Category: The Lab

TL;DR: Keyword research misses semantic topics that AI systems naturally cite. Embedding-Guided Expansion uses neural embeddings to discover these gaps—topics semantically adjacent to your content that keyword tools can’t find. By analyzing the “gravitational pull” of your core content in latent semantic space, you find 5-10 new topics per core article. These topics compound: each new article attracts 3-5x more AI citations than traditional keyword research would suggest.

The Keyword Research Blind Spot

Traditional keyword research is about volume and intent. You find keywords humans search for (search volume) and infer user intent (commercial, informational, navigational).

This works for traditional SEO. It fails for AI citations.

Here’s why: AI systems don’t synthesize responses around keyword clusters. They synthesize around semantic concepts. When an AI generates an answer, it’s pulling from a latent semantic space where topics cluster by meaning, not keyword volume.

Example: Keyword research for “data warehouse” finds:

• Data warehouse (120K searches/month)
• Snowflake data warehouse (45K)
• Redshift vs Snowflake (8K)
• How to build a data warehouse (15K)
• Cloud data warehouse (22K)

You write articles for these keywords. Reasonable. Traditional SEO plays.

But keyword research misses:

• Data mesh (semantic neighbor: distributed data architecture)
• Lakehouse architecture (semantic neighbor: hybrid storage)
• Data governance patterns (semantic neighbor: data quality, compliance)
• Streaming analytics (semantic neighbor: real-time data)
• dbt and data transformation (semantic neighbor: ELT, data preparation)

These aren’t keywords humans search for at scale (lower volume). But AI systems treat them as semantic neighbors to “data warehouse.” When an AI generates a comprehensive answer about modern data architecture, it pulls from all six topics. You wrote content for only three.

Result: Competitors with content on data mesh, lakehouse, and dbt get cited. You get cited partially. You’re incomplete.

Embedding-Guided Expansion: The Method

Instead of keyword research, use semantic expansion. Here’s the process:

Step 1: Compress Your Core Content

Take your best, most-cited article. Compress it into 1-2 paragraphs that capture the essence. Example:

Core article: “Modern Data Warehouses: Architecture, Cost, and ROI”
Compression: “Modern cloud data warehouses (Snowflake, BigQuery, Redshift) replace on-premise systems. They cost $50-200K/month but reduce analytics latency from weeks to minutes. Typical ROI timeline is 18 months.”

Step 2: Generate Embeddings

Use a text embedding model (OpenAI’s text-embedding-3-large, Cohere, or Anthropic’s Claude) to vectorize your compressed content. This creates a mathematical representation of your core topic in latent semantic space.

Step 3: Discover Semantic Neighbors

Generate embeddings for adjacent topics. Find topics whose embeddings are closest to your core content’s embedding. These are semantic neighbors—topics that naturally cluster with yours in latent space.

Example topics to embed and compare:

• Data mesh
• Lakehouse architecture
• Data governance
• Real-time analytics
• Data lineage
• ETL vs ELT
• Data quality frameworks
• Analytics engineering
• dbt and transformation
• Cloud cost optimization

Embeddings reveal which topics are semantically closest (highest cosine similarity) to your core content.

Step 4: Rank by Semantic Distance + Citation Potential

Not all semantic neighbors are worth content. Rank them by:

• Semantic distance (how close to your core content)
• Citation frequency (do AI systems cite content on this topic?)
• Competitive density (how many competitors already have good content?)
• Audience fit (does this topic align with your user base?)

Example: “Data mesh” has high semantic distance, high citation frequency, moderate competitive density, and strong audience fit. Worth writing. “Blockchain for data warehousing” has low semantic distance, low citation frequency, low density. Skip it.

Step 5: Map Content Clusters

Group your discovered topics into clusters. Example cluster around “data warehouse”:

Cluster 1 (Architecture): Lakehouse, data mesh, streaming analytics
Cluster 2 (Implementation): dbt, data transformation, ELT vs ETL
Cluster 3 (Operations): Data governance, data quality, data lineage
Cluster 4 (Economics): Cost optimization, pricing models, ROI

Now you have a content map. Not based on keyword volume. Based on semantic relatedness and citation potential.

Step 6: Build Content Systematically

Write articles for each cluster. Link them internally. The cluster becomes a web of lore around your core topic. AI systems recognize this as comprehensive, authoritative coverage. Citations compound across the cluster.

