Baotou Marathon

I ran the Baotou Marathon this past Sunday—4 hours and 32 minutes total.

It was my first full marathon in an urban area at subplateau elevation (~1,050 meters). I misjudged the pace: after 32 km, my energy collapsed completely. The final 10 km were walked—fast, but walked.

For the first 32 km, I held a steady pace of ~5:20/km. But I underestimated how quickly glycogen depletes at that altitude—even though it’s not “high” by mountain standards, the body doesn’t adapt instantly without prior exposure. For safety, I downshifted to power-walking the rest.

Still, I finished—safely, and with a new kind of respect for terrain-aware pacing.

A different kind of awe came later that afternoon, driving to the “First Village on the Yellow River.” Standing beside the river—watching its turbid flow, feeling its scale—was visceral in a way no photo or video conveys.

I asked an AI to draft a poem about it. Tinkered with the lines a bit—and liked the result:

Sweat still damp from the full marathon,
Dark clouds hang low over barren hills beside the river.
My body loosens with the muddy current,
Its voice carried by wind straight to my heart.

dd03ffa7d0bd177547907bca7a110e49.jpg 6533fcf7c1502c278f30132de74ce21f.jpg

GEO Keynote at the 4th Digital Intelligence Summit

At the 4th Digital Intelligence Summit, I delivered a keynote on GEO (Generative Engine Optimization). Here are the core takeaways:

  1. AI is now the mainstream information gateway: Over 600 million users rely on AI to search, understand problems, and compare companies. Businesses must adapt—not just to being found, but to being understood and chosen in AI-native workflows.
  2. The age of Agent-assisted decisions has arrived: Consumer decision logic has shifted four times: from functional → preference-based → value-driven → and now Agent-mediated. All four coexist—but Agents increasingly handle information synthesis, comparison, and action initiation. Companies must ensure their product specs, brand narratives, and value claims all feed into this new loop.
  3. How Agent decisions actually work: Humans define goals, constraints, budgets, and preferences; AI interprets intent and surfaces options; Agents organize alternatives, evaluate trade-offs, and—within authorized scope—trigger transactions or service connections. To succeed, businesses need clear reasons to recommend, verifiable facts, and executable entry points—that’s GEO in practice.
  4. Four choke points in AI’s choice path: AI must (a) understand user intent, (b) retrieve multi-source data, (c) assess credibility, and (d) generate answers + drive next steps. At each stage, businesses risk being invisible, incomprehensible, or untrustworthy. GEO means auditing your presence across all four.
  5. User queries are getting hyper-specific: People ask about location, family size, budget, delivery windows, and product specs—all in one prompt. Procurement teams add lead time, certifications, system integration, capacity, and fulfillment risk. Businesses must structure facts around these concrete conditions so AI can answer: Who is this for? What does it satisfy? Why should anyone believe it?
  6. GEO’s essence: Turning real-world value—products, tech, services—into digital evidence: discoverable, understandable, trustworthy, and actionable. That conversion isn’t one-time. It must be continuous, so AI can reliably surface your facts across diverse questions and contexts—and embed them in users’ cognition and choices.
  7. The three gates to AI visibility:
    • “Seen”: Information must be findable and accessible.
    • “Understood”: Clear subject, complete facts, consistent expression.
    • “Trusted”: Reliable sources, verifiable evidence, timely updates.
      All three must align—or your data won’t make it into AI-generated answers.
  8. Who benefits most from GEO? Organizations where decisions carry high stakes, involve long evaluation cycles, vary widely across use cases, and rest on rich factual foundations. B2B: supports supplier vetting, solution comparison, and case validation—generating qualified leads. B2C: enables brand comparisons, risk assessment, and personalized recommendations—driving consultation, store visits, and conversion.
  9. Tackling hallucination & quality control: One proven approach is a multi-layer AI QA architecture. In GEOFlow, for example, you can route content through atomic fact checks, knowledge-slice validation, or full knowledge-base audits—then apply auto-approval, human review, or hard blocking based on task sensitivity, risk, and cost.
  10. The GEO feedback loop: Build a self-correcting growth cycle—starting from real user questions → establish a baseline of known gaps → fill missing facts, sources, evidence, and ownership → distribute across channels → monitor performance → feed insights back into the next round. Each iteration accumulates reusable, business-grade assets: questions → facts → sources → answers → leads.
  11. Measuring GEO impact: Track progression along the evidence-to-revenue chain: Is your company appearing in AI shortlists? Are your sources and facts being cited? Is your brand recommended—and does that trigger action? Finally: do those actions convert to leads, opportunities, and revenue? Establish baselines, measure repeatedly, segment by platform and scenario, and close attribution loops to quantify real business value.

