Checkout, price formatting and product copy are where Arabic breaks in ways that cost money rather than looking wrong. These are the checks that catch them, and the one case where an automated scan cannot reach you at all.
A form flips cleanly and the card field keeps its left-to-right numerals, which is correct. What does not flip is the step indicator built from physical margins, so the progress bar runs backwards against the flow of the page. Nobody reports it; they abandon.
Arabic-Indic digits, a currency symbol on the wrong side, a decimal separator that changes meaning between locales. A price is the one string on a storefront where a reader who hesitates does not come back to check.
Category pages translate, and the long tail does not. Size charts, care instructions, shipping tables and the strings inside a third-party review widget stay English on a page that is otherwise Arabic.
Arabic runs longer than English for the same content, so a line item that fits at every English width clips at the same widths in Arabic. It is measured per breakpoint because it is a per-breakpoint defect.
Deterministic checks with the evidence attached: a screenshot, the selector, and the change that fixes it. Run at three breakpoints, because a defect that appears at one width is a defect at that width.
A storefront changes weekly. Monitoring re-scans on a cadence and mails the owner when the score drops or a new critical appears, so a regression is caught by the scan rather than by a customer.
An expiring link to the findings, with what was measured stated on it. The person who signs off on the work is rarely the person with the account.
Commerce sites are the most likely to sit behind a bot check, and a bot check does its job: it refuses an automated browser, including ours. When that happens the scan says so instead of scoring a challenge page, and the failure routes you to a human audit. A score of a page we never rendered would be worse than no score.
One page, no account. If bot protection stops it, the result says so and offers the audit instead.
Deciding between approaches first? Compare the four.