morphiq-track（形态轨迹）

Pipeline Position

Step 4 of 4 — measurement + flywheel.

• Input: Build Output (JSON) from morphiq-build + MORPHIQ-TRACKER.md (persistent state).
• Output: Delta Report (JSON) → loops back to morphiq-rank.
• Owns: MORPHIQ-TRACKER.md — generates on first run, updates every run.
• Owns: morphiq-track/ state directory — JSON state layer for prompts, results, citations.
• Drives: 3 ongoing workflows (Content Optimization, Content Creation, Query Fanout Expansion).
• Data contract: See PIPELINE.md §4 for the Delta Report, §5 for MORPHIQ-TRACKER.md, §6 for the JSON State Layer.

Purpose

Morphiq Track is the measurement and flywheel skill. It queries AI providers to measure brand visibility, computes GEO scores and Share of Voice, tracks deltas over time, and drives three ongoing workflows that feed back into the pipeline.

Workflow

Step 0: Initialize or Load State

Check if morphiq-track/manifest.json exists in the project root.

• Missing (first run): Proceed to Step 1. The state directory will be created.
• Present (subsequent run): Load morphiq-track/prompts.json directly — this contains the full prompt set with config, metadata, and tracking state. Skip to Step 2. If recommendations.cooldown_days has elapsed since recommendations.last_generated, generate 20 new recommendations via create-prompts.py --state-dir morphiq-track/ --refresh.
• Migration (tracker exists but no state dir): Parse MORPHIQ-TRACKER.md §8 to bootstrap prompts.json, parse §7 to bootstrap citations.json. See references/state-layer.md Migration section.

For state layer specification, read references/state-layer.md.

Step 1: Generate Prompts

First run only. Generate 50 prompts across 5 GEO categories:

Apply quality rules per category. Add temporal markers to 70%+ prompts. Include entities in comparison/technical prompts.

Run scripts/create-prompts.py --state-dir morphiq-track/ --brand {brand} --category {category} --competitors {competitors}. This writes morphiq-track/prompts.json and initializes morphiq-track/manifest.json.

For taxonomy, fanout profiles, and generation rules, read references/prompt-taxonomy.md.

Step 2: Query AI Providers

Distribute prompts evenly across 4 providers. Execute using scripts/run-queries.py --state-dir morphiq-track/ --mode execute. This reads prompts from morphiq-track/prompts.json, writes versioned results to morphiq-track/results/track-{date}.json, and updates morphiq-track/manifest.json.

Mandatory requirements for every query:

1. Full response text. Store the complete response — never truncate. morphiq-build's content creation workflow requires the full text for analysis.
2. Sub-query extraction. For each provider that exposes tool calls, extract the search queries the model issued. These feed Workflow C (Query Fanout Expansion) and invisible SoV.
3. Citation deduplication. After collecting citations per response, strip UTM/tracking params from URLs and deduplicate. Track citation_weight (number of times each URL was cited).
4. Retry on transient failure. Retry once with 2-second delay before marking as error. This handles rate limits.

Provider-specific requirements:

• OpenAI: Iterate response.output for items with type == "web_search_call" — extract the query field into sub_queries[]. This reveals GPT's site: operator searches and two-phase research pattern.
• Perplexity: Citations are a Perplexity-specific field. Check response.citations, then response.model_extra["citations"], then response.__dict__["citations"], then response.choices[0].message.model_extra["citations"]. The OpenAI-compatible client puts unknown API fields in model_extra.
• Anthropic: Tool config must be {"type": "web_search_20250305", "name": "web_search", "max_uses": 5}. Try model claude-sonnet-4-5-20250514 first, fall back to claude-sonnet-4-20250514. Response content blocks include text (final answer), web_search_tool_result (search results with URLs), and server_tool_use (the search call). Extract text only from text blocks; extract citations from both text block inline citations and web_search_tool_result block content.
• Gemini: Grounding metadata returns vertexaisearch.cloud.google.com redirect URLs. Follow the redirect to get the real URL. If redirect fails, use the grounding_chunk.web.title as fallback domain: {url: proxy_url, title: title, resolved_domain: title}.

For full provider config and response pipeline, read references/provider-strategies.md. For selection rules and distribution, read references/query-targets.md.

Step 3: Analyze Responses

5-step pipeline: extract raw response/citations/sub-queries → structured analysis (using the agent's reasoning capabilities) → brand mention validation (exact → TLD → LLM judge) → competitor filtering → entity normalization.

Input requirements for analysis: Each response must include the full response text, deduplicated citations with citation_weight, and extracted sub-queries. The analysis uses the config block from the prompts file for brand, domain, and competitors — never hardcoded values.

Step 4: Compute GEO Score

GEO = mean(provider_scores)
Weighted GEO = (Organic × 0.45) + (Competitor × 0.22) + (How-to × 0.22) + (Brand × 0.11)

Thresholds: ≥60 Excellent, ≥40 Good, ≥20 Fair, ≥10 Poor, <10 Very Poor.

For GEO methodology, read references/query-targets.md.

Step 5: Compute Share of Voice

Three SoV tiers:

Track Conversion Gap = Influence SoV − Citation SoV.

For SoV methodology, read references/share-of-voice.md.

Step 6: Compute Deltas

Compare against previous snapshot using scripts/diff-results.py --state-dir morphiq-track/. The script reads manifest.json to auto-resolve the current (runs[0]) and previous (runs[1]) results paths, and reads morphiq-track/citations.json for previous citation state. Flag changes >5 points. Generate flagged actions for regressions, losses, displacement, and conversion gaps.

For delta methodology, read references/delta-scoring.md.

Step 7: Update State Layer and MORPHIQ-TRACKER.md

State layer updates (JSON — source of truth for track-owned data):

1. Rebuild morphiq-track/citations.json from current results + previous citation state (gained/lost/stable)
2. Update morphiq-track/prompts.json tracking fields (mentioned, cited, best_provider, runs_tracked, last_run)
3. Update morphiq-track/manifest.json updated_at

Tracker updates (markdown — user-facing dashboard): Project state layer data into MORPHIQ-TRACKER.md sections 5-9 and 14 (SoV, SoV Trend, Citations, Prompts, Competitors, Run History). Update remaining sections (1-4, 10-13) per tracker-spec.md rules.

For tracker specification, read references/tracker-spec.md. For state layer specification, read references/state-layer.md.

Step 8: Produce Delta Report

Assemble JSON (PIPELINE.md §4): SoV metrics, citations, per-provider data, competitors, flagged actions, content queue. Loops back to morphiq-rank.

Three Ongoing Workflows

Workflow A: Content Optimization

1. Identify pages with declining SoV or lost citations
2. Feed to morphiq-build (existing content path)
3. Re-track to measure impact

Workflow B: Content Creation

1. Collect prompts where brand is absent
2. Identify competitor citation sources
3. Generate content briefs for missing coverage
4. Feed to morphiq-build (new content path)

Workflow C: Query Fanout Expansion

1. Run scripts/analyze-fanout.py --state-dir morphiq-track/ with optional --scan-report for page inventory and simulated queries
2. Script extracts sub-queries from latest track results, merges with scan simulated queries (fills Perplexity/Gemini gap)
3. Compares against site page inventory to identify unanswered sub-queries
4. Extracts competitor citation sources for each unanswered sub-query
5. Generates content briefs prioritized by citation weight (site: 2x, citation-producing 1.5x, silent 0.5x)
6. Output feeds Delta Report content_creation_queue via --fanout flag on generate-report.py
7. Update MORPHIQ-TRACKER.md §12 (Query Fanout Coverage) and §13 (Content Creation Queue) with new entries

Reference Files