Recommendation Audit

TL;DR: Run the Recommendation Audit when you have a Shopify product and a buyer question. Colter baselines the product evidence, identifies one bounded correction, and records whether the same prompt changes in the same named answer engine.

Recommendation Audit is Colter's merchant product. Check, Fix, Test, Lens, and Verify remain available as technical capabilities inside the audit and through the CLI, API, SDK, and MCP server.

The Four-Stage Loop

Stage	Question	Colter capability
Baseline	What public product evidence can an agent retrieve for this exact buyer question?	Check and recommend-audit
Fix	What single evidence gap is most likely to make the answer less accurate or complete?	Fix guidance
Prove	What did the same named engine actually answer before and after the correction?	Content-hashed answer observations and, when needed, Test
Monitor	Did the unchanged prompt gain, retain, lose, or still omit the product or target citation?	Comparable reruns and, when useful, Lens

What You Need

• one Shopify storefront
• one to three exact product pages
• one or more real buyer questions
• one named engine for each observed answer: ChatGPT, Google AI Mode, Perplexity, Claude, or Other

The audit is read-only. It does not change Shopify, publish product content, scrape private answer history, or ask for a model credential.

Web Workflow

1. Open Recommendation Audit.
2. Enter the Shopify store, exact product page, buyer question, and named engine.
3. Review the public-evidence baseline and one recommended correction.
4. Capture the real engine answer locally and import it for hashing and bounded observation.
5. Apply the correction outside Colter only after review.
6. Rerun the unchanged buyer question in the same engine and compare the observation.

CLI Workflow

colter recommend-audit https://store.example.com \
  --product https://store.example.com/products/example \
  --intent "best lightweight travel backpack under $150" \
  --json

Without an observed-answer file, the command reports a readiness baseline only. A readiness score is not proof of placement, a mention is not a recommendation, and a changed score is not proof that an answer changed.

Evidence Boundary

Colter keeps three classes of output separate:

• Measured: public storefront facts retrieved by Colter.
• Observed: the merchant, product, and citations present in an imported answer from one named engine and prompt.
• Inferred: a correction priority or mechanism that must still be tested.

Success is a comparable answer delta: gained, retained, lost, or still absent. The audit does not promise that a correction will cause a model to recommend a product.

Technical Capabilities

• Check creates the public-evidence baseline.
• Fix turns an evidenced gap into a bounded correction.
• Test answers a runtime question only when public evidence is insufficient.
• Lens monitors observed agent traffic when ongoing measurement is warranted.
• Verify produces conformance evidence for technical teams.

Next Steps

• Run Recommendation Audit
• For Merchants
• For Agencies
• Developer Quickstart