metehan.ai Research & SOPs

Appearance

AI Prompt Tracking ​

Fresh

This content reflects Metehan Alp's prompt tracking methodology as of 2025-2026.

The biggest gap in AI search optimization is measurement. You can optimize content for AI citations, but how do you know which prompts users are actually typing? This methodology transforms Google Search Console data into actionable prompt intelligence.

Core Problem ​

Most AEO practitioners guess which prompts to track. They write content targeting prompts they think users ask, but have no systematic way to discover what users actually type into ChatGPT, Perplexity, or Google AI Mode.

The problem is structural: AI platforms do not provide query analytics the way Google Search Console does for traditional search. There is no "AI Search Console."

GSC Data Transformation Approach ​

The insight: Google Search Console queries are evolving to look more like AI prompts. As users shift behavior between Google and AI platforms, the same intent surfaces in both places. GSC data becomes a proxy for AI prompt discovery.

The Transformation Process ​

  1. Extract GSC queries -- Pull all queries driving impressions and clicks for your domain
  2. Filter for conversational patterns -- Identify queries that resemble natural language prompts (longer queries, question formats, comparison structures)
  3. Cluster by intent -- Group similar queries into intent clusters that map to how an AI system would decompose them
  4. Generate prompt variants -- Transform clustered queries into the prompt formats users likely use on AI platforms
  5. Validate against AI platforms -- Test generated prompts on ChatGPT, Perplexity, and Google AI Mode to see if your content appears

3-Layer Technical Implementation ​

Layer 1: Gemini Embedding ​

The first layer uses Google's Gemini Embedding model to convert GSC queries into vector representations. This enables semantic clustering -- grouping queries by meaning rather than keyword overlap.

Process:

  • Export GSC query data (query, impressions, clicks, position)
  • Run each query through the Gemini Embedding API
  • Store embeddings in a vector database
  • Cluster embeddings using cosine similarity thresholds
  • Each cluster represents a prompt intent group

Layer 2: Gemini Flash ​

The second layer uses Gemini Flash to transform query clusters into natural language prompts. Gemini Flash takes the clustered keywords and generates the conversational prompts a user would likely type into an AI platform.

Process:

  • Feed each query cluster to Gemini Flash
  • Prompt: "Given these search queries, generate the 5 most likely conversational prompts a user would type into ChatGPT or Perplexity to get this information"
  • Validate that generated prompts capture the full intent of the cluster
  • Score prompts by estimated frequency (based on the combined impressions of the source queries)

Layer 3: BrightData Scraper ​

The third layer uses BrightData's scraping infrastructure to validate prompts against live AI platforms.

Process:

  • Take the generated prompts from Layer 2
  • Submit each prompt to ChatGPT, Perplexity, and Google AI Mode via BrightData
  • Capture the AI response for each prompt
  • Check whether your domain is cited in the response
  • Record citation position, context, and exact text used

Limitations and Honest Assessment ​

Honest Limitations

This methodology is a proxy, not a direct measurement. Important caveats:

What This Does Well ​

  • Discovers prompt intents you would never have guessed
  • Creates a systematic pipeline from data to actionable prompts
  • Validates citations across platforms at scale

What This Does Not Do ​

  • It does not capture actual user prompts -- It generates probable prompts from GSC data
  • GSC is an imperfect proxy -- Users may prompt AI differently than they search Google
  • Prompt behavior is evolving rapidly -- Today's patterns may not hold in 6 months
  • Citation validation is point-in-time -- AI responses change with every model update

Known Gaps ​

  • Private/personalized prompts are invisible
  • Multi-turn conversations (follow-up prompts) are not captured
  • Platform-specific prompt styles (e.g., users prompt Claude differently than ChatGPT) are not differentiated
  • The method works best for informational queries; transactional and navigational prompts are harder to predict

Stakeholder Communication Shift ​

This methodology changes how you report on AI visibility to stakeholders.

Before Prompt Tracking ​

  • "We think users ask about X on ChatGPT"
  • "We wrote content targeting AI prompts"
  • "We have no way to measure AI search performance"

After Prompt Tracking ​

  • "GSC data shows 847 conversational queries in our niche, clustered into 23 prompt intent groups"
  • "We validated our citation status for the top 50 prompts across 3 AI platforms"
  • "We are cited in 34% of validated prompts on ChatGPT, 28% on Perplexity, 41% on Google AI Mode"
  • "These 12 prompt groups represent the highest-value optimization targets based on impression volume and current non-citation"

The shift is from speculation to data-informed estimation. It is not perfect measurement, but it is significantly better than guessing.

Further Reading ​

  • CiteMET Method -- Distribution strategy for the content you discover needs optimization
  • RRF Top-n Playbook -- The ranking math behind why your content gets cited or not