Verifex is Vohkus's internal website audit tool. Enter any domain or page URL and it produces a full diagnostic report in about ninety seconds: an overall score out of 100, eleven individual scores covering design, copy, conversion, user experience, performance, SEO, AEO, GEO, trust, psychology and clarity, and most importantly the reasoning behind every number and a complete, ranked fix plan for each category. The intent is practical: a web team should be able to read a report, work through the fix lists, re-audit, and watch the scores move. Audit history is tracked per page so before-and-after progress is visible. Access is by sign-in link, restricted to Vohkus and Injekt Media email addresses, and every audit is logged.
Verifex combines three engines. Google's PageSpeed Insights (the Lighthouse engine, run on Google Cloud) loads the actual page under mobile conditions and returns real laboratory measurements like load times, Core Web Vitals, and category scores for performance, SEO, accessibility and best practices, along with a rendered screenshot. Verifex's own technical inspector then parses the page's source code directly and runs around twenty deterministic checks: title and meta description lengths, canonical tags, Open Graph and Twitter cards, H1 structure, image alt text, JSON-LD structured data validity, robots.txt, sitemap presence, and whether AI crawlers such as GPTBot and ClaudeBot are permitted. Finally, Claude Sonnet 4.6 (Anthropic's AI model) receives the screenshot, the live page content and all of the measured facts, and performs the qualitative analysis. Accounts and audit logging run on Supabase; hosting is on Vercel.
Because the measured layer contains no opinion. Performance and SEO scores come straight from Google Lighthouse, the same tool Google itself provides to developers. They are labelled "measured" on the dashboard. The technical checks are parsed facts: when Verifex says a title tag is 64 characters or a page has two H1s, that is read directly from the code and can be verified by anyone in seconds by viewing the page source. These numbers are repeatable: run the audit again and the facts stay the facts (speed metrics vary slightly run to run, as they do in any lab test, because real network conditions vary).
The AI aspect of this app has been trained to judge the craft of the page the way a consultant would: is the design coherent, the copy persuasive, the path to action clear, the content answerable and citable by search and AI engines? It works under strict evidence rules built and tested during development: it must look at a real screenshot and the actual fetched page before judging; it may never claim something is absent without verifying, including checking obvious sub-pages; every score must cite the specific evidence behind it; fix lists must be complete and concrete, with generic advice banned; famous brands get no credit and no penalty, because the page is being judged, not the business.
These scores are evidence-grounded expert judgment, not physical measurements. Like two consultants reviewing the same site, a re-run may differ by a few points, and occasionally a judgment can be argued with, which is why every claim is written to be checkable against the page itself, and why the report separates what is measured from what is analysed. The right way to use the analysed scores is the way you'd use a trusted reviewer: read the evidence, fix the gaps, re-audit, and let the results show the improvement.