Why Big Tech Is Entering GEO

Lately, Tencent launched Answerbit, Alibaba upgraded Wanxiao Zhi 3.0 with GEO capabilities, and Baidu, Meituan, and 360 have all entered the space—only ByteDance remains unannounced.

This isn’t fragmentation—it’s validation. When giants collectively step in, they’re not just building tools—they’re certifying the category.

Even more telling: In May, Xinhua News Agency rolled out the Xinhua GEO Agent Platform—a signal from China’s most authoritative media institution that GEO isn’t hype. It’s necessary infrastructure.

The market will be large—and fiercely pluralistic. Not one winner-takes-all platform, but a thriving ecosystem: data providers, service layers, implementation partners, toolkits, standards bodies, consultants, trusted sources, compute infra, end-to-end solutions, cross-border support, and open-source communities.

Each layer spawns further specialization—by industry, company size, maturity, and use case.

So the strategic question isn’t “Can I compete with Tencent?” It’s: Where do I hold a differentiated or total-cost advantage?

Smaller players, for instance, can go deep where big tech won’t—or can’t: delivering white-glove GEO implementation, vertical-specific evidence curation, or hands-on coaching for mid-market teams. In a blooming garden, there’s room for many strong perennials.

The Most Important Thing

How do you identify the most important thing—right now?

It’s shaped by environment, resources, role, current constraints, and even your vantage point.

And it shifts—constantly.

It’s inherently relative. And deeply subjective.

Which makes acting on it anything but simple.

Two mental habits help:

  1. Zoom out—then zoom back in: Step up at least two levels above your immediate context. See patterns, dependencies, and leverage points. Then return to execution with sharper clarity. This solves how to see.
  2. Anchor daily action: Every day, deliberately do one thing that advances your current priority thread. Not random “important” tasks—just one, aligned with your near-term line of effort. Consistency builds perception, discipline, and momentum. This solves how to act.

Together, they form a rhythm: perspective → intention → action → reflection → repeat.

GPT-6 Astra: Early Takeaways

I reviewed OpenAI’s official GPT-6 Astra documentation and several independent benchmarks—focusing on complex, multi-step task performance. Key observations:

  • Astra sits firmly in the top tier of reasoning models—especially strong in programming-agent workflows.
  • Its standout capability is continuous execution: chaining coding, desktop interaction, and web browsing into unified workflows. Official API supports 1.05M-token context.
  • Progress spans both training (improved pretraining + RLHF) and runtime (better long-context state retention, tool orchestration, and memory management). Translation: stronger problem-solving and better task persistence.
  • Cost calculus is shifting: Astra’s per-token price is ~2.5× GPT-5.6 Sol—but token usage per complex task drops to ~⅓. So while unit cost rose, total task cost may fall for high-value, multi-step jobs.

Bottom line: For production-grade automation—especially involving code, UI interaction, and dynamic research—Astra is worth piloting now.

Website GEO Diagnostic Skill

I built an open-source skill to audit any website for GEO readiness:

How it works:

  1. Starts at the homepage, reads robots.txt, fetches up to 5 sitemaps, and crawls navigation—building a candidate list of ≤500 normalized URLs.
  2. Classifies pages into 10 types: Home, Brand/Org, Products/Services, Categories/Content Hubs, Articles/News, Guides/Docs, Case Studies/Proof, Comparisons/Lists, Pricing/Transactions, FAQ/Support/Contact. For each type, it selects representative pages using fixed weights:
    • Type confidence (35%)
    • Navigation prominence (25%)
    • URL template representativeness (20%)
    • Sitemap coverage (10%)
    • Freshness (10%)
  3. Scores each page across 8 dimensions:
    • Accessibility & Discoverability (15%)
    • Architecture Coverage (10%)
    • Entity Clarity (12%)
    • Answerability (15%)
    • Evidence & Citability (16%)
    • Authority & Trust (12%)
    • Structuredness & Extractability (12%)
    • Freshness (8%)
  4. Runs 19 granular checks per page—recording raw values, thresholds, weights, status, evidence IDs, and remediation steps.
  5. Generates a rich, visual report: crawl access matrix, readiness funnel, page-type coverage, internal-link graph, entity matrix, answer-depth scatter plot, Schema tree diagram, and more.

Get it:

  1. GitHub repo: GEOHub
  2. Sample report: doc.laoyao.cn
  3. Deploy: Tell any AI: “Clone and deploy GEOHub”
  4. Run: After deployment, say: “Run GEO diagnosis for www.geoflow.me”
    Or CLI: geo-seo-hub site-diagnose --url <URL> --output runs

GEOFlow 3.0 Is Live

GEOFlow 3.0 is now publicly deployed, with updated architecture and documentation: