AttestorAmazon ad-compliance evidence
Case fileAurora
File note. Amazon pages and data are fictional and drawn in your browser; the compliance rules are ported from the real pipeline. Nothing leaves this page.

Exhibit board

Every Sponsored placement Aurora booked across Amazon, captured as a filed exhibit and ruled against the media plan. The sweep runs the moment this page opens.

Evidence store

What the unattended job leaves behind is one folder per placement, per run: the screenshot and the scraped metadata.json, kept so any claim about what ran on Amazon can be proven on demand.

Compliance report

Auto-written after each sweep: the document the audit team used to assemble by hand. Marketplace by marketplace: what the media plan promised versus what was actually live on Amazon.

How it worked

This started by shadowing an audit team and noticing what hurt: to prove a brand's booked ads were really running on Amazon (correct Sponsored placement, keyword and marketplace), they were manually screenshotting each one, hand-typing the details, and filing it, across several countries. The tool automated the entire loop.

By hand, beforeSearch each booked keyword · screenshot the Sponsored slot · type out the ASIN, placement and position · file it somewhere · repeat per marketplace, per country · fall behind.
On a schedule, afterOpen · screenshot · scrape the served ad · rule against the media plan · file the evidence · auto-write the report. Unattended, across every marketplace.

the loop · once per booked placement

open the keyword / product page→ screenshot the proof→ scrape the served ad→ check vs the media plan→ file the evidence→ auto-write the report
  1. A headless browser fleet opened each placement, per marketplace
    The media plan became a job list: for every marketplace (amazon.it, .de, .es, .co.uk…), every booked keyword or product page and the Sponsored unit it was meant to run. A pool of headless Chrome instances (driven with Selenium) opened each search or product page in the right locale and waited for the ad slots to render.
  2. It screenshotted the proof and scraped the served ad
    A capture for the record, then it read the DOM the audit team had been transcribing by hand: was the Sponsored placement actually present, which ASIN it promoted, its slot (top-of-search, product page), whether the “Sponsored” label was shown, and whether the product still held the Buy Box and was in stock.
  3. Ported human judgement into rules
    The checks the team ran in their heads became code: is the ad serving for the booked keyword at all? correct ASIN? in the booked slot? is the required “Sponsored” disclosure present? still inside the booked flight? still winning the Buy Box, or is spend leaking on an out-of-stock listing? Each rule emitted a pass, watch or breach with a plain-language reason.
  4. Filed tamper-evident evidence
    Everything landed in a predictable tree, evidence/<run>/<marketplace>/<placement>/, holding the screenshot plus metadata.json, so proof of any placement, in any market, on any date could be retrieved on demand instead of hunted for.
  5. Wrote the report itself
    A per-marketplace compliance summary with every breach linked to its evidence, plus a standing promised-versus-actual check that ran day after day, across countries, without anyone remembering to.
PythonSeleniumheadless Chromeweb scraping multi-marketplacescheduled jobevidence storereport generation

This demo is a faithful rebuild, not the original code. The production tool drove a real browser fleet against live Amazon pages across several markets; here those pages are drawn in-browser as SVG so the demo runs anywhere, offline. The rules, the evidence shape and the report are ported straight across. Client confidential: the brand, ASINs and marketplaces shown are invented.