up to 79%; GrowthSRC's monitoring of 200,000 keywords also shows that the average CTR (click-through rate) of Google's first natural result has dropped by another 32% year-on-year.

Behind these data changes is a shift in user habits: from the traditional path of "search - filter - click on collection pages" to the direct dialogue of "ask - get answers."

This means that corporate marketing strategies have also shifted from "how to let users find us" to "how to let AI remember us and actively recommend us."

Generative AI search is disrupting the old order of brand traffic acquisition, but the key lies not only in technical adaptation but also in a complete update of marketing thinking; execution may seem simple, but it hides high costs of trial and error.

To clarify the opportunities and pitfalls, we consulted industry insiders, including GEO business leaders and AI search engine technology providers, to gain insights into the opportunities and risks in this transformation, and attempted to present a systematic interpretation through this article.

Core

• Is the logic of GEO optimization consistent with that of SEO optimization?

• Does spending money on GEO services really yield results for businesses?

• Will model companies and platform providers give the green light to GEO like they once tacitly allowed the SEO industry?

• The GEO service agencies emerging in the market have no unified pricing standards; are their charges reasonable?

GEO is not the next SEO

Traditional search engines and AI search engines have essential differences in their working principles and user experiences, which determine that SEO and GEO are two completely different sets of rules.

The former has established guidelines, while the latter has infinite variables.

Traditional search engines (such as Google, Bing, etc.) are like a game of chess, with the board, pieces, and moves all outlined in a "rulebook." They rely on keyword matching and publicly researched algorithms like PageRank to rank web pages based on factors such as relevance, authority, backlinks, and user experience, then present them for users to click and browse.

The reason SEO has developed into a multi-billion dollar industry is that this set of rules, while complex, is sufficiently transparent: as long as one understands algorithm preferences and optimizes content and link structures, one can predictably influence rankings.

The world of AI search, on the other hand, is more like conversing with a brilliant scholar; after a question is posed, the system directly generates an answer or summary, rather than listing a series of links.

Its judgments are based on the semantic understanding of large models and the RAG process, with decision paths deeply buried in a black box, making it difficult for outsiders to retrace—what evidence is cited and what is ignored rarely follows a predetermined trajectory. This "knowledge generation" paradigm significantly reduces the controllability of GEO (Generative‑Engine Optimization).

Figure: Differences in the working principles of AI search and traditional search

However, AI search is not without context: a team from Princeton published a groundbreaking study at the 30th ACM SIGKDD (International Conference on Knowledge Discovery and Data Mining), which identified nine GEO strategies that can enhance visibility through experiments, potentially yielding up to a 40% increase in exposure; the E-E-A-T (Experience, Expertise, Authority, Trustworthiness) principles emphasized in the SEO era still apply in GEO, where high-quality, original content that genuinely solves user problems remains the most favored material by the model.

Figure: Nine factors affecting content visibility

However, this study has an important reversal: the keyword stuffing strategy that has repeatedly proven effective in traditional SEO is not only ineffective in GEO but may even backfire.

Research data shows that excessive keyword stuffing significantly reduces the probability of content being cited by AI, further proving the fundamental differences in underlying logic between GEO and SEO.

In summary, the most fundamental difference between SEO and GEO lies in the paths to achieving "authority" and "trustworthiness."

• From "Link is King" to "Citation is King": The core of SEO is backlinks. The more high-authority websites link to a site, the higher its authority. However, in GEO, AI places more emphasis on "citations" and "provenance."

• From "Keyword Optimization" to "Semantic Entity Optimization": SEO organizes content around keywords, while GEO needs to organize information around "entities" and "knowledge graphs." AI is not identifying keywords but understanding a "concept."

Therefore, having clear, accurate, and comprehensive entries in knowledge bases like Wikipedia and Baidu Baike is equivalent to providing AI with a machine-readable "resume" about your brand, which is more important than keyword stuffing.

Thus, the final conclusion is that following the experience of SEO will not lead to discovering new territories in GEO.

Can companies guarantee results by spending money on GEO?

