Three Generations of GEO Evolution — From Manual Operations to AI Self-Optimization
Have you noticed a pattern: almost every technology goes through a similar evolutionary path —
At first, humans do it manually, then tools help, and finally machines do it themselves.
This is true for e-commerce, marketing, and GEO as well.
Since GEO was first proposed in 2023 (now 2026), it has completed three generations of evolution in just three years.
Which generation is your team currently in?
When do you need to evolve to the next generation?
This article will help you see the full picture at once.
I. The Complete Picture of GEO's Three Generations
| Generation | Period | Name | Core Characteristics | Human Role |
|---|---|---|---|---|
| First generation | 2023-2024 | Manual GEO | Human manual analysis, manual optimization | Executor |
| Second generation | 2024-2025 | AutoGEO | Algorithms auto-extract rules, tools assist | Decision-maker |
| Third generation | 2026+ | AgenticGEO | AI agents autonomous closed loop | Supervisor |
Core trend: Human involvement decreases, AI autonomy increases.
Each generation doesn't "replace" the previous one — it "upgrades" it. First-generation methods still work, they're just no longer efficient enough.
II. First Generation: Manual GEO (2023-2024) — The "Handicraft Workshop" Era
What Happened?
In 2023, a team from Princeton University published the foundational paper in the GEO field (arXiv:2311.09735, KDD'24), proposing 9 strategies that "can influence AI recommendations." At that time, GEO was a brand-new concept — nobody knew whether it could be done or how to do it.
First-generation GEO practitioners were like "pioneers":
- Manually searching industry topics on ChatGPT to see how AI answers
- Manually analyzing who and what AI cited in its responses
- Manually adjusting their content to see if AI "liked" them more the next month
- Then summarizing their experience into "playbooks" to share with the industry
Typical First-Generation Workflow
┌─────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐
│ Manual Search │ → │ Manual Analysis │ → │ Manual Editing │ → │ Manual Verification │
│ (ChatGPT) │ │ (Excel spreadsheet) │ │ (Editing articles) │ │ (Re-querying) │
└─────────┘ └──────────┘ └──────────┘ └──────────┘
The entire process was all hands-on. One person could optimize about 5-10 articles per week and analyze 3-5 topics.
Limitations of the First Generation
- Extremely low efficiency: Human capacity is the bottleneck; scale is impossible
- Inconsistent results: AI responses differ each time, making it hard to judge whether optimization actually worked
- Hard to replicate success: Success depends on "feel" — a different person may not get the same results
- Limited coverage: A team can only cover a limited number of industry topics
But the First Generation Had an Irreplaceable Value
"Tactile intuition."
The first batch of manual GEO practitioners developed an instinct for AI preferences through the "hands-on" process — what content formats AI prefers, what phrasings AI is more willing to cite, which topics' AI answers are most susceptible to influence. This intuition remains valuable as "qualitative judgment" even after automated tools became available.
Should You Still Do GEO Manually?
If a company is just starting GEO now, it's not recommended to begin with the first generation. Jump directly to the second generation, using tools to assist while humans focus on strategy and review.
But there's one exception: if AI answer quality in your industry is poor and information is scarce, first-generation manual analysis still has value — because your industry understanding may be deeper than any tool.
III. Second Generation: AutoGEO (2024-2025) — The "Industrial Automation" Era
The Key Turning Point
Between 2024-2025, the AutoGEO paper from Carnegie Mellon University (CMU) was accepted by ICLR 2026. The paper's core contribution: letting algorithms automatically learn from data "what types of optimization are most effective" instead of having humans guess on their own.
This generation's characteristics are:
- Automated rule extraction: Tools automatically analyze which content dimensions (structure, formatting, credibility signals) have the greatest impact on citations
- Systematic monitoring: Continuously tracking citation changes in AI answers
- Quantifiable optimization suggestions: Tells you how to change and what improvement to expect
- Dramatically improved efficiency: One person can manage hundreds or thousands of topics
Typical Second-Generation Workflow
┌─────────────┐ ┌─────────────┐ ┌──────────────┐
│ Monitoring tool │ → │ Tool analyzes data │ → │ AI generates suggestions │
│ tracks AI citations │ │ extracts rules │ │ Human confirms & executes │
└─────────────┘ └─────────────┘ └──────────────┘
Humans moved to the "confirmation" stage. The tool says "this content needs more tables and authoritative citations" — the human reviews it, agrees it makes sense, and executes.
Key Changes in the Second Generation
- From "experience-driven" to "data-driven": Not "I think AI likes this" but "the data tells me AI likes this"
- From "point optimization" to "batch optimization": Optimization experience from one article can be applied in bulk to similar content
- From "post-hoc verification" to "pre-prediction": Tools can predict how much citation share will increase after a specific content change goes live
New Role Born in the Second Generation: GEO Analyst
The biggest change in this generation was the emergence of the "GEO Analyst" role — not a pure content writer, nor a pure technical engineer, but someone who "understands GEO monitoring data, can interpret tool reports, and can formulate optimization strategies."