Why Embeddings Find What Keywords Miss

Keywords are explicit. “Data warehouse” = human searches for that string. Search volume is measurable.

Semantic relationships are implicit. “Data mesh” and “data warehouse” don’t share keywords, but they’re semantically related (both about data architecture). Embedding models understand this. Keyword tools don’t.

When an AI system writes a comprehensive answer about data platforms, it’s pulling from semantic space. If you have content on warehouse, mesh, lakehouse, governance, and transformation, you’re represented comprehensively. If you only have content on warehouse (keyword-driven), you’re partially represented.

Embedding-Guided Expansion fills those gaps systematically.

Real Example: Analytics Platform Company

Before Embedding Expansion:

Company created content for top 10 keywords: data warehouse (yes), Snowflake (yes), cloud analytics (yes), BI tools (yes), etc. Total: 10 articles.

AI citation analysis (via Living Monitor): 240 citations/month. Competitors getting 800-1200.

Embedding Expansion Applied:

Team embedded their core “data warehouse” article. Discovered semantic neighbors:

1. Data mesh (similarity: 0.84)
2. Lakehouse architecture (0.81)
3. Data governance (0.79)
4. Real-time analytics (0.76)
5. dbt and transformation (0.74)
6. Data lineage (0.71)
7. Analytics engineering (0.68)
8. Cost optimization (0.65)
9. Streaming platforms (0.62)
10. Data quality frameworks (0.60)

They wrote 8 new articles (skipped 2 due to low priority).

After 3 months:

Total citations: 1,200/month (5x increase). Why the compound effect?

1. Each new article got cited 40-80 times/month individually.
2. The cluster (original article + 8 new ones) got cited more frequently because AI systems recognize comprehensive coverage.
3. Internal linking amplified citation frequency (when cited, the entire cluster gets pulled in).

After 6 months:

Citations plateaued at 2,800/month. They discovered a second layer of semantic neighbors and started a second cluster around “data transformation.” Repeat the process.

The Recursive Process

Embedding Expansion is not one-time. It’s a system:

1. Create article cluster (10-15 related pieces)
2. Monitor citations for 60 days
3. Analyze which articles get cited most
4. Re-embed the highest-citation articles
5. Discover a new layer of semantic neighbors
6. Create a second cluster
7. Repeat

This recursive process compounds. After 6-12 months, you’ve built a semantic web of 50+ articles, all discovered through embeddings, not keyword research. Your citation frequency is 5-10x higher than keyword-driven competitors.

Technical Implementation

Option 1: In-House

Use OpenAI’s text-embedding API or open-source models (all-MiniLM-L6-v2). Cost: $0.02 per 1M tokens. Build a Python script that:

1. Embeds your content
2. Embeds candidate topics
3. Calculates cosine similarity
4. Ranks by similarity + other factors
5. Outputs ranked topic list

Timeline: 2-3 days to MVP.

Option 2: Use Existing Tools

Some content intelligence platforms offer semantic topic discovery (e.g., Semrush, MarketMuse). They’re not perfect (their algorithms aren’t transparent), but they’re faster than building in-house.

Option 3: Manual Process

If you understand your domain well, list 20-30 candidate topics manually. Re-read your core articles. Which topics naturally appear in them? Those are semantic neighbors. Rank by citation frequency (use Living Monitor).

Why This Works for AI Systems

AI systems are trained on web-scale data. They learn semantic relationships between topics automatically. When they generate responses, they navigate latent semantic space.

If your content is comprehensive within that semantic space, you win. If you’re missing semantic neighbors, you lose—even if you rank well for keywords.

Embedding-Guided Expansion is how you ensure comprehensive semantic coverage. It’s how you become the canonical source across an entire topic domain, not just one keyword.

Next Steps

1. Pick your strongest article (highest traffic, highest citations via Living Monitor).
2. Compress it into 1-2 paragraphs.
3. Embed it. Embed 20 candidate topics. Calculate similarity.
4. Rank by similarity + citation potential.
5. Write articles for the top 8-10 semantic neighbors.
6. Monitor citations for 60 days.
7. Repeat the process for your next cluster.

Read the full guide for the complete framework. Then start embedding. The semantic gaps in your content are worth 5-10x more citations than keyword research would ever find.

