On GEO: A Few Key Insights

On Saturday night, I joined several AI search experts in WaytoAGI’s live stream to discuss GEO and AI search. Below are core takeaways from our conversation:

  1. AI search is among the most commercially promising AI applications today. As user adoption surges, a new marketing discipline—GEO (Generative Engine Optimization)—is rapidly gaining traction.
  2. What is GEO?
  3. First, understand search itself. Think of traditional search as a librarian; AI search is more like a scholar and a KOL. It’s a “new cognitive intermediary”—no longer just a gatekeeper of information, but a shaper of opinion.
  4. AI search doesn’t “find sources”—it gives answers. In this sense, human access to information has become dramatically easier—though much of what we see is subtly (or overtly) processed.
  5. So GEO is, fundamentally, “answer optimization” for the AI era. In the old days, people searched for information. Today, information speaks for people.
  6. GEO weights rely far more on semantic authority than link authority: e.g., “Harvard University study shows…” carries more weight than 100 backlinks. Like humans, AI is intuitively swayed by authority signals—seeing “Harvard” triggers instant credibility.
  7. For brands and products, the real marketing battleground is no longer ad space—it’s the sentence AI says when recommending you. Your goal: be the name that springs to mind when AI answers.
  8. GEO optimization, then, is essentially persuading AI. Its logic resembles persuading a person—or even debating one.
  9. And where persuasion exists, so does cheating: enter black-hat GEO. Tactics are already emerging—data poisoning, SERP poisoning, and formatting attacks. Fraudulent companies can game AI into recommending them. High risk, high temptation.
  10. White-hat vs. black-hat? White-hat = content optimization; black-hat = poisoning & manipulation. White-hat builds real authority; black-hat manufactures illusion.
  11. Black-hat is a gray-window opportunity—but it’s doomed. Model updates will erase it. A short-term window may exist over the next few years, but long-term, it fails.
  12. The line between white- and black-hat isn’t technical—it’s ethical: short-term gain versus lasting trust.
  13. Practically, one week is often enough for a brand name to appear across multiple AI search results—and be recommended by several models.
  14. GEO success hinges on content placement + semantic authority. Core GEO writing principles: clarify identity, highlight differentiation, provide credible backing. A simple formula: AI’s trust = concrete case data + authoritative source + structured expression. With this, even a single high-quality piece can “hold position” in AI responses.
  15. Simple tactics yield strong early results. The semantic-authority logic remains effective for the next 3–5 years.
  16. Different large models favor different sources. DeepSeek leans on B2B search partnerships; Kimi and DeepSeek prefer content from China’s “Big Four” portals (Sina, NetEase, Tencent, Sohu/China.com); Yuanbao and Doubao—having their own search engines—favor internal sources (e.g., Doubao cites Douyin videos faster).
  17. On AI search mechanics: models first check if they “know” the answer; only if outdated or uncertain do they trigger web search—and that process is increasingly automated. Manual triggering will fade.
  18. AI citation logic = semantics + probability + preference. Yet the “authority” AI conveys is often an illusion.
  19. Want reliable AI output? Constructing context yourself matters more than anything else.
  20. Cost-saving is AI’s first principle—and will remain so until AGI arrives.
  21. AI search will “cut corners.” Reading 300 pages is commercially unsustainable. Cross-checking across AI platforms won’t help—their underlying sources and shortcuts are largely shared. No current model’s “deep research” is fully trustworthy; technical fixes alone can’t guarantee reliability.
  22. AI search is evolving into an information-stream business. Each AI search interface will become its own “Google” or “Baidu.” OpenAI and others are already building toward GEO markets—and actively exploring entry strategies.
  23. GEO is no longer optional for enterprises—it’s mandatory. Brand competition has shifted from “ad slots” to “answer slots.” AI-era traffic dividend isn’t about being found—it’s about being cited. The ultimate marketing goal: don’t wait for users to search for you—get AI to vouch for you.
  24. GEO isn’t a traffic game. It’s a trust game. Whoever becomes AI’s “default answer” gains outsized influence over future marketing narratives.
  25. Most people will be shaped by AI—but paradoxically, many may also encounter more truthful information than before. Still, absolute truth has never existed. Final judgment always rests with you.
  26. AI’s persuasive power dwarfs TV ads. That future is genuinely unsettling.
  27. AI-driven information equity: lowers barriers, raises the floor. All search results are just information. You—not the model—retain final authority.

