Brand Protection Monitoring at Scale

Brand protection monitoring is a detection problem with a legal output. The system has to find listings that infringe, prove what it found in a form that holds up, and avoid burying the team in false positives.

That last requirement is the one that decides whether a monitoring programme actually functions.

What is worth detecting

Four violation classes, each with different detection logic and different evidence needs.

Counterfeits. Listings offering replica or look-alike goods under your mark. Detection is keyword and image based, and the listing itself is the evidence.

MAP violations. Authorised resellers advertising below the manufacturer's minimum advertised price. Detection is price comparison against the policy, scoped to authorised sellers.

Unauthorised sellers. Listings from sellers not in your distribution network, particularly on marketplaces where the listing claims to be your product.

Brand and content misuse. Your marks used in titles, images or descriptions without authorisation, including in sponsored placements.

Price comparison alone does not identify violations, because a low price is not proof of anything. Every class needs corroboration from a second signal: seller identity, image similarity, or the listing's own claims.

The detection pipeline

Five stages, and each has a failure mode.

1. Discovery. The listings to evaluate. From marketplace search, category browsing, or a watchlist of product identifiers. This stage needs broad coverage, and it is where the address layer matters most, because results vary by region and by seller network.

2. Normalisation. Parsing each listing into a comparable record: product identifier, seller, price, currency, shipping terms, images, and the region the listing is viewable from.

3. Scoring. Comparing against expected values. Keyword match on the title, perceptual hash against your official product images, price deviation from the policy or the authorised range, seller against the authorised list.

4. Evidence capture. For candidates above a threshold, capture a complete, timestamped record: the rendered page, the price at that moment, the seller details, the images, and the region the capture came from.

5. Review and action. Human confirmation, then takedown request, seller communication or escalation.

The error that breaks the programme is skipping stage 4, then being unable to act on a confirmed violation because the evidence is a parsed row rather than a page.

Capture from the right vantage point

Marketplace content is region-dependent. Listings visible in one country may not appear in another, prices are localised, and shipping terms change by origin.

Three requirements for evidence:

• An address in the market you are monitoring. A capture from elsewhere may show different listings, a different price, or nothing at all.
• A stable session per region for the duration of a sweep, so the vantage point is consistent across the evidence set.
• Timestamp, region and exit address recorded with every capture, because the evidence is only meaningful with its vantage point attached.

This is the same locality requirement as any regional collection, described in Residential Proxies for Brand Protection and, more generally, in E-Commerce Data Scraping.

False positives

The reason programmes get abandoned. A detector that flags every reseller is a detector nobody reads.

Practical controls, in order of effect:

• Require two signals. Price deviation alone is not a violation. Price deviation plus an unrecognised seller is worth reviewing.
• Maintain an authoritative seller list and suppress matches from it, rather than scoring them.
• Use image similarity with a threshold, not exact matching, and tune it against a labelled set of your own official imagery.
• Scope MAP checks to the policy. A price below MSRP is not a MAP violation.
• Review the false positive rate explicitly. Track it, and treat a rising rate as a bug.

A programme that surfaces tens of genuine candidates a week and is read is worth more than one that surfaces thousands and is filtered into a folder.

Cadence

Monitoring cadence follows how fast the risk moves, which differs by class.

The counterfeits row is the one with a time constraint: listings can be removed by the seller or the marketplace before a slow sweep catches them, so evidence capture needs to happen on the same run as detection.

Reporting and workflow

Three properties make the output usable by a legal or brand team.

Complete records. Listing URL, seller identity, captured page, images, price, timestamp, region, and the specific policy or right at issue. One record per candidate, not a summary row.

Auditable state. Each candidate moves through detected, reviewed, actioned, resolved. Without states, the same listing is re-reported weekly.

Volume discipline. Report confirmed violations only. A takedown request carrying a hundred false positives damages the relationship with the platform and makes the next request slower.

Infrastructure summary

Hosting ranges are a poor fit here in a specific way: they receive content that differs from the consumer view, which makes the evidence wrong rather than merely blocked. The reasoning is in Why Residential IPs Are Trusted.

For the solution context, see E-Commerce. For the address layer, Residential, and for the collection method, Collecting Amazon Product and Review Data.