{
“@context”: “https://schema.org”,
“@type”: “Article”,
“headline”: “Embedding-Guided Content Expansion: How Neural Networks Find Topics Your Keyword Research Misses”,
“description”: “Use semantic embeddings to discover topics adjacent to your content that keyword research can’t find. Build comprehensive semantic coverage and compound A”,
“datePublished”: “2026-03-30”,
“dateModified”: “2026-04-03”,
“author”: {
“@type”: “Person”,
“name”: “Will Tygart”,
“url”: “https://tygartmedia.com/about”
},
“publisher”: {
“@type”: “Organization”,
“name”: “Tygart Media”,
“url”: “https://tygartmedia.com”,
“logo”: {
“@type”: “ImageObject”,
“url”: “https://tygartmedia.com/wp-content/uploads/tygart-media-logo.png”
}
},
“mainEntityOfPage”: {
“@type”: “WebPage”,
“@id”: “https://tygartmedia.com/embedding-guided-content-expansion-how-neural-networks-find-topics-your-keyword-research-misses/”
}
}

TL;DR: AI isn’t replacing writers—it’s augmenting them. The Ghost Writer Protocol is about using AI as a collaborative muse, not a content factory. The key: humans provide the soul (voice, intention, judgment), machines provide the stamina (research, structure, iteration). Best results come when you stop treating AI as a writer and start treating it as a very smart research assistant who can also edit.

The False Choice: AI vs. Authenticity

The question every writer asks when they first encounter AI for creative work: “Won’t using AI dilute my voice?”

It’s the wrong question. The real question is: “How do I use AI to amplify my voice?”

I spent the first few months of working with AI on creative projects terrified of this exact thing. I’d built a particular voice over years—direct, densely researched, willing to go against consensus. Would giving AI a role in my workflow hollow that out?

The answer was no. The opposite happened. Integrating AI into my writing process made my voice stronger, not weaker. Here’s why, and how to make it work for your writing.

The Three Phases of AI-Assisted Writing

Phase 1: Ideation and Research Scaffolding

This is where AI is most valuable and least threatening to your voice. You’re not asking AI to write. You’re asking it to think alongside you.

I start every article with a research phase. Rather than manually searching and reading, I use AI to:

• Map the landscape of existing ideas on the topic
• Identify gaps and contradictions in conventional wisdom
• Generate research questions I hadn’t considered
• Organize information into a knowledge structure
• Play devil’s advocate against my assumptions

The output isn’t content. It’s scaffolding. It’s the thinking work that usually takes 40% of my writing time. By offloading this to AI, I have more mental energy for the thing only I can do: deciding what’s actually true, what matters, and why.

Phase 2: Structural Outlining

Once I know what I want to say, I give AI a constraint: “Here’s my thesis. Here’s my voice guidelines. Here’s what I want readers to feel. Generate 5 different structural approaches.”

I don’t use any of them as-is. But seeing the options forces me to articulate my own structural intuition. “No, this works better. This section should move here. This argument lands harder if we front-load it.”

This is where the Exit Schema concept becomes crucial. The constraints (your voice, your thesis, your intended outcome) are what make the AI’s structural suggestions valuable.

Phase 3: First Draft Writing and Iteration

Here’s where most people use AI wrong. They ask it to write the article. Then they edit it. Then it still sounds like AI.

Instead: you write the opening. You set the tone. You make the first argument. Then you bring AI in to extend your voice, not replace it.

In practice, this looks like:

• You write the opening 300 words in your voice
• You give AI those words as a context sample and say: “Continue this. Maintain this voice.”
• You edit what it produces, fixing anything that drifts from your tone
• You write the next key argument or transition yourself
• You loop back to AI for sections that are more research-heavy or require more scaffolding

This isn’t laziness. It’s collaborative intelligence. The sections you write contain your authentic voice. The sections AI generates (always guided by your voice samples) fill in the research-heavy connective tissue. Readers experience the whole thing as authentically yours—because the critical thinking and voice are authentically yours.

Maintaining Authentic Voice: Technical and Philosophical

The technical side: feed AI examples of your writing at the beginning of every creative session. Not just instructions about your voice—actual paragraphs you’ve written. Show it the sentence length you prefer, the vocabulary, the cadence, the way you structure an argument.

The philosophical side is more important: own your judgments. AI can help with research, structure, and execution. But the thing that makes the work authentically yours is your judgment about what’s true, what matters, and what’s worth saying.

