Marketing

Auditing a Topic's Search Results with an Agent

• Choosing a Reader for Each Page Type: HTML, PDF and JS-Rendered
• Choosing Chunk Size and Overlap for a Search Audit Index
• Routing Errors Versus Prompt Errors: Telling Them Apart
• Using an AI Agent to Audit a Topic's Search Results
• What Source Pages Cost to Fetch and Read
• When a Search Audit Should Run on Retrieval, Not a Long Prompt
• A Canary Row: Routing a Slice of Traffic Through the Edited Table
• A Confusion Matrix Workflow Your Team Can Read
• A Cost Ceiling for a Full Site Audit
• A Fetch Budget per Source Page: Retry and Render Limits
• A Fetch Budget per Source Page: Setting Retry and Render Limits
• A Human Review Step for Search Gap Recommendations
• A Reader Fallback Ladder That Ends at a Human
• A Regression Set for a Support Index: Queries and Answers
• Alerting on Injection Attempts Without Drowning the On-Call
• Caching Fetched Page Text: Cache Keys, Expiry and What to Store
• Choosing a Chunking Strategy for Support Pages That Get Edited
• Comparing an Embedding Index to BM25 on the Same Query Set
• Fallback Ladders That Keep a Reasoning Trace Straight
• Keeping a Retrieval Index Fresh When Pages Change Weekly
• Keeping the Index in Step When a Page Is Rewritten
• Logging an Audit: Labels, Reviewers and Scores in Postgres
• Logging Reader Cost and Quality So You Can Compare Them
• Logging Search Audits So a Model Can Explain What Changed
• Measuring Retrieval Hit Rate Before You Trust the Audit
• Storing Fetched Page Text and Its Reader in Postgres
• Storing Page Hashes and Chunk Vectors in Postgres
• Storing Query Logs So a Golden Set Can Be Sampled Later
• What a Search Audit Costs in Model Tokens and Crawl Requests
• What a Vector Store Costs per Thousand Queries
• When a Cheaper Reader Wrecks the Verifier: A Quality Gate
• When a Reviewer's Override Becomes a Calibration Label
• When a Second Judge Disagrees: Three-Way Log
• When an Audit Needs a Reranker, Not More Context
• When to Alert on a Stale Index: Thresholds and a Runbook
• Writing the Labeled Query Set Before You Tune Anything