M

MarqVision

The models are the product, with nothing conventional beneath them. MarqVision sells detection and enforcement at a scale no human team can reach. It scans more than 1,500 channels across 118 countries, compares findings with a brand's own product data at SKU level, analyzes video frame by frame, and generates the notices that follow. The named capabilities are all machine learning. They include Atomic Product Detections, described as using generative AI for SKU level precision, and a multi agent system that decides whether an infringement is valid and classifies the enforcement grounds.

There is also an LLM framework that generates search keywords in any language, image recognition and text analysis, and models the company says are trained on successful enforcement data. There is no content asset underneath: the raw material is marketplace listings the crawler gathers, which only the models make usable. The stated alternative is not other software but the manual monitoring the product replaces.

Accuracy is measured, published and explained, and one clause in the agreement undercuts it. Three figures appear: 97 percent on enforceable item detection using SKU level comparison, 99.8 percent across more than 48,000 domain impersonation incidents, and 94 percent SKU matching accuracy on unauthorized sales. The company says openly that these are internal benchmarks. The method is described: SKU level comparison with genuine product data, detection on several signals, and a tiered model that routes borderline cases to analysts.

A published piece names failure modes, setting out what a false positive costs, including misrepresentation liability under 17 U.S.C. 512(f), tortious interference and suspension of platform reporting. Only one figure carries a sample size and none carries a test set. Section 2.4 of the terms also says Marq Vision assumes no liability for the quality, accuracy or validity of the data gathered by the platform and that the customer shall in no event rely on it. The marketing measures accuracy and the agreement disclaims reliance on it.

Real review surfaces are published, and the point at which the system acts alone is not. The company says a human reviewer confirms every significant decision and describes a tiered model that routes only borderline cases to analysts. Each classification comes with its reasoning and a risk score, so a team can see why a determination was made. Two customer controls are contractual. Section 8.3 of the terms requires qualified customer representatives to review, validate and verify all information transmitted with due care, and to report claims only with a good faith belief in their truth.

The whitelist gives the customer a standing list of sellers it does not want enforced against. Against that sits the speed the product is sold on: 15 minutes from detection to enforcement, median responses under five hours for fake domains, responses in seconds for piracy, and enforcement actions started automatically. Both cannot be fully true at once. Significant and borderline are never defined, so nothing published says which determinations reach a person.

The customer list is large and spans categories seldom seen together. Roughly forty logos run across automotive (Nissan, GM), electronics (Panasonic), consumer goods (Henkel, Persil, Miele), beauty and fashion (Lush, Wella, Stussy, Courreges, Lemaire, Ader Error), Korean commerce and finance (Coupang, Shinsegae, Toss, KT&G), and gaming and media (Smilegate, Class101). Six case studies are published and five people speak with name and title.

They include Tae Hyun Kim, Director of Ecommerce at Miele, Ryan Dahlstrom, Global Director of Digital Commerce at Darn Tough Vermont, and John Belcaster, General Counsel of MSCHF. Operational figures are plentiful: 180 times faster enforcement for its own teams, 50 percent faster than traditional providers and a 15 minute mean time to resolution. It also reports more than ten million sellers tracked, dozens of offline raids in Shenzhen, Guangdong and Yiwu, and one million dollars of counterfeits seized. Method and dates are missing. No figure carries a measurement basis or date, and none is tied to a named customer.

One part is strong and the rest are thin or point the wrong way. Section 5 of the terms is a proper mutual confidentiality regime with notice before compelled disclosure, and it survives termination. Section 4.2 confirms the customer owns all data it sends and grants no license beyond providing the services, and section 4.3 limits Marq Vision's license to the term and to service delivery. Against that, section 2.5 grants a broad right to use information the customer submits, expressly including information about its customers, distributors, inventory and products.

The uses include improving the platform, developing updates, and improving the business models and services of Marq Vision, its subsidiaries, its vendors or its other customers. Anonymized data is excluded from confidential information at section 5(d) and owned by Marq Vision under section 4.1. Nothing addresses separation between customers, and neither document gives a retention or deletion position beyond general reasonableness.

Privilege and work product are not mentioned. Section 1.2 disclaims any attorney client relationship, but that leaves the question open for a customer whose in house counsel routes enforcement through the platform.

The agreement draws the line clearly and the marketing crosses back over it. Section 1.2 of the terms, headed Authorization; No Legal Representation, says the services include no legal representation and that Marq Vision and its staff do not and cannot give legal guidance or advice. It says the services never create an attorney client relationship, that soft notices are not a substitute for legal claims, and that legal questions should go to the customer's own counsel.

Where offline services involve proceedings, third party law firms or agencies bring them. That is a complete statement of what the product is and is not, backed by the customer warranties in section 8.3 requiring due care in review and a good faith belief before any claim is reported. Against it, the anticounterfeit page claims the AI driven solutions comply fully with the ABA's ethics guidelines, ensuring the highest standards of legal integrity in every action.

The American Bar Association's rules govern lawyers, and the vendor's own agreement says it is not acting as one. No jurisdiction limits are stated.

One transparency commitment is published and nothing stands behind it. The anticounterfeit page says a customer will always have full visibility into when and how AI is used to enforce and protect their brand, which is a governance principle rather than a feature. The impersonation material adds that every classification carries its reasoning and a risk score, which explains individual decisions. Nothing else is published.

There is no responsible AI page, framework or set of principles, no person or function named as accountable for model behavior, nothing on what is tested before a model change ships, and no certification such as ISO 42001. Nothing addresses uneven output. That is a live question for a system working across 118 countries and 1,500 channels, where detection quality plausibly varies by language, script, marketplace and product category. Here a false positive is not an inconvenience but a takedown against a real business.

A trust center exists and almost nothing beneath it is stated in a way a reviewer could test. Section 1.3 of the terms commits Marq Vision to using information security techniques consistent with good industry standards, and the privacy policy states that all appropriate technical, organizational and administrative security measures are used, both of which are reasonableness language rather than a described control set.

Retention is acknowledged and never quantified, with the privacy policy saying data is kept no longer than reasonably necessary and the terms containing no deletion obligation for customer property on termination. No subprocessor is named anywhere: the privacy policy describes agents and affiliated businesses generically. No incident or breach notification practice is published on any surface. Access control is addressed only as the customer's responsibility to protect its own password.

A trust center hosted on Vanta is linked from the footer and would ordinarily carry the control detail; it renders in the browser and returned no readable content, so it may say more.

A complete allocation is published, and the exposure the product creates sits with the customer. Section 10.1 caps Marq Vision's total liability at the amounts paid in the twelve months before the event, and section 10.2 excludes consequential damages for both sides, with indemnification obligations carved out. Section 9.1 gives the customer an indemnity for third party claims that the platform infringes intellectual property.

Section 8.2 warrants compliance with law, noninfringement of the platform and absence of malicious code. The catch is what the indemnity excludes. Section 9.1(d) removes any claim arising from the customer's use of the services to request removal of content or send soft notices, which is the core function of the product. Section 9.2 requires the customer to indemnify Marq Vision for actions taken at its instruction and for any consequence of a party being wrongly left off the whitelist.

Section 8.4(b) says Marq Vision is never responsible for enforcement against an entity that should not have been targeted if it was not properly whitelisted. Wrongful takedown liability, which the vendor's own material names as the central risk, is allocated to the brand.

No integration into the systems legal or IP work already lives in is published, and none is claimed. No IP management or docketing system is named, no case management or document management system, no connector list, no API or developer documentation, and no integrations page exists in the navigation or the footer. What the platform does connect to is the enforcement estate rather than the customer's own systems, with the privacy policy noting that third-party account credentials may be provided so that account information transmits into the customer's MarqVision account, which is marketplace access rather than workflow integration.

Trademark management is delivered through a separate product on a separate domain, and nothing describes data moving between it and the enforcement platform.

Nothing published addresses where the platform runs or how customers are separated within it. No cloud provider is named, no hosting region or country is stated, no residency option is offered, nothing separates processing from storage, and no tenancy or isolation model is described. The only geography published is the office list: Los Angeles, San Francisco, New York, Seoul, Paris, Shanghai and Tokyo. The only cross border statement is the privacy policy's reference to the EU-U.S. Privacy Shield Framework.

That reference is itself a problem, since Privacy Shield was invalidated in 2020 and replaced in 2023, and the policy carrying it took effect in February 2026. For a platform operated from Delaware with offices in Seoul, Shanghai and Paris, processing brand and seller data across 118 countries, the lack of any residency statement is a material gap.

A trust center is linked, and no attestation is stated on any page that renders. The footer links directly to a Vanta trust center for MarqVision, a real route rather than a badge image, which suggests a monitored compliance program. Nothing else a reviewer needs is published. No standard is named on the site itself: neither SOC 2 nor ISO 27001 appears on the home page, the product pages, the terms or the privacy policy, so a buyer learns of any certification only by leaving the site.

No auditor, certificate number, report period, scope or trust services criteria is published, and no penetration testing or vulnerability program is described. The Vanta portal renders in the browser and returned no readable content, so it may say more. The vendor's own pages state no certification.

The capabilities are named in proprietary terms and nothing identifies what sits underneath. The published vocabulary is extensive: Atomic Product Detections using generative AI, a multi agent AI that decides whether an infringement is valid, an LLM framework that generates keywords in any language, image recognition and text analysis, and models said to be trained on successful enforcement data. None is attributed. No model, version or provider is named, and nothing says whether any component is built in house or reached through a third party, which a brand handing over its product catalog and seller data would ask.

No subprocessor list exists in the terms, the privacy policy or anywhere else; the privacy policy describes third parties only as agents and affiliated businesses. Nothing says where inference runs, and there is no commitment to notify customers when the models change. Trademarked capability names stand in for provider identification.

The pricing structure is published in the terms in detail, and no figure appears anywhere. Fees sit on an order form. Section 3.1 says they are based on services purchased rather than actual use and are not refundable, so this is a committed subscription rather than a consumption model. Under section 7.4 the term renews automatically for twelve month periods, with thirty days notice to end it. Two escalation terms are published.

Section 3.5 says fees may be revised on renewal, resulting in an increase every renewal term. Section 3.6 caps increases on agreements longer than a year at seven percent on each anniversary unless the order form says otherwise. Late payment carries one percent a month, with suspension possible after fifteen days. Coverage is itself commercial: the order form sets the regions, and section 8.4(a) keeps discretion over which marketplaces come first.

No number, tier or unit is published, no pricing page exists, and the entry routes are a free brand scan and a demo request.

Coverage is documented by industry and by right rather than by buyer role, and on its own terms it is thorough. Seven industry pages cover beauty, fashion, automotive, pharmaceuticals, food and beverage, network marketing and natural health products, each with its own counterfeiting patterns. Six solution lines are documented separately. The anticounterfeit page says the platform protects trademarks, copyrights, patents and designs.

Reach is given as more than 1,500 channels across 118 countries, with offices in seven cities and the site in English, Korean and Japanese. What is missing is the buyer. No page addresses legal, IP counsel or brand protection teams as roles. The people named in customer stories are mostly in commerce and sales, with one general counsel, so a reader cannot tell whether the legal function or digital commerce buys the platform. No portfolio size is stated and nothing says where coverage stops.