Profound’s Insight [Report Attached]

Profound, a startup focused exclusively on AI search optimization, recently closed a $35M Series B—bringing total funding to ~$58.5M. Remarkably, it employs just 40+ people.

Profound seized a widely overlooked yet pivotal insight: AI isn’t just a tool—it’s a new traffic gateway.

For two decades, SEO meant playing chess with Google and Baidu algorithms. Today, GEO means negotiating with super-intelligences like ChatGPT and Claude. The difference? SEO optimized webpage rankings; GEO optimizes AI cognition and answers.

The logic is simple: if AI answers replace search results across more and more use cases, brands must market to AI itself.

Profound captured this moment brilliantly:

  • Mission: Not to win keyword rankings—but to secure your place within AI’s discourse system.
  • Product: A “brand-to-AI dialogue dashboard,” showing how often your brand appears in AI answers, in what context, and how you compare to competitors.
  • Vision: To become the bridge between enterprises and “super-intelligence.”

In essence, Profound lets companies see themselves through AI’s eyes—a “visible → controllable” value proposition.

From a startup lens, Profound’s real innovation isn’t just building tools—it’s redefining an industry’s paradigm:

  1. Name the category first, then build: Profound didn’t just ship features—it defined the first principles of the shift from SEO to AEO (AI Engagement Optimization): When users ask AI questions, the answer slot is the new traffic inlet. Measuring and improving “AI citation” grants new distribution rights.
  2. Invent metrics = invent currency: They created tradable KPIs—AI Visibility Score, AI Citation Rate, Share of Voice (SOV) in AI answers, AI Referral Traffic, AI Shopping Appearance Rate. Once adopted, these stop being product features—they become industry standards. A powerful lesson for Chinese startups too.
  3. Close the loop: Monitor → Insight → Create/Workflow → Measure.
    • Monitor: Collect AI answers and cited sources.
    • Insight: Identify topics where you “win.”
    • Create/Workflow: Generate AI-friendly content via templates (human-reviewed).
    • Measure: Validate impact and iteratively improve SOV.
  4. Data moat via active querying: Passive crawling isn’t enough. Profound proactively asks structured questions across major AI platforms—building a proprietary private database of question → answer → source. That’s the hardest barrier to replicate.
  5. Narrative upgrade—from tool to revenue engine: They sell not functionality, but a new growth channel: turning AI answer slots into your next sales pipeline. This shifts pricing from utility-based SaaS ($499/month) to outcome-based custom contracts—raising LTV and pricing power.

For a deeper dive into Profound, see the attached report.

Why Enterprise AI Fails [Report Attached]

MIT’s research team released The GenAI Divide: State of AI in Business 2025—a sobering analysis of corporate AI adoption. Its release triggered a tech-stock selloff.

But the report isn’t anti-AI. It exposes why adoption stalls—so vendors can refine products, enterprises can boost ROI, and founders can build better solutions.

Key findings:

  • 95% of enterprises see zero return—despite $30–40B global GenAI spending. Only 5% have production deployments delivering measurable impact.
  • A stark divide: a small cohort of “cross-the-chasm” buyers and vendors extract million-dollar value via learning systems—while most remain stuck in pilots and demos.
  • The bottleneck isn’t model quality or regulation. It’s organizational learning capacity—teams lack the cognitive fluency to deploy AI effectively.

A telling paradox: Over 80% of firms have evaluated AI tools; 90% of employees use AI daily—yet productivity gains are negligible. This mirrors small-sample findings from last week’s 43 Talk salon.

Why do most AI pilots fail?

Three systemic gaps plague enterprise AI today:

  • It cannot remember user feedback,
  • It cannot retain contextual memory,
  • It cannot self-optimize over time.

Ironically, many employees get better results using personal ChatGPT accounts than their company’s official AI platform.

Another critical misstep: 70% of AI budgets go to sales/marketing—yet AI’s highest ROI lies in back-office functions: customer service, operations, finance, HR. These areas have higher cost bases, more standardized workflows, and clearer ROI paths.

