The challenge
A site migration silently breaks content in dozens of subtle ways: a product page inherits another product's copy, a spec number changes, an image swaps to the wrong product, a contact form stops rendering, a heading truncates, a datasheet mis-transcribes a model number. Reviewing ~200 pages by eye takes days and misses things. The goal was human-level, word-by-word comparison at scale — reliably, and presented so a developer can act immediately.
It is hard because the two sites use different templates (naive text comparison drowns in nav/footer noise), staging serves malformed HTML that breaks standard parsers, many differences are intentional redesigns rather than defects, the worst defects (wrong images, broken layout, missing forms) are invisible to any text diff, and datasheets exist as PDFs on live but HTML pages on staging.
Three evolving passes
Sitemap + text
Header/footer-stripped word comparison plus SEO, links, and structure checks. Surfaced launch blockers: site-wide noindex, 98 product URLs needing 301 redirects, missing canonicals, broken links.
Full mapping + text classification
Authoritative URL pairing from the client's crawl exports, word-diffed, then an 83-agent workflow classified each change with adversarial verification. Found 21 confirmed issues — but text can't see images or layout.
Human-level: rendered + vision
Every page rendered in a real browser (visible text + full-page screenshot); datasheets checked against their source-PDF pages. A 286-agent vision workflow then looked at both screenshots and read both texts for every unit, catching visual defects the text passes were blind to.
How the v3 pipeline works
- Map — reconcile live vs staging URL lists, apply exclusions, pair each in-scope page to its staging twin (247-row mapping).
- Render — 331 units (177 page pairs + 154 datasheet↔PDF pairs) across 6 parallel browser workers → 510 screenshots + 463 PDF page-images + visible text.
- Compare — vision agents read each unit's screenshots and texts, returning a structured verdict with exact expected-vs-actual values.
- Verify — every flagged defect is adversarially re-checked; false positives (shared chrome, intentional edits, render artifacts) are dropped.
- Present — a filterable dashboard where each issue expands to a description, the visual note, the exact change, and one-click links to the staging and live pages.
What it found
Of 1,000 live HTML pages, 214 were in scope after excluding Spanish, knowledge-center, reps, states, and listing pages. The v3 pass confirmed 158 issues (51 high, 66 medium), including two product pages serving 404s, a vacuum-pump page showing a dryer's copy and image, blank expert headshots, 37 pages missing their contact form, wrong distributor contact details, and 129 datasheet-vs-PDF transcription errors.
Systematic patterns (fix once, clears all)
- "Balmore" → "Baltimore" footer typo — 91 pages
- Dropped toll-free number — 85 pages
- Missing / broken images — 41 pages
- Missing contact forms — 37 pages
- Truncation typos from an Elementor widget bug ("contro", "nfrared", "ou're") — 22 pages
- Wrong document codes copied from another manual — 17 pages
Deliverables
Three dashboards — v3 (human-level), v2 (text), v1 — plus companion files: the full report, the URL-mapping spreadsheet, the raw verdicts JSON, and two browsable screenshot galleries (shots and pdfimg). Every dashboard issue is fix-ready: a plain-English description, the exact live-vs-staging values, and one-click Open-staging / Open-live buttons.