Case Study 01
AI-Powered SEO Content Production System
Three-Stage Workflow from Research Thread to Deployable Article
The Problem
Manual SERP analysis for a single keyword meant visiting each qualifying page by hand, stripping out directories and irrelevant listings by eye, copying and pasting every heading and body section into a separate document, counting images manually, cataloguing heading patterns by hand, and running word cloud tools to identify semantic term frequency across multiple pages. On a conservative estimate, a rough outline took hours. Anything thorough enough to actually write from took days. Then came the writing. Drafting, editing, and refining a single article took additional days. The only automated step in the entire process was word count. Scaling this was essentially unthinkable.
The System
Three stages, each using AI differently.
Stage 01
Research Thread
A conversational AI session used to map the entire domain. Specifically, how SERP analysis tools like SurferSEO and WriteWave evaluate search rankings, what factors they weight most heavily, and why. The AI was used as research infrastructure, not as a writing shortcut. Output: synthesized domain intelligence that would have taken days to compile manually.
Stage 02
Prompt Generation
AI directed to encode the synthesized intelligence from Stage 1 into a reusable, production-ready prompt. Rather than typing constraints from memory (which would have been slow, incomplete, and dependent on recall), AI researched those constraints and wrote the prompt itself. Output: a structured prompt functioning as a repeatable methodology.
Stage 03
Deployment and Benchmarking
The prompt was deployed on a real keyword, dry needling therapy Houston. The system autonomously searched the SERP, applied filtering rules to exclude directories and irrelevant listings, and analyzed eight qualifying organic pages. Output included a full optimization blueprint, three title tag options, three meta description options, URL slug and schema recommendations, and a fully optimized ~2,500-word article with FAQ content, image placeholders, internal and external link suggestions, and copy-paste-ready schema markup. The output was then benchmarked across multiple AI platforms (the same prompt run through Claude and Grok, alongside Grok without the prompt and a third-party writing tool) to evaluate quality differences and identify gaps that warranted further iteration.
Key Architectural Decision
The research thread came first by design, not by habit. Before writing a single instruction, the AI was directed to do something specific. Investigate what tools like SurferSEO and WriteWave were actually measuring, identify the factors they weighted most heavily, and compile everything into a constraint set to feed directly back into a prompt. Deep working knowledge of SEO was already present. That wasn't the gap. The gap was that spelling out every relevant factor manually, from memory, would have been slow, incomplete, and dependent on what happened to be recalled on a given day. Having AI research and generate those constraints meant the output would be more thorough than anything produced manually, reviewable for accuracy rather than constructed from scratch.
There was a second reason, one that mattered more in the long run. Because Thomas had enough SEO experience to evaluate the output, the research could be judged as trustworthy. But the same process works in fields with no prior experience at all, as long as there are reliable benchmarks to anchor the research. Established tools, recognised industry standards, reputable services. Any credible reference point AI can investigate and draw from. The pattern stays the same. Find the benchmarks, extract the critical factors, encode them into a prompt, generate the output, evaluate against what is known or verifiable. The methodology doesn't require expertise in the subject. It requires knowing how to find and use the right reference points, then directing AI accordingly. The research thread wasn't the warm-up. It was the foundation the prompt was built on.
The Retrospective
Post-Build Reflection
After deployment, a design flaw surfaced in the location placeholder scaling system. The system used placeholders with a find-and-replace kit to deploy the same article across multiple locations. The conclusion on reflection was straightforward. Sound concept in other contexts, wrong for SEO specifically. Running the full prompt per keyword produces better-tailored output than find-and-replace location swapping, because each keyword deserves its own fresh SERP analysis rather than a rehashed article built for a different location.
The cross-platform benchmarking surfaced a second, more nuanced gap. Comparative scoring revealed that one platform's output outperformed on local signal specificity. It had surfaced a genuinely local environmental detail that read as assumed knowledge rather than templated content (the kind of observation that signals real-world familiarity with a place). The prompt was producing solid local context but nothing with that level of specific grounding. The fix was precise and constrained. A lateral context research step was inserted into the workflow, directing the system to surface local environmental, demographic, or cultural factors relevant to the service before writing began. Hard limits were built in. Findings integrated as one sentence or clause each, placed where they arise organically, never as dedicated sections, never forced. The goal was a touch of genuine local intelligence, not breadth.
Both retrospectives are preserved deliberately. They demonstrate that the system was evaluated critically after build, not just shipped.
The So What
Any workflow that requires synthesising existing research into a structured, repeatable output faces the same bottleneck. Too manual to scale, too dependent on individual effort and recall to be reliable. The same architectural approach applies to a cancer researcher mapping existing trials to identify untested combinations and scope future proposals. It applies to a marketing team reverse-engineering high-performing ad copy across emotional, logical, and structural categories to generate testable headline variations systematically. It applies anywhere the raw material is research, the goal is a structured output, and the current process is a human doing it slowly by hand.
The domain changes. The framework doesn't. And because the research phase is handled by AI working from credible benchmarks, deep domain expertise isn't a prerequisite. That transferability is the point. Build it once, deploy it in any domain where research feeds a structured output.
Without a computer science background, he reverse-engineered a production-grade AI workflow, identified what the system needed to do, directed AI to build it, stress-tested it against real data, found two design flaws and fixed both, and handed off something a non-technical user could run in five minutes.
Live Demo
Try the system with your own keyword.
Enter any keyword and the CS1 methodology runs live — SEO analysis, optimization blueprint, title and meta options, and a full ~2,500-word article with FAQ content and schema markup.
Open the demo →