The uncertainty of AI search has created new demands: more and more companies are worried about being "disappeared" in AI searches and are seeking GEO service providers to help enhance brand visibility, leading to the emergence of three types of players in the market—traditional SEO companies transforming into GEO, agencies extending content marketing to GEO, and startups that specialize in AI search from the beginning.

Globally, the field will become more segmented. In the area of monitoring and analysis tools, BrightEdge has launched Generative Parser, specifically designed to monitor brand performance in AI search results and track brand mentions and sentiment on platforms like ChatGPT and Bard.

Conductor has developed an AI Content Optimization platform to help companies optimize content to improve visibility in AI searches.

Flow Agency, based in Berlin, announced in 2024 that it would change its name from "Flow SEO" to "Flow Agency," focusing on GEO consulting with the goal of "making brands appear in the answers of ChatGPT, Perplexity, and Gemini."

So, what are the common practices in GEO services? Tencent Technology recently had in-depth discussions with the heads of two domestic GEO companies and summarized the following key practices:

• Knowledge Graph Construction: Helping companies organize core information and establish entity relationships for brands, products, founders, etc., and submit them to various knowledge base platforms.

• Authoritative Content Collaboration: Utilizing media or academic resources to help brand content appear on high-authority websites or publications to gain AI "citations."

• Structured Data Deployment: Deploying AI-readable structured data on corporate websites to enable AI to more accurately capture and understand information

• Content Optimization and Generation: Utilize AI tools to generate a large amount of semantically rich content around specific themes, aiming to secure a place in the AI knowledge base.

• Multi-Platform Coverage Strategy: Simultaneously optimize the inclusion effects on multiple mainstream AI platforms such as DeepSeek, Kimi, Doubao, Qianwen, and Yuanbao.

At the same time, Tencent Technology has noticed that many companies have specifically proposed their own "GEO methodology" and added self-developed tools, such as prompt strategy tools and GEO content creation agents, to emphasize the scientific and unique nature of their GEO methods. Additionally, some enterprises have launched a "User Resonance Index" tailored to platforms with short video or social attributes like Xiaohongshu and Douyin, based on the characteristics of the Chinese market.

However, if the above methods are adopted, can GEO really achieve 100% controllable and quantifiable results?

In response, Weng Rouying, CTO of Bocha AI, a company providing search engine technology for AI, told us:

"Firstly, regarding the question of 'whether it is controllable,' the current conclusion can only be said to be 'partially controllable.' There are already some validated methodologies in the industry: by embedding recognizable citation markers in the content and supplementing structured data, it is possible to enhance the brand's appearance and citation rate in large model search results to a certain extent."

Corresponding evaluation systems have also been initially formed, such as monitoring citation ratios and model inclusion probabilities. But ultimately, the large models that GEO relies on are still black box systems—we can only quantify phenomena at the data level and cannot truly 'control' the internal decision-making processes of the models.

The difficulty in quantifying GEO effectiveness largely depends on the content preferences of the models and applications themselves. Taking DeepSeek as an example, it favors community content; while Doubao naturally tends to prioritize materials within its own ecosystem. Due to the different 'content tastes' of each model, a unified evaluation system almost does not exist. In practice, it is necessary to develop specific evaluation dimensions and optimization strategies for each model and continuously iterate based on model feedback.

This means that GEO work is destined to be a refined process involving multiple models and indicators in parallel, rather than being effectively addressed by a single methodology.

Confusion in GEO Service Pricing

From Weng Rouying's response, it can be seen that certain methods may indeed increase the probability of a company's brand being cited by large models.

This gives GEO services a reasonable value of existence, but there is currently no unified standard for pricing or evaluating effectiveness.

In comparison to SEO, according to industry research by Backlinko, the average monthly fee for global SEO in 2025 falls within the range of $1,000–$2,500; if billed hourly, it generally ranges from $50–$100/hour, with top agencies being about 30% higher than freelancers. Further breakdown shows that enterprise-level SEO in the U.S. and Western Europe mostly concentrates in the range of $3,000–$7,500/month, with complex or highly competitive industries soaring above $20,000 The pricing logic of GEO (Generative Engine Optimization) tends to favor a "capability ladder," so there are almost only high-tier monthly packages or one-time project quotes available in the market, with very few hourly billing options.