When I use AI in my writing process, I’m making more conscious decisions about these things, not fewer. I’m delegating the stamina work so I can focus on the thinking work.

The Prosthetic Muse Concept

Here’s the mental model that changed how I think about this: treat AI as a prosthetic muse.

A prosthetic isn’t a replacement for a limb. It’s an amplification. It extends your capability. It lets you do things you couldn’t do before, but in a way that’s still authentically you using it.

AI is the same. It’s not trying to be the writer. It’s trying to be the part of you that can:

• Research 10 sources simultaneously while you think about the argument
• Generate 20 opening sentences so you can pick the one that lands
• Maintain paragraph continuity while you focus on logical flow
• Catch inconsistencies and tighten prose while you focus on ideas

These aren’t the things that make writing authentically yours. They’re the infrastructure. The voice, the judgment, the intention—that’s all you.

The Mistake Everyone Makes

Most people use AI as a content factory. They give it a prompt and hope it produces something publishable with minimal editing. This approach:

• Produces generic, AI-sounding content
• Requires massive editing to make it authentic
• Dilutes your voice rather than amplifying it
• Wastes the actual advantage AI provides

Instead, use AI as a research partner and structural collaborator. Your voice should be the dominant signal in every piece you publish. AI should be invisible except for the efficiency gains you gain from it.

When someone reads your work, they should think: “This person thinks deeply about this topic and writes beautifully.” They shouldn’t think: “Oh, this is AI-assisted.” And they won’t—because the voice is authentically yours.

Building Your Ghost Writer Protocol

Here’s how to implement this in your own writing:

1. Define your voice guidelines: Write 3-4 paragraphs that are peak-you. Give these to AI as reference every single time.
2. Map your writing process: Where do you spend the most time? (Usually research and iteration.) That’s where AI adds the most value.
3. Set structural constraints: Define the format, the sections, the flow before you start writing. This is your Exit Schema.
4. Write the critical sections yourself: Openings, theses, key arguments, conclusions. Your voice in these sections sets the tone for the whole piece.
5. Collaborate on the rest: Use AI to extend your voice, fill research gaps, maintain structure. But curate ruthlessly.
6. Edit for voice authenticity: Your final pass should be about ensuring the whole piece sounds like you, not about fixing AI mistakes.

This protocol transforms AI from a threat to your authenticity into a tool that amplifies it. You’re not losing your voice. You’re delegating the grunt work so you can focus on the thinking and judgment that actually makes your voice valuable.

And the work gets better. Not in spite of using AI. Because of it.

{
“@context”: “https://schema.org”,
“@type”: “Article”,
“headline”: “The Ghost Writer Protocol: How to Use AI as a Creative Partner Without Losing Your Voice”,
“description”: “AI isn’t replacing writers—it’s augmenting them. The Ghost Writer Protocol shows how to use AI as a collaborative muse: humans provide the soul (voi”,
“datePublished”: “2026-03-30”,
“dateModified”: “2026-04-03”,
“author”: {
“@type”: “Person”,
“name”: “Will Tygart”,
“url”: “https://tygartmedia.com/about”
},
“publisher”: {
“@type”: “Organization”,
“name”: “Tygart Media”,
“url”: “https://tygartmedia.com”,
“logo”: {
“@type”: “ImageObject”,
“url”: “https://tygartmedia.com/wp-content/uploads/tygart-media-logo.png”
}
},
“mainEntityOfPage”: {
“@type”: “WebPage”,
“@id”: “https://tygartmedia.com/the-ghost-writer-protocol-how-to-use-ai-as-a-creative-partner-without-losing-your-voice/”
}
}

TL;DR: AI tool outages, rate limits, and billing walls are a weekly reality in 2026. The professionals who maintain “airplane projects” — offline-capable, deep-work tasks ready to deploy the instant cloud tools fail — never lose a productive hour. The ones who don’t lose 2-4 hours doomscrolling and refreshing status pages.

The Fragility Problem

If you’ve built your workflow around Claude, ChatGPT, Gemini, Midjourney, or Cursor, you’ve experienced it: the 2 PM outage that kills your afternoon. The billing wall that hits mid-project. The DDoS event that takes down an entire provider for 3 hours. The API rate limit that throttles your automation pipeline to zero.

