How the audit works, and what it can’t tell you

An audit compares one ad with the page it links to and asks: does the page continue the reason someone clicked? This page explains every step, the fixed weights, the checks against your original material and the limits. Current pipeline version: 2026-10-22.3; every saved audit records the version and AI model that produced it.

1. What the AI reads

• The ad: the image, a carousel’s cards in order, or up to ten frames of a video (densest in the first seconds), its on-screen text, the voice-over transcribed with timestamps, and the ad text you paste.
• The page: its text as a visitor’s browser receives it, in order, with headings and buttons marked. Bot checks, login pages and cookie walls stop the audit before any AI work.
• Your context: the channel (social visitors arrive with only the ad in mind, search visitors with a query) and the target market.
• Not seen: the page as an image. That’s why visual continuity is shown but not scored.

2. The message-match score

The AI rates each dimension from 0 to 100 with a one-sentence reason. The overall score is a weighted mean computed in code, so the weights are fixed and the same for everyone:

The score is a content assessment. It is not a conversion rate and doesn’t predict one.

3. Checks against your original material

• Every quote is located. Quotes in findings are searched in the ad text, the voice-over (with the second it’s said) and the page. Paraphrases are marked “not verbatim · check”; passages the speech recognition couldn’t make out are marked “unclear audio”.
• Rule-based consistency checks run after every audit: figures that appear in no input, units that changed (percent vs percentage points, days vs business days), promises the source negates, offer conditions that went missing, ad and page asking for different actions, a different product variant.
• Promise coverage puts each promise of the ad next to the page line that shares most of its key words. It’s word overlap, not meaning: a pointer for where to look.
• Page readiness is checked from the page HTML, without AI and separate from the score. Accessibility: declared language, zoom allowed, alt text, named links and buttons, labelled form fields, one main heading, heading order, media with sound, keyboard order and the buy button's contrast (the automatically checkable part of WCAG 2.2; keyboard use, focus order and meaning still need a person). Machine readability: noindex, schema.org Product data with price, currency, availability and an identifier, the shown rating in that data, meta description, canonical, link previews, and whether robots.txt blocks AI crawlers such as GPTBot, ClaudeBot or PerplexityBot.
• You can correct the audit. Fix a misread hook, variant or offer, add page content the crawl missed, and the comparison is recalculated as a new version.

4. Generated pages

• Every factual claim must cite an exact quote from the ad, the page, the product data or your approved claims. Numbers must appear in a source; units can’t change; negated statements can’t become promises; reviews must exist on your page.
• Pages declare their language, keep zoom, name the product photo, keep button text readable against the button color and carry schema.org Product data built only from your shop's captured name, description, images, price, availability and a rating the page shows.
• Your brand rules apply: forbidden wording blocks publishing and store import, restricted claims need approval, and kinds of change you ruled out aren’t recommended.
• Before publishing, a check shows where the buttons actually lead (product, collection, homepage, checkout), insecure images, mobile first-screen lengths and offer end dates. Publishing needs your explicit confirmation of the claims.

5. What isn’t measured

• No conversion uplift is promised or predicted. Results come from tests you run; the experiment log shows rates, revenue and data-quality warnings, not statistical significance.
• Published-page analytics count views, visitors per day (cookieless, so a returning visitor counts again on a later day), scroll depth, active time and button clicks. Orders come only from the Shopify order sync and only when the visit came through the page's button.
• AI output can be wrong. Every claim, price and guarantee needs a human check before it goes live.

6. What connects and what doesn’t

See a full run in the example audit, or try the live demo.