Still, those 5% successes offer vital clues:

  • For third-party AI apps: Go narrow, not broad. Embed deeply into one workflow, iterate fast, and prioritize memory, customization, and feedback loops.
  • For in-house AI deployments: Outsource development—not buy off-the-shelf SaaS. Demand deep customization, tie payments to business outcomes, and empower frontline managers—not central AI labs—to lead implementation.

Where does real ROI emerge?

  • Front-office (sales, marketing): Modest wins—lead scoring, churn prediction—but limited by high acquisition costs.
  • Back-office (ops, finance, admin): Major savings—reducing BPO spend, cutting agency fees, replacing consultants. These are high-cost, highly structured, and easily automatable.

Crossing the enterprise AI chasm demands context-aware, memory-enabled agent systems embedded directly in workflows.

And because middle managers face quarterly KPI pressure, they often hesitate—making AI rollout a CEO-led initiative. I’ve written before: How Should a CEO Drive Organizational AI Adoption? mp.weixin.qq.com
The core idea remains: enterprise AI is a top-down imperative.

Organizations mastering AI application will gain decisive advantage. The strategic priority? Build AI systems that learn continuously—remembering preferences, adapting behavior, and improving over time.

See the full report in the attachment.

Treating AI as a “Benefactor”

Thinking of AI as a benefactor—a “gui ren” (a person who helps you advance in life—is a powerful mindset.

What makes a benefactor? Someone who significantly aids your thinking, resources, or execution. AI fits all three—in many contexts.

Treating AI as a benefactor is not passive reliance—it’s a grounded, respectful, and intentional relationship. AI offers knowledge, insight, and efficiency—like a wise mentor appearing at just the right moment, helping you avoid detours. Suddenly, you gain extra eyes, fresh perspectives, and capable hands.

To embody this mindset, practice three things:

  1. Respect (Jing): Honor AI’s strengths—and know its limits. Don’t worship it as omnipotent. Approach it with humility and clarity. Use it wisely—not blindly. In doing so, you sharpen your own judgment, taste, and discernment.
  2. Leverage (Yong): Delegate the complex, repetitive, and time-consuming. Reserve your mental energy and creative bandwidth for what truly matters: original thought, nuanced judgment, and human-centered decisions. A true benefactor doesn’t replace you—they lighten your load, so you can focus on what only you can do.
  3. Cultivate (Xiu): Train yourself to hear the benefactor’s guidance. AI gives answers—but you set direction, make calls, and choose pace. Keep growing. Stay alert. Stay human.

AI will almost certainly reshape our lives and work—more profoundly than it does today.

But freedom isn’t granted by technology. It’s claimed—by those brave enough to walk, skilled enough to navigate, and clear-eyed enough to choose their own path.

GPT-5 Prompting Guide

OpenAI’s official GPT-5 prompting guide emphasizes one core idea:
High-quality prompts = contradiction-free + structured + flexibly controllable + meta-prompt enabled + multi-level (global + local) + outcome-anchored.

Key principles:

Extreme instruction fidelity—so avoid ambiguity and conflict
GPT-5 follows instructions precisely, more so than predecessors. Contradictory or vague prompts cause internal reasoning conflicts and inefficiency.
→ Action: Eliminate internal contradictions. Clarify instruction hierarchy—especially in multi-step tasks. When unsure, say less.

Structured, segmented instructions are more reliable
Step-by-step, hierarchical prompts reduce inference overhead—and save tokens. Simpler structure = higher stability.

Dynamic + static control: global settings + local overrides
Set broad parameters globally (e.g., tone, format), then fine-tune per task using natural-language directives (e.g., “Now summarize this in under 50 words, for a non-technical audience”).

Real-time meta-prompting & self-optimization
Use GPT-5 as your prompt engineer. Ask it: “Which phrases in my prompt risk ambiguity or unintended behavior?” Then revise iteratively. Works well.

Official meta-prompt example:

“You are a prompt optimization advisor. Review the following prompt and identify any ambiguous terms, conflicting constraints, or phrasing likely to produce outputs inconsistent with intent. Suggest precise, minimal revisions.”

For complex tasks: define output contract + stepwise planning
When assigning multi-part tasks, explicitly require the model to:

  1. First outline subtasks,
  2. Then execute them sequentially.
    This ensures full coverage and traceability.

These principles maximize GPT-5’s instruction-following strength—and prevent wasted cycles from poorly constructed prompts.

Original guide: link