Should You Adopt the Second Generation?
For most companies, the second generation is the most realistic choice right now. Because:
- The first generation is already outdated — efficiency is too low
- The third generation is still early — commercial tools aren't mature
- The second generation strikes the right balance between "efficiency" and "controllability"
IV. Third Generation: AgenticGEO (2026+) — The "Autopilot" Era
What Is AgenticGEO?
In March 2026, a team from Beihang University formally proposed the concept of AgenticGEO (Agent-based GEO), which the industry considers GEO's "self-driving" era.
The core of AgenticGEO is a complete "perception-decision-action-learning" closed loop, run fully autonomously by an AI Agent:
- Perception layer: The Agent continuously monitors multiple AI products (ChatGPT, DeepSeek, Doubao, Perplexity, etc.) for their answers about your brand and industry topics, capturing citation data in real-time
- Decision layer: The Agent analyzes current citation share, identifies content gaps, and automatically formulates optimization strategies
- Action layer: The Agent auto-generates content, adds Schema markup, publishes to your website, and distributes to external platforms
- Learning layer: The Agent tracks citation changes after optimization, validates strategy effectiveness, stores experience in a knowledge base, and guides the next round of optimization
The human-Agent relationship: Humans no longer do operations — they only set "boundaries" — setting goals ("increase our citation share to 20%"), setting scope ("optimize only the website and Zhihu"), setting limits ("don't fabricate data, don't plagiarize"), then letting the Agent execute and report regularly.
Key Differences Between the Third Generation and the First Two
| Dimension | 1st Gen (Manual) | 2nd Gen (Tool-Assisted) | 3rd Gen (Agentic) |
|---|---|---|---|
| Who analyzes? | Human | Tool | AI Agent |
| Who decides? | Human | Human | AI Agent |
| Who executes? | Human | Human | AI Agent |
| Who verifies? | Human | Tool + Human | AI Agent |
| Human role | Executor | Decision-maker | Supervisor |
| Topics manageable | 10-50 | 100-1,000 | 1,000+ |
Risks of the Third Generation
AgenticGEO looks ideal, but there are two risks to watch for:
Risk 1: AI polluting AI.
Imagine: Agent A (responsible for optimizing brand content) writes some content, Agent B (the AI platform's RAG system) crawls and cites that content, then Agent A further optimizes based on B's citations — creating a closed loop. If A's content contains a tiny error, that error could be amplified within the loop.
Risk 2: Cascading errors.
One Agent's erroneous decision could trigger a series of incorrect content modifications. For example, the Agent misjudges that "AI prefers table format" (when the real reason might be another content factor), then batch-converts all content to tables, actually lowering citation rates.
The current industry consensus is: the full version of the third generation still needs 12-18 months to mature. Current "AgenticGEO" products are generally "semi-automated" — the Agent handles analysis and suggestions, but critical execution still requires human confirmation.
V. Which Generation Should Your Company Be In Now?
This is a very practical question. My recommendation:
Scenario A: Just starting out, limited resources
- Don't do the first generation
- Start directly with the second generation: use existing free/low-cost GEO monitoring tools
- Humans handle content strategy and review; tools handle data analysis and performance tracking
Scenario B: Some GEO foundation, team of 2-5
- Primarily use the second generation: paid monitoring tools, systematic data operations
- Begin exploring the third generation: trial AgenticGEO tools, explore "semi-automation"
- Goal: "let tools handle 80% of data analysis work for humans"
Scenario C: Large enterprise, thousands of product lines
- The second generation isn't enough — must start planning for the third generation
- Because managing thousands of product pages is beyond what humans and tool assistance can handle
- Introduce AgenticGEO for automatic optimization of "content-intensive topics"
- But maintain human review mechanisms to prevent cascading errors
How to Judge Your Generation
A simple method: how many hours per week do you spend on GEO-related operations?
- > 80% spent on "manually querying AI, manually analyzing, manually editing content" → you're still in the first generation
- > 50% spent on "reading tool reports, adjusting content based on suggestions" → you're already in the second generation
- > 50% spent on "setting strategy and goals, reviewing Agent output" → you've entered the third generation
Understanding GEO's three generations of evolution isn't about chasing the latest trend. Quite the opposite — it's about helping you find the generation that's right for you.
The "tactile intuition" honed in the first generation is precious, but don't rely on it for scaling.
The "efficiency" offered by the second generation is necessary, but don't treat tools as the end goal.
The "future" promised by the third generation is tempting, but don't over-rely on it before it matures.
GEO's evolution won't stop here. The fourth generation — Multi-Agent GEO — may disrupt current thinking again before long. But regardless of how it changes, GEO's core logic remains the same: do content well, do credibility well, do coverage well. These are GEO's "constants." On top of the constants, use ever-evolving tools to improve efficiency.