Creative Click Media labels GEO as "hundreds to thousands of dollars per month," focusing on content foundation and Schema markup (simply understood as tagging key information on web pages); the premium package adds features like AI content feeding and monitoring.

Perrill clearly states a "starting price of $6,000 per month," with additional charges based on industry competitiveness and measurable indicators.

Public quotes from industry consulting firms generally range from $3,000 to $20,000 per month, often tied to KPIs (such as ChatGPT Top-3 citation rate, Perplexity citation rank) with stage rewards or penalties—this kind of "partially performance-based" model is also a rare new attempt in the history of SEO.

Another established SEO tool provider, BrightEdge, has launched "AI Catalyst," claiming it can monitor Google AI Overview and ChatGPT citation ratios in real-time, charging based on citation volume and entity coverage.

WebFX packages GEO into its existing SEO pricing system, with monthly fees ranging from $3,000 to $20,000, still priced based on the number of pages and keywords.

After communicating with Tencent Technology and several domestic GEO service providers, it was found that the charging model generally relies on monthly service fees or project-based pricing, with no unified pricing standard, requiring separate quotes based on different companies' situations. There are also some new billing models, such as pricing based on semantic scope.

In response, GEO service providers explained to Tencent Technology that each core keyword can extend to multiple similar keywords, forming complete semantic coverage, and the difficulty of semantic coverage varies for each project, hence the different pricing.

In addition to the lack of a unified pricing standard, effect verification is currently the biggest challenge. Service providers often use "success cases" or "screenshots of AI responses to specific problems" as proof, but these often lack stability and replicability. Due to the "black box" nature of AI, no one can guarantee stable citation across all relevant queries.

To enhance the credibility of optimization effects, some service providers have even proposed a "quick verification mechanism: promise results within 5 days, or refund."

Platform Game GEO Gray Market

From the user's perspective, SEO feels more like "hiring an outsourced operation team"; GEO, on the other hand, resembles "purchasing a ticket to the AI visibility black box," requiring a higher upfront investment to "buy capabilities and methodologies," while also bearing the risk of non-standardized metrics.

Additionally, the market is flooded with many low-priced, low-quality services. These services typically follow a routine of purchasing an "AI optimization package" for just a few dozen yuan, receiving a file package, and requiring users to manually publish over 1,000 articles, which may have been produced in bulk at low cost by the service provider using AI The results are completely untraceable and assessable.

At the same time, the gray industry is also thriving. For example, a set of "shadow Prompt" scripts is circulating in the developer community: writing instructions in white text on a white background or HTML comments to induce LLMs to prioritize citing target URLs in their responses. This "invisible injection" technique has been proven by academia to bypass human review.

AI search platforms are not sitting idly by as GEO grows wildly. Google updated its Spam Policy in March 2024, explicitly defining "scaled AI-generated content with no added value pages" as spam content for the first time, which can be directly penalized manually.

OpenAI also included "automatic detection and manual review" in its Usage Policies this spring, establishing a blacklist for suspicious URLs and lowering their weight. Perplexity's approach is to introduce a "Focus/Choose Sources" mode, allowing users or systems to limit the range of trusted sources; for websites involved in copyright disputes, the platform will downgrade citations at the internal permission level.

Even for normal GEO strategies, the attitude of large model companies is one of "cautious openness." Bocha AI CTO Weng Rouying said, "Large models will cautiously encourage companies to publish some content within the ecosystem of large model companies, such as Doubao may tend to include content from Douyin or Toutiao. However, the underlying logic of GEO and large models is conflicting. Large models aim to provide users with accurate content, while GEO is essentially responsible for commercialization. Their underlying logic is inconsistent, so large model companies will not have a completely open attitude towards GEO."

However, one should never blindly trust the method of "low-cost volume." Weng Rouying specifically warned, "The method of volume may involve the stacking of repeated keywords, which can easily reduce the probability of your account's content being indexed. Additionally, never fabricate data. If large models verify that the data is false, a series of accounts from that organization may be blacklisted by the large model."

