---
title: GPT & Perplexity Prompt Trace
category: product
entity_type: skill
price: $15
canonical: https://forgehouse.ai/skills/gpt-perplexity-prompt-trace/
lang: en
hreflang_alt: https://forgehouse.ai/tr/skiller/gpt-perplexity-prompt-trace/
last_updated: 2026-06-20
---

# GPT & Perplexity Prompt Trace

> AI Search Forensic Intelligence

A forensic intelligence layer for AI search that reverse-engineers which prompts cite which of your pages across ChatGPT, Perplexity, Claude, Gemini, and AI Overviews. Where standard analytics go dark on AI referrals, this builds a reverse citation index and a winning-prompt catalog, telling you exactly which question formulas earn citations and which ones leave a content gap. It is the difference between knowing your citation rate and knowing what to write next.

## Use cases
- Reverse-trace the prompts that cite a specific page
- Build a per-page winning-prompt catalog across multiple engines
- Win/loss cohort analysis to find which intent types earn citations
- Content gap discovery from prompts that fail to cite you
- Competitor citation tracking when an engine cites them instead
- Weekly citation diff to catch losses after model updates

## Benefits
- Turn vague AI-search visibility into a concrete sprint brief
- Find the exact question types and formats that trigger citations
- Spot citation loss early after AI engine model refreshes
- Give clients a tangible 'cited by N prompts this week' value signal

## What’s included
- Prompt fingerprint JSON schema with intent type, entity focus, and format request
- Five-intent taxonomy (definition, how-to, comparison, best, troubleshoot) with cite behavior
- Playwright + API batch trace runner across five engines with rate-limit control
- PostgreSQL reverse citation index with GIN index for fast URL lookups
- Statistical win/loss analysis with position-weighted scoring
- Weekly winning-prompt email report and free manual mini-trace method

## Who it’s for
For SEO and AEO teams managing many pages who need to know not just whether AI engines cite them, but exactly which prompts to win next.

## How it runs
Which prompts get your pages cited by AI engines stops being a guess. Every week a prompt catalog fires at five engines, each citation lands in a fingerprint database, and statistics separate winning patterns from losing ones.
1. Builds a prompt catalog from 5 intent types (definition, how-to, comparison, best-of, troubleshooting) crossed with target entities, each prompt carrying 3-10 wording variations to catch phrasing-dependent behavior
2. Fires the batch at five engines (ChatGPT, Perplexity, Claude, Gemini, AI Overviews) under rate-limit-aware concurrency; engines without an API get scraped through a headless browser
3. Logs every response as a fingerprint record: which URLs were cited, at what position, in what sentiment context, written append-only so a run can never corrupt history
4. Builds the reverse index in the database so a single query answers which prompts cite this page, with position-weighted scoring because a first-position cite carries roughly 5x the traffic of a fifth
5. Runs win/loss cohort statistics (chi-square, Mann-Whitney) across prompt features to prove which patterns actually trigger citations, instead of guessing
6. Diffs weekly snapshots against the prior week, alerts on 15%+ citation loss, and ships the winning and losing prompt tables as the next sprint's content brief

## FAQ
### We mostly care about Google AI Overviews, is that covered?
Yes. Tracing runs across five engines: ChatGPT, Perplexity, Claude, Gemini, and AI Overviews: and the per-page winning-prompt catalog is built engine by engine, so you can read the AI Overviews results in isolation.

### AI engines don't report referrals, so how can you know which prompts cite us?
You define a fingerprinted prompt catalog and the batch runner executes it against the engines via Playwright and APIs, writing results into a PostgreSQL reverse citation index. It's a controlled experiment you run yourself, not access to anyone's hidden data.

### Does it show what real users actually type into ChatGPT?
No, that data isn't available to anyone. What it shows is which question types in your own prompt set earn citations for your pages, and which ones reveal a content gap to write against.

## Price
$15, one-time, no subscription. VAT included.

Related guide: [AI and LLM engineering](https://forgehouse.ai/guides/ai-llm-engineering/)