In 2025-2026, AI tool fragility isn’t an exception — it’s a structural feature. Every major AI provider has experienced multi-hour outages. Rate limits are tightening as demand outpaces capacity. And the more deeply you integrate AI into your workflow, the more catastrophic each outage becomes.

The Airplane Projects framework treats this fragility as a routing problem, not a crisis. When your primary AI tools go down, you don’t stop working. You switch tracks to a pre-loaded, offline-capable task — the same way you’d shift to deep work on an airplane where you never expected internet access in the first place.

The Framework

An Airplane Project has three qualities: it requires zero internet connectivity, it advances a meaningful business objective, and it can be picked up and put down in 2-12 hour blocks without significant context-switching cost.

For content professionals and agency operators, the strongest Airplane Projects are:

Offline writing and editing. Pre-download your research materials, briefs, and reference documents. When AI tools go dark, open Obsidian, Typora, or iA Writer and draft the pieces that require human judgment — opinion articles, case study narratives, strategy memos. These are the pieces that AI assists but shouldn’t author, and they benefit from the enforced deep focus that an offline environment creates.

Local AI experimentation. Ollama and LM Studio run language models entirely on your machine. When cloud APIs fail, your local models keep running. Use downtime to test prompts, fine-tune local models on your content style, or build automation scripts that will accelerate your workflow when the cloud comes back. We’ve built entire agent armies using Ollama during cloud outages that later became production tools.

Code and automation work. VS Code works offline. Python works offline. Your WordPress REST API scripts, data processing pipelines, and automation tools can all be written, tested (against local mocks), and refined without any cloud dependency. An afternoon of offline coding often produces cleaner code than a connected session because there’s no temptation to ask the AI to write it for you.

Strategic planning and architecture. The best system designs happen on paper or in Excalidraw (which runs locally). When your AI tools go down, pull out your notebook or whiteboard and design the architecture for your next project. Our Site Factory architecture was sketched during a 4-hour Claude outage. The enforced disconnection from execution let us think structurally instead of reactively.

The Implementation

Maintaining Airplane Projects isn’t a habit — it’s a system. Every Friday, spend 15 minutes on three preparation steps.

Pre-download. Save any research materials, PDFs, documentation, or reference content you might need for your current projects to a local folder. If you’re mid-project on content for a client, download their brand guidelines, competitor analyses, and any data files to your machine.

Queue offline tasks. Identify 1-2 tasks from your project list that can be completed without internet. Write them on a physical sticky note or in a local text file. These are your runway tasks — ready for immediate takeoff when the cloud goes dark.

Test your local tools. Verify that Ollama is running and your preferred local model is downloaded. Open your offline writing app and confirm your files are synced locally. Check that your code editor has the extensions and dependencies it needs without fetching from the internet.

The Psychological Advantage

The real value of Airplane Projects isn’t productivity during outages — it’s the elimination of anxiety about outages. When you know you have 8 hours of meaningful work queued that requires zero cloud dependency, an AI outage notification goes from “my afternoon is ruined” to “I’ll switch to my offline queue.”

This is the same psychological principle behind the Expert-in-the-Loop architecture: building systems that gracefully degrade rather than catastrophically fail. Your personal productivity stack should be just as resilient as your enterprise AI infrastructure.

Keep 1-2 airplane projects in your back pocket at all times. When the cloud goes dark, you don’t stop working. You just change altitude.

{
“@context”: “https://schema.org”,
“@type”: “Article”,
“headline”: “Airplane Projects: The Productivity Framework for When Your AI Tools Go Down”,
“description”: “AI tool outages are a weekly reality in 2026. The Airplane Projects framework keeps 1-2 offline-capable deep-work tasks ready so you never lose a productive hou”,
“datePublished”: “2026-03-30”,
“dateModified”: “2026-04-03”,
“author”: {
“@type”: “Person”,
“name”: “Will Tygart”,
“url”: “https://tygartmedia.com/about”
},
“publisher”: {
“@type”: “Organization”,
“name”: “Tygart Media”,
“url”: “https://tygartmedia.com”,
“logo”: {
“@type”: “ImageObject”,
“url”: “https://tygartmedia.com/wp-content/uploads/tygart-media-logo.png”
}
},
“mainEntityOfPage”: {
“@type”: “WebPage”,
“@id”: “https://tygartmedia.com/airplane-projects-the-productivity-framework-for-when-your-ai-tools-go-down/”
}
}