# AI Answer Trace

> Ask Claude, ChatGPT, and Gemini a question and capture the full evidence trail behind each answer: the search queries each engine ran, the pages it retrieved, and the sources it cited. The raw material of GEO measurement. Needs AI engine API keys, not an Xpoz account.

- Skill: `xpozpublic/ai-answer-trace-2` (Agent Skill)
- Install (CLI): `npx skillmds@latest add xpozpublic/ai-answer-trace-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/xpozpublic/ai-answer-trace-2/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Integrations & APIs
- Author: XPOZpublic (https://skillmd.com/u/xpozpublic)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/xpozpublic/ai-answer-trace-2

---


# AI Answer Trace

**Not just what the engines say: which sources made them say it.**

Asks the major AI engines a question exactly the way a real user would, with live web search on, and captures the complete evidence trail. Every trace records three layers: the answer, the search queries the engine silently ran with the pages each search retrieved, and the URLs the answer actually cited (with answer spans where the API provides them).

## Setup

Set whichever engine keys you have; each script runs independently, one key is enough to start:

```bash
export ANTHROPIC_API_KEY=...   # Claude trace
export OPENAI_API_KEY=...      # ChatGPT trace
export GEMINI_API_KEY=...      # Gemini trace
```

## Trace Process

### Step 1: Run the question per engine

Fetch the three self-contained trace scripts (Claude, ChatGPT, Gemini) from the [full skill source](https://github.com/XPOZpublic/xpoz-agent-skills/blob/main/skills/ai-answer-trace/SKILL.md), which carries them inline with per-engine dependency headers; save and run them locally. Each saves a structured JSON trace per run: answer, per-search queries and retrieved pages, cited URLs. Default 2 samples per engine (1 for a quick look, 3 for a stable read); phrase questions the way a real user would, not keyword-style.

### Step 2: Engine quirks, handled at capture time

- **Gemini**: grounding URLs are expiring Google redirects; the script resolves each to its real destination.
- **ChatGPT**: retrieved sources only appear when explicitly requested; `?utm_source=openai` suffixes are stripped from retrieved sources (cited URLs keep them; strip before exact-URL comparisons).
- **Claude**: citations scattered across content blocks are reassembled in answer order.

### Step 3: Aggregate cited domains

A deterministic aggregator counts cited domains across engines and samples.

### Step 4: Analyze

- **Presence**: does the brand appear in each answer, named early or as an afterthought?
- **Citation path**: which cited URL carried it there (own domain, directory, community thread, roundup)?
- **Winners**: which domains dominate, and what surface types are they?
- **Near-misses**: retrieved-but-never-cited pages show what the engine considered and rejected (Claude and ChatGPT; Gemini's retrieved set is essentially its cited set).
- **Stability**: sources recurring across samples are signal; one-sample citations are noise until they repeat.

## Tips

Keep the trace JSONs; re-runs against the same questions are real before/afters | Feeds `geo-visibility-check` (verdicts) and `geo-reddit` (cited-thread mining) | Weekly program: [geo-seo-agent](https://github.com/XPOZpublic/geo-seo-agent)