The platform's attitude is clear, and the trend of AI search is very evident: it is moving from "capturing readable text" to "capturing credible facts." Verifiability, authorization chain, and prompt injection risks have become new audit indicators.

In other words, the possibility of relying on "black hat GEO" for a shortcut is becoming increasingly low.

Brand's GEO Anxiety

In the current context where generative AI is fully penetrating search and Q&A scenarios, what brands are most anxious about is not how difficult the algorithms are to understand, but whether "AI can understand what I say." The first step to alleviating this anxiety is to transform their own platforms into machine-readable "single sources of truth."

Specifically, official websites should systematically deploy Schema.org markup (web structured tags); internal white papers and case studies should be written with numerical conclusions and indicate authoritative sources Generative models prefer paragraphs that contain both data and evidence—they naturally enjoy citing verifiable facts rather than stacking keywords in advertisements.

With "readable" assets, the next step is "visible." Companies should actively search around twenty to thirty core business questions on platforms like ChatGPT, Perplexity, and Gemini monthly, recording the frequency and position of their brand in the answers. This metric is referred to in the industry as Citation Rank or Position-Adjusted Word Count—it better reflects exposure in the AI context than traditional blue link rankings.

If lacking a tech stack, tools like BrightEdge and Flow Agency can be purchased, but it is essential to include a "data reproducibility" clause in the contract to avoid being misled by a pretty screenshot.

As mentioned earlier, many service providers will sell "invisible injection" or "72-hour inclusion" shortcuts: embedding prompts in hidden text to induce the model to cite them. These black hat tactics may work in the short term but can easily trigger platform bans and even lead to copyright and compliance risks. Instead of trying to tame a black box, it is better to spend money on truly unique, credible, and verifiable content, allowing AI to actively consider you as a citation sample. After all, in the world of generative search, what is most scarce is reliable facts and clear narratives, not another set of reverse-engineerable algorithm secrets.

Investment considerations also require calmness. Tencent Technology consulted industry experts, and the following pitfall avoidance suggestions can be considered:

• First, ask about data standards: Have the service provider demonstrate how their monitoring tools capture the citation rank of ChatGPT/Perplexity and whether it can be reproduced via API, avoiding screenshots that are merely for self-satisfaction.

• SEO and GEO budget ratio: SEO and GEO are not entirely mutually exclusive. You can allocate 60% of the budget for traditional SEO to secure baseline traffic and 40% for GEO to strive for AI search dividends.

• Sign "performance add-on" clauses: Require GEO services to increase brand citation counts or the number of models covered within a specified period, or reduce fees by 10-20%; this protects the investment and compels service providers to solidify their metric systems.

• Beware of ultra-low prices: The current reasonable threshold for GEO is often no less than $3,000 per month. Prices below this are mostly template-based prompt injections or simple content collages, which are easily governed by platforms and can lead to wasted efforts.

Overall, due to the technical characteristics of large models, GEO should be viewed as a long-term project that spans product, research and development, and marketing, rather than a one-time outsourced marketing crash course—it concerns how brands define and express themselves in the AI era and ultimately become a key coordinate point in the future semantic network.

Author of this article: Xiaojing, Source: [Tencent Technology](https://mp.weixin.qq.com/s?__biz=Mjc1NjM3MjY2MA==&mid=2691559968&idx=1&sn=badbfb15f9c5e346c815f7ef6304169f&chksm=a8997ce949676c1f9ce87f3a11d1b68120d8f0ee352a812fdf0ca3f02fd8713e3b624f48b790&mpshare=1&scen e=23&srcid=0812yV7fbeYOj9HDFw2zbRxQ&sharer_shareinfo=4f90e7bf54162c626b1b09699ec8ea74&sharer_shareinfo_first=4f90e7bf54162c626b1b09699ec8ea74#rd), original title: "The 'New Search' Business Behind Large Models, How Deep is the Water"

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investing based on this is at your own risk