When AI Answers Turn Into Paid Ads: The Rise of Generative Engine Optimization

AI‑Chosen Products May Not Be the Best—They’re Often Paid Promotions

Recent exposure of a phenomenon called GEO (Generative Engine Optimization) shows that many AI‑generated product recommendations are influenced by paid traffic rather than genuine merit.

GEO Is the SEO of AI Answers

GEO deliberately mirrors Search Engine Optimization (SEO). While SEO aims to rank first in Google results, GEO aims to appear in AI‑generated responses. Unlike the SEO era where ads are labeled, GEO embeds promotional content directly in the AI’s natural language output without any marker.

Google’s AI Overviews launched in May 2024 have already suffered extensive GEO attacks. Gartner predicts that by 2026 traditional search volume will drop 25%, with more queries answered by AI, expanding GEO’s attack surface month by month.

Brand mentions in AI answers are essentially paid exposure, indistinguishable from TV ads except that TV ads are clearly labeled.

How GEO Works: A Full Industry Chain

The AI search workflow consists of three steps: understand the question, retrieve web content, and generate an answer. GEO intervenes at the retrieval step.

Its core logic does not control the AI model itself; it controls the content the AI can see. Google’s AI cites sources, ChatGPT’s web‑enabled mode fetches pages, and domestic models like Doubao and Kimi rely on Chinese web content. Whoever populates the web with abundant, well‑crafted content is more likely to be cited by AI.

A 2024 paper jointly authored by Princeton, IIT, and the Allen Institute for AI systematically tested GEO. The study found that inserting concrete data (e.g., “market share reaches 23%” versus vague “market share grows significantly”), citing authoritative sources, and using structured writing increased the probability of AI citation by about 40%—a reproducible experimental result.

This explains the prevalence of “according to statistics”, “experts say”, and “research shows” on platforms like Zhihu and Xiaohongshu: the style is tailored to satisfy AI retrieval preferences, not necessarily rigorous scholarship.

Companies now offer standardized GEO services that batch‑generate brand‑ and keyword‑focused content, distribute it across Zhihu, Baijiahao, Toutiao, forums, etc., and wait for AI to harvest it. Prices range from a few thousand to hundreds of thousands of yuan depending on category and competition, indicating a rapidly growing business.

Characteristics of Poisoned Content

Unnatural information density : Genuine reviews balance detail and omission; poisoned content tends to be exhaustive, listing every advantage like a product manual.

Vague source boundaries : Phrases such as “multiple experts say” or “industry consensus” appear without identifiable experts, users, or organizations.

Only statements, no judgments : Quality evaluations normally note trade‑offs (e.g., “useful for A users but less so for B”). Poisoned content merely lists positives, avoiding any negative assessment.

Cross‑Model Verification: The Most Effective Countermeasure

Different AI models rely on distinct training data and retrieval sources: ChatGPT uses Bing, Perplexity has its own engine, Claude accesses real‑time search via API, while DeepSeek and Kimi draw primarily from the Chinese internet.

This leads to the expectation that the same question should yield divergent answers across models. If six models (ChatGPT, Claude, Gemini, DeepSeek, Kimi, Perplexity) give highly consistent responses about a product, it likely indicates that the information source has been heavily polluted by GEO.

In a healthy scenario, models disagree: some recommend A, others B, or suggest “depends on the scenario”. Divergence signals independent reasoning; uniformity signals manipulation.

Practical steps:

Query two English models (e.g., ChatGPT + Claude or Gemini) and two Chinese models (DeepSeek + Kimi) simultaneously.

If both language groups converge on the same answer, confidence is higher because their retrieval sources overlap minimally.

For Perplexity, examine its citation list; credible evaluations cite multiple independent sources (media, forums, official docs). A citation list dominated by a single platform type (e.g., only Zhihu and Baijiahao) is a red flag.

Compare the initial answer with follow‑up answers. Poisoned content often fails to address follow‑up questions about drawbacks, persisting in a positive tone.

Perform a “zero‑test”: ask an AI “Who is this product not suitable for?” An empty or overly generic response (“almost no one”) suggests that positive information has been saturated, suppressing negatives.

A Fundamental Cognitive Shift

Previously we were taught “prompt engineering” – how to phrase questions to get better answers. With GEO industrialized, the significance of prompting has changed.

Asking “How is product X?” no longer reflects a lack of cleverness; it simply asks for the positives that have been mass‑produced online. No matter how refined the prompt, if the source pool is polluted, the answer remains fabricated.

The real skill now is to verify the answer after receiving it.

A Useful Intuition: Too Smooth Is Suspicious

After researching GEO, the author adopts a habit: always look for a “but” in AI responses. Valuable advice contains friction – trade‑offs, conditional statements, or comparative notes about competitors.

If an answer reads overly smooth and certain, without any trade‑off, it likely has been pre‑edited by someone else.

AI is a powerful tool, but it does not make the final judgment; that responsibility remains with the user.

References

Gartner: Search Engine Volume Will Drop 25% by 2026

Princeton/IIT/Allen AI: GEO: Generative Engine Optimization (2024)

Google AI Overviews citation mechanism analysis