Huski.ai vs MarqVision: how they compare in 2026

The image recognition model is the product and the founding insight, not a layer over an existing service. The company was built in 2020 around applying visual similarity matching to trademark work, and the vendor states the technology eliminates the need for design search codes, which is the precise claim: the USPTO design coding system exists because computers could not compare images, and a model that compares them directly replaces the taxonomy rather than accelerating a search within it. The same recognition engine runs across all four products, matching filed applications in watch, clearing candidate designs in search, and identifying infringing product listings in e-commerce including images the vendor states are blurred or obscured. Remove the model and there is no search, no watch and no brand protection, only a docketing tool. Distinguished from Corsearch at B in this category: Corsearch is a decades old search and services business that added AI, and this vendor exists because of it.

The models are the product and there is nothing conventional beneath them. What MarqVision sells is detection and enforcement at a scale no human team can reach: scanning more than 1,500 channels across 118 countries, comparing findings against a brand's own product data at SKU level, analysing video frame by frame, and generating the notices that follow. The named capabilities are all machine learning: Atomic Product Detections described as using generative AI for SKU-level precision, a multi-agent system that decides whether an infringement is valid and classifies the enforcement grounds, an LLM framework that auto-generates search keywords in any language, image recognition and text analysis, and models the company states are trained on successful enforcement data. There is no content asset underneath, since the corpus is marketplace listings the crawler gathers and only the models make usable, and the stated comparator is not other software but the manual monitoring the product replaces. Verified 2 September 2026.

The mechanism is described and demonstrated, and nothing measured is published. Independent coverage from a named legal technology journalist describes running an image search against the USPTO database and receiving visually similar results within seconds, and reproduces a worked example in which the tool identified a filed design as similar to an existing Louis Vuitton mark. A demonstrated example in third party coverage is better evidence than a claim, and it is credited as such rather than as a measurement. What is absent is any figure: no precision or recall on visual similarity matching, no false negative rate on clearance, no comparison against design search code retrieval which is the incumbent method the vendor says it replaces, no evaluation and no test set. The gap is consequential for the same reason recorded on Corsearch: in clearance the failure mode is a silent miss, and a confusingly similar mark the model does not return produces an opinion that is wrong and looks complete. Checked the home page, the solutions pages, the FAQ and independent coverage on 29 Aug 2026.

Accuracy is measured, published, and unusually well contextualised, and one clause in the agreement undercuts all of it. Three figures appear across the product lines: 97 per cent on enforceable item detection using SKU-level comparison, 99.8 per cent across more than 48,000 domain impersonation incidents, and 94 per cent SKU-matching accuracy on unauthorised sales, with the company stating openly that these are internal benchmarks. The mechanism behind them is described rather than asserted, covering SKU-level comparison against genuine product data, multi-signal detection and a tiered model routing borderline cases to analysts. Failure modes are named in more detail than any vendor in this corpus manages, in a published piece that sets out what a false positive costs and names misrepresentation liability under 17 U.S.C. 512(f), tortious interference and platform reporting suspension as the consequences. What holds this below the top band is that only one figure carries a sample size and none carries a test set, and more importantly that section 2.4 of the operative terms states that Marq Vision assumes no liability for the quality, accuracy or validity of the data gathered in or by the Platform and that in no event shall the customer rely on it. The marketing measures accuracy and the agreement disclaims reliance on it. Verified 2 September 2026.

The professional is structurally in the loop and no oversight model is published. The product is sold exclusively to trademark attorneys, paralegals and legal professionals, with the vendor stating plainly that IP attorneys and their staff are the end users, so every output reaches a qualified reviewer by design rather than by policy. Output is framed as input to professional work throughout, with search results feeding a report and a customised opinion letter the attorney issues, and watch results feeding a decision on whether to file a Letter of Protest or opposition. What is not published: no statement of what runs unattended, no confidence indication on a visual similarity match, no threshold governing what surfaces in a watch alert, and no description of what happens when the recognition engine is uncertain about a blurred or obscured listing image in brand protection, where an automated conclusion could support an enforcement action against a third party seller. Structural oversight is real and is not a substitute for a published model.

Real review surfaces are published and the threshold at which the system acts alone is not. On the oversight side the material is substantive: the company states that a human reviewer confirms every significant decision, describes a tiered human-in-the-loop model that routes only borderline cases to analysts, and returns the reasoning and a risk score with each classification so a team can see why a determination was made. Two customer-side controls are contractual rather than optional. Section 8.3 of the terms requires the customer to have qualified representatives using due care to review, validate and verify all information transmitted, and to report claims only where it holds a good faith belief in their veracity. The whitelist gives the customer a standing exclusion list of sellers it does not wish to enforce against. Against that sits the speed the product is sold on, with 15-minute detection-to-enforcement, median responses under five hours for fake domains, response times described as seconds for piracy, and enforcement actions initiated automatically. Both statements cannot be fully true at once, and what the conflict shows to be missing is the definition: significant and borderline are never defined, so nothing published states which determinations reach a person. Verified 2 September 2026.

Independent journalism and a demonstrated product, with no customer evidence at all. The strongest element is coverage by a named legal technology publication that reviewed the product at launch, described the founders and their reasoning, ran the search itself and published a worked example of a result, which is independent verification that the core capability functions as described. Company facts are checkable: founded 2020, based in Middletown, Delaware, with named cofounders. Independent directory and review listings exist. Against that: no law firm or customer is named anywhere in located material, no case study, no usage figure, no funding announcement and no award were found, and no outcome or efficiency claim is published with a figure beyond the pricing comparison. Held at C rather than D because independent hands on coverage is a materially better evidence base than the self authored material that left IP Author at D in this category, and below B because nothing establishes adoption.

The reference base is one of the largest in the index and spans categories that rarely appear together. Roughly forty logos run across automotive with Nissan and GM, electronics with Panasonic, consumer goods with Henkel, Persil and Miele, beauty and fashion with Lush, Wella, Stussy, Courreges, Lemaire and Ader Error, Korean commerce and finance with Coupang, Shinsegae, Toss and KT&G, and gaming and media with Smilegate and Class101. Six case studies are published and five individuals speak with name and title, including 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 published in quantity: 180 times faster enforcement for its own teams, 50 per cent faster than traditional providers, 15-minute mean time to resolution, more than ten million sellers tracked, dozens of offline raids in Shenzhen, Guangdong and Yiwu, and one million dollars of counterfeits seized. What holds this below A is method and dating: no figure carries a measurement basis or a date, and none is tied to a named customer. Verified 2 September 2026.

Nothing located. The product holds candidate marks and unlaunched brand designs uploaded for clearance, which are commercially sensitive before filing, and it generates opinion letters, which are the archetypal attorney work product in trademark practice and the document that becomes contested in later infringement litigation on the question of what the client knew before adoption. Nothing published addresses attorney client privilege, work product, the confidentiality of a clearance search or of a generated opinion letter, or the conflicts question that arises when one platform serves many firms clearing marks against each other's clients. The vendor's exclusive focus on legal professionals makes the omission more surprising than it would be for a general enterprise tool. Checked the home page, the solutions pages, the FAQ and the site navigation on 29 Aug 2026.

One limb is genuinely strong and the rest are thin or point the wrong way. Section 5 of the terms is a proper mutual confidentiality regime with a compelled-disclosure notice commitment, and it survives termination. Section 4.2 confirms the customer owns all data it transmits and that no licence is granted beyond providing the services, with section 4.3 limiting Marq Vision's licence to the term and to service provision. Against that, section 2.5 grants a broad right to use customer-submitted information, expressly including information about the customer's customers, distributors, inventory and products, for improving the platform, developing updates and improvements, and improving Marq Vision's, its subsidiaries', its vendors' or its other customers' business models and services. Anonymised data is carved out of confidential information entirely at section 5(d) and is owned by Marq Vision under section 4.1. Nothing addresses segregation between customers, no retention or deletion position for customer data appears in either the terms or the privacy policy beyond a general reasonableness statement, and neither privilege nor work product is mentioned, which is defensible given section 1.2 disclaims any attorney-client relationship but leaves the question unanswered for a customer whose in-house counsel routes enforcement through the platform. Verified 2 September 2026.

A clear scope statement exists and no professional framework sits behind it. The vendor states unambiguously that Huski is for trademark attorneys, paralegals and other legal professionals, that IP attorneys and their staff are the end users, and that clients benefit indirectly, which is an explicit restriction of the user population to qualified professionals and the reason this grades above the D recorded on most of this index. Selling only to lawyers is a real answer to the unauthorised practice question, since the tool never reaches the public. What is absent is everything else: no statement that generated output including the customised opinion letter is not legal advice, no positioning on the attorney's duty of competence over a machine generated opinion letter issued under their own name, and no engagement with any bar guidance. Generating an opinion letter in a few clicks is the sharpest version of that question in this category, and it is not addressed.

The agreement draws the line clearly and the marketing crosses back over it. Section 1.2 of the terms, headed Authorization; No Legal Representation, states that the services do not include any legal representation, that Marq Vision and its personnel do not and cannot provide legal guidance or advice, that in no event shall the services constitute the creation of any potential or actual attorney-client relationship, that soft notices are not a substitute or replacement for any legal claims, and that any legal questions should be directed to the customer's own counsel. It also states that where offline services involve proceedings, those are brought by third-party law firms or agencies. That is a complete statement of what the product is and is not, and it is reinforced by the customer warranties at section 8.3 requiring due care in review and a good faith belief before any claim is reported. What pulls against it is a marketing claim on the anti-counterfeit page that 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 this vendor's own agreement says it is not acting as one, so the claim asserts compliance with a framework the terms say does not apply to it. No jurisdiction limits are stated. Verified 2 September 2026.

Nothing published about how the models are governed, evaluated or monitored. No AI policy, no model card, no bias or fairness testing, no evaluation methodology, no accuracy monitoring, no drift statement, no named governance body, no ISO 42001 and no EU AI Act positioning were located. The untested risk is specific to image recognition and worth naming rather than listing generically: visual similarity models are known to perform unevenly across image types, and this engine is applied to logos, product photographs and deliberately obscured listing images alike, with results feeding both clearance opinions and enforcement actions against third party sellers who never see the model. Nothing published indicates that performance across those conditions has been examined. Checked the home page, the solutions pages, the FAQ, the published rankings material and the site navigation on 29 Aug 2026.

One transparency commitment is published and no machinery stands behind it. The anti-counterfeit page states that 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 and is more than most records in this corpus offer. The impersonation material adds that every classification carries its reasoning and a risk score, which is explainability at the level of an individual decision. Beyond that nothing was located. No responsible AI page, framework or set of principles exists, no individual or function is named as accountable for model behaviour, nothing describes what is evaluated before a model change ships, and there is no certification such as ISO 42001. Nothing anywhere addresses uneven output, which is a live question on a system operating across 118 countries and 1,500 channels where detection quality plausibly varies by language, script, marketplace and product category, and where a false positive is not an inconvenience but a takedown against a real business. Verified 2 September 2026.

No stewardship position located. Nothing states whether uploaded candidate marks, logo images, client portfolios or search histories are used to train or improve the recognition models, no retention period is published, and no deletion right is described. The gap has a distinctive edge on this product: the platform ingests unfiled logo designs at the clearance stage, which are both commercially sensitive and precisely the kind of labelled image data that improves a visual similarity model, so the incentive alignment is visible and the position is unstated. Docketing additionally holds entire client portfolios uploaded in one click, and the Customs and Border Protection integration means recordation status data flows through the platform. Checked the home page, the solutions pages including docketing and brand protection, the FAQ and the site navigation on 29 Aug 2026.

A trust centre 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, organisational 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 was located on any surface. Access control is addressed only as the customer's responsibility to protect its own password. A Vanta-hosted trust centre is linked from the footer and would ordinarily carry the control detail; it renders client-side and returned no readable content, which is recorded as a retrieval limit rather than an absence. Verified 2 September 2026.

No published position located. Nothing was found on liability for AI output, warranty, service levels or remedy. Three distinct exposures arise from this product and none is addressed. A clearance search that misses a confusingly similar mark supports an opinion letter the attorney signs, and the opinion letter is the artifact a client relies on to launch a brand. A docketing system that fails to surface a deadline causes a lapsed registration, which is a classic malpractice claim and is not covered by any published service level. And a brand protection match that wrongly identifies a legitimate listing as counterfeit supports an enforcement action against a third party seller with no described recourse. Checked the home page, the solutions pages, the FAQ, the pricing material and the site navigation on 29 Aug 2026.

A complete allocation is published and the exposure the product actually creates sits entirely with the customer. Section 10.1 caps Marq Vision's aggregate liability at amounts paid in the twelve months before the occurrence, and section 10.2 excludes consequential damages mutually with indemnification obligations carved out. Section 9.1 gives the customer a real indemnity for third-party claims that the platform infringes intellectual property, and section 8.2 warrants performance in compliance with law, non-infringement of the platform, and absence of malicious code. That is more structure than most records here. The limitation worth a buyer's full attention is what the indemnity excludes. Section 9.1(d) removes from Marq Vision's indemnity any claim arising from the customer's use of the services to request removal of content or to send soft notices, which is the core function of the product, and 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 omitted from the whitelist. Section 8.4(b) states expressly that Marq Vision is in no event responsible for enforcement against an entity that should not have been targeted if it was not properly whitelisted. Wrongful takedown liability, the risk the vendor's own published material identifies as the central one, is allocated to the brand. Verified 2 September 2026.

One genuinely uncommon integration and none of the ordinary ones. The vendor states integration with US Customs and Border Protection to track recordation statuses, which connects the platform to the government system through which recorded marks are enforced at the border, and no other record on this index names a customs or enforcement authority integration. One click portfolio upload with automatic population of clients and trademarks from filing record details implies a working connection to USPTO filing data rather than manual entry. What is absent is everything a firm would ask about: no IP management or docketing system integration, which matters less here because this product is itself the docketing system, no document management system, no practice management connector, and no API or export documentation was located beyond single click report generation. Checked the docketing page, the solutions pages, the home page and the site navigation on 29 Aug 2026.

No integration into the systems legal or IP work already lives in was located, and none is claimed. Checked the home page, the anti-counterfeit, brand protection, content protection, impersonation and MarqLaw pages, the MARQ AI terms in full, the privacy policy in full and the complete site footer on 2 September 2026. 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. Verified 2 September 2026.

Nothing located. No hosting provider is named, no region or data residency commitment is published, and no deployment options are described. The vendor is a small United States company serving United States trademark practice against the USPTO, so residency is a narrower question than for the multinational vendors elsewhere in this category, and it remains unanswered for a platform holding unfiled logo designs and complete client portfolios. Checked the home page, the solutions pages, the FAQ and the site navigation on 29 Aug 2026.

Nothing published addresses where the platform runs or how customers are separated within it. Checked the home page, the anti-counterfeit, brand protection, content protection, impersonation and MarqLaw pages, the MARQ AI terms in full and the privacy policy in full on 2 September 2026. No cloud provider is named, no hosting region or country is stated, no residency option is offered, nothing distinguishes processing from storage, and no tenancy or isolation model is described. The only geography published is corporate, being offices in Los Angeles, San Francisco, New York, Seoul, Paris, Shanghai and Tokyo, and 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 for this axis rather than an answer to it, 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 absence of any residency statement is a material gap. Verified 2 September 2026.

No certification, attestation, trust centre or security page was located. No SOC 2 of either type, no ISO 27001, no named auditor, no penetration testing partner, no encryption statement and no security documentation of any kind was found across the pages read. Under the three tier test the artifact is absent rather than gated. Two records in ip-and-patents now sit at D on this axis, this one and Corsearch, and both are the trademark vendors, while all five patent vendors publish at least one named certification. Recorded as an observation rather than a conclusion, since two records is a small sample, and flagged for the category note. Held at D rather than higher because unlike IP Author, where a published Trust Center existed but could not be read, no trust surface of any kind was located here to hold a grade open for. Checked the home page, the solutions pages, the FAQ, the pricing page and the site navigation on 29 Aug 2026.

A trust centre is linked and no attestation is stated on any page a reader can render. The footer carries a direct link to a Vanta-hosted trust centre for MarqVision, which is a real route rather than a badge image, and its existence indicates a compliance programme being monitored. What is absent is everything the reader needs from it. No standard is named anywhere 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 a 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 programme is described. The Vanta portal renders client-side and returned no readable content on 2 September 2026, which is recorded as a retrieval limit on this reading rather than as an absence on the vendor's part, and the grade rests on what the vendor states on its own surfaces, which is nothing. Verified 2 September 2026.

Nothing located. No foundation model provider, model family or version is named, no distinction is drawn between proprietary and third party models, and no subprocessor list was found. The vendor describes advanced image recognition and AI powered brand recognition throughout, which characterises capability rather than provenance, and given the image recognition engine is the entire technical proposition, whether it was built in house or licensed is the single most material supply chain question about this product and it is unanswered. Checked the home page, the solutions pages, the FAQ and independent coverage on 29 Aug 2026.

The capabilities are named in proprietary terms and nothing identifies what sits underneath. The published vocabulary is extensive: Atomic Product Detections described as using generative AI, a multi-agent AI that determines whether infringement is valid, an LLM framework auto-generating keywords in any language, image recognition and text analysis, and models stated to be trained on successful enforcement data. Not one of those is attributed. No model is named, no version, no provider entity, and nothing states whether any component is built in-house or reached through a third party, which is the question a brand handing over its product catalogue and seller data would ask. No subprocessor list exists in the terms, the privacy policy or anywhere else, the privacy policy describing third parties only as agents and affiliated businesses. Nothing states where inference runs, and no commitment exists to notify customers when the model set changes. Trademarked capability names in place of provider identification is what this band describes. Verified 2 September 2026.

THE SECOND A ON THIS AXIS IN 54 RECORDS, after Descrybe, and the only one in ip-and-patents. Everything a buyer needs is published without contacting anyone: a stated entry price of $299 per month, a dedicated pricing page, the unit of charge stated as unlimited text and image clearance searches rather than metered per search, discounts stated for twelve month commitments, and explicit terms of no contracts and no credit card required to start. The vendor also positions its price against the market openly, and independent coverage corroborates the comparison, noting that competing trademark search tools commonly charge $100 to $800 per search. A firm can calculate total cost of ownership and compare it against incumbents from published material alone. Graded A because that is the top of this axis. Held short of a perfect record because the full tier structure above the entry price was not read in this pass and it is unclear which of search, docketing, watch and brand protection the entry price includes.

The structure is published in the terms in real detail and no figure appears anywhere. Fees sit on an order form and are stated at section 3.1 to be based on the services purchased rather than actual usage, and to be non-refundable, which tells a buyer this is a committed subscription rather than a consumption model. The term renews automatically for successive twelve month periods under section 7.4 with thirty days notice of non-renewal. Two escalation provisions are published and one is unusually candid: section 3.5 states plainly that fees may be revised on renewal, resulting in an increase every renewal term, and section 3.6 caps increases on multi-year agreements at seven per cent on each anniversary unless the order form says otherwise. Late payment attracts one per cent per month with suspension available after fifteen days. Coverage scope is itself commercial, with the order form specifying regions and section 8.4(a) reserving discretion over which marketplaces are prioritised. What is absent is any number, tier or unit: no pricing page exists, and the entry routes are a free brand scan and a demo request. Verified 2 September 2026.

Lifecycle coverage is complete and jurisdictional coverage is narrow and largely unstated. Within its scope the product spans the whole trademark lifecycle in one platform: clearance search, docketing with deadline management, watch on new filings, Office Action handling, and post registration enforcement through e-commerce brand protection, which is broader than most single vendors in this category attempt. E-commerce coverage is characterised specifically, naming Shopify powered independent stores as a claimed unique capability with the reasoning stated, that Shopify has no centralised marketplace and over three million stores operate on custom domains. Held at C because jurisdictional coverage appears to be United States only: the USPTO is the sole trademark office named across search, watch and docketing, the Customs and Border Protection integration is United States specific, and no international registry, Madrid Protocol handling or foreign filing coverage was located. Compare Corsearch at A in this category with 190 registries and 1,100 databases stated. A single office platform is a legitimate design choice and is a real coverage limit for any firm with international clients.

Coverage is documented by industry and by right rather than by buyer role, and on its own terms it is thorough. Seven industry pages are published for beauty, fashion, automotive, pharmaceuticals, food and beverage, network marketing and natural health products, each a category with distinct counterfeiting patterns. Six solution lines are separately documented. The rights covered are stated explicitly, with the anti-counterfeit page saying the platform protects trademarks, copyrights, patents and designs, and geographic reach is given precisely as more than 1,500 channels across 118 countries with offices in seven cities and the site published in English, Korean and Japanese. What is missing is the buyer. There is no page addressed to legal, IP counsel or brand protection teams as roles, and the named individuals in customer stories are overwhelmingly commerce and sales, with a single general counsel among them, so a reader cannot tell from the site whether the platform is bought by the legal function or by digital commerce. No portfolio size band is stated and nothing describes where coverage stops. Verified 2 September 2026.

Silent. The quoted line is the vendor's clearest statement about who the platform serves and it says nothing about what happens to what they upload. No statement in either direction was located on whether candidate marks, uploaded logo images, client portfolios or search histories are used to train or improve the recognition models. The incentive alignment is unusually visible here and is why the silence is recorded pointedly: the core technology is a visual similarity engine, and unfiled logo designs uploaded for clearance are precisely the labelled image data that improves such a model. Docketing additionally ingests entire client portfolios in one click. Recorded as silent, not as a negative commitment. Checked the home page, the FAQ, the solutions pages including docketing and brand protection, and the site navigation on 29 Aug 2026.

Training is described in product material and is not addressed in the agreement, which is what puts it in this value rather than a contractual one. The content protection page states the quoted position, describing an anti-piracy method built on models trained on successful enforcement data. A published article on false positives goes further, describing a tiered human-in-the-loop model in which borderline cases route to analysts and their decisions are then fed back into training to shrink the review queue over time. Enforcement data and analyst decisions are generated in the course of work performed on customer matters using customer-supplied product data, so the material customers contribute is in the training loop even though neither statement uses the phrase customer data. The agreement does not name training at all. Section 2.5(b) of the MARQ AI terms grants a broad right to use information the customer submits, expressly including information about the customer's customers, distributors, inventory and products, for improving platform performance, developing updates and improvements, and improving Marq Vision's and its subsidiaries', vendors' or other customers' business models and services, and separately allows any anonymised data to be used for developing new products. That is a wide improvement right that does not name training, so it does not by itself move the value.

Not addressed. No retention period is published for uploaded images, search queries, generated reports and opinion letters, or docketed portfolio data, and no deletion right is described. Retention is implicit in the product design rather than stated as a policy: docketing is a system of record that holds client portfolios and deadlines by function, and generated reports and opinion letters persist for later use, and nothing states for how long, under what terms, or what happens when a firm stops subscribing, which matters for a month to month product sold with no contract. Checked the docketing page, the home page, the FAQ and the site navigation on 29 Aug 2026.

Retention is acknowledged and never quantified. The privacy policy states that personal information is kept as long as necessary to fulfil the purpose for which it was collected or to comply with legal or regulatory requirements, and no longer than reasonably necessary, with no period attached to any category. The terms are silent on the point in both directions: section 7.3 sets out the effect of termination as the licence ending and the customer ceasing use and destroying documentation, and imposes no deletion or return obligation on Marq Vision for customer property. Nothing states how long detection records, evidence packages, seller intelligence or enforcement history are held, and no retention setting is offered. Two provisions push the other way and belong on the record. The privacy policy states that some information may remain in records after account deactivation, and that aggregated data derived from or incorporating personal information may continue to be used even after the customer updates or deletes it. Section 4.1 of the terms confirms Marq Vision owns anonymised data outright.

Not addressed. No permission model, client level access restriction or tenant segregation description was located. The vendor describes a dashboard giving a bird's eye view of all clients, their trademarks, related office actions and detected risks, which is a firm wide view across the whole book of business and raises rather than answers the question of whether access can be restricted by client or matter within a firm. The cross customer question is equally open: this platform serves many trademark firms clearing marks against each other's clients, and nothing describes what separates one firm's searches and portfolios from another's. Checked the docketing page, the watch page, the home page and the site navigation on 29 Aug 2026.

Checked the home page, the anti-counterfeit, brand protection, content protection, impersonation and MarqLaw pages, the MARQ AI terms in full and the privacy policy in full on 2 September 2026. Nothing describes segregation between customers or isolation of one brand's data from another's. No tenancy model is stated and no access boundary is described beyond the customer's own obligation to protect its password. The question has a specific edge on this product that the published material does not reach. Section 2.5(b) of the terms permits use of information a customer submits, including data about its customers, distributors, inventory and products, to improve the business models and services of Marq Vision's other customers, and the platform maintains a global seller database of more than ten million entries built from enforcement work. A brand would reasonably want to know what separates its catalogue and channel intelligence from a competitor's on the same platform, and nothing published answers that.

Not addressed. No government or law enforcement request clause, no commitment to notify a customer before producing their data, and no transparency report were located. The question has a concrete edge on this record that it lacks elsewhere: the platform integrates directly with US Customs and Border Protection to track recordation statuses, which is an operational relationship with a federal enforcement agency, and nothing published describes what data flows in that direction or what would happen if an agency requested more. Checked the docketing page, the brand protection page, the home page and the site navigation on 29 Aug 2026.

The commitment is in the operative agreement and it is more specific than most. Section 5 of the MARQ AI terms permits disclosure of confidential information where required by valid court order, judicial process or regulatory authority, and then requires the receiving party to use best efforts to preserve confidentiality and to give the quoted notice promptly, unless restrained from doing so by court order. Naming the nature, scope and contents as what must be disclosed to the customer is more than the bare notification most records here offer. The obligation is mutual and survives termination. Two limits belong on the record. Anonymised data is excluded from the definition of confidential information at section 5(d), so the notice commitment does not reach it. And the privacy policy takes a different and broader position for personal information, reserving the right to access, read, preserve and disclose any information Marq Vision believes necessary to comply with law or court order, to enforce its agreements, or to protect its own rights, property or safety, with no notice commitment attached.

Named and public, with no licence basis stated and none obviously needed. The searched corpus is identified specifically as the USPTO trademark database, including the design and image records, which is a single named public authority rather than an aggregated set of undisclosed databases, and the vendor's central technical claim is built on it, being that image recognition replaces the USPTO design search code taxonomy. Naming one authoritative source is a clearer provenance position than a large unenumerated database count, and it is credited. The e-commerce side names monitored platforms including Shopify powered independent stores. What is absent: no statement of the licensing or access basis for USPTO data, no update lag or refresh frequency despite watch depending on catching applications at filing, and no statement of how e-commerce listing data is obtained or how current it is.

The sources are identified by platform and scale and no rights basis is stated for any of them. Coverage is published as more than 1,500 marketplaces, social media platforms, websites and app stores across 118 countries, with individual platforms named throughout the product material including Amazon, Alibaba, Taobao, TikTok, BiliBili, YouTube, Google and Meta, and a global seller database of more than ten million entries. The customer's own genuine product data is the comparison set for SKU-level matching. What is absent is the legal basis on which any of it is gathered. Nothing states whether marketplace listings are collected under platform agreements, APIs, brand registry programmes or by crawling against those platforms' terms of use, and the distinction matters commercially as well as legally, since the company separately markets its marketplace partnerships in Southeast Asia and China and holds direct takedown authority on YouTube, which implies negotiated access on some channels and not others. No licence, agreement or permission is described for the underlying listing data anywhere.

Own treatment signal, and this CONFIRMS the patent versus trademark split identified on Corsearch. Status is a first class function of this product rather than an optional layer: docketing tracks live deadlines and critical dates across a portfolio with a daily email, the dashboard surfaces office actions and detected risks, portfolio data is automatically populated from filing record details, and Customs and Border Protection recordation statuses are tracked, which is status monitoring at the enforcement layer. Watch exists to catch newly filed applications at filing, which is currency in the most direct sense. Held at own treatment signal rather than higher because no citator or status authority is named as a distinct source, nothing describes how current the status data is or how often it refreshes, and no explicit statement says that dead, abandoned or cancelled marks are flagged as such in search results, which is the specific case that matters for a clearance opinion.

Checked the home page, the anti-counterfeit, brand protection, content protection, impersonation and MarqLaw pages, the MARQ AI terms and the privacy policy on 2 September 2026. No public material addresses subsequent history, treatment or good law checking, and none is claimed. The product does not retrieve primary law, so a citator has nothing to operate on and the limb does not bite in its usual form. The analogous currency question in this product class is whether the intellectual property right being enforced is still valid and in force, and it is addressed only as the customer's responsibility rather than as a system check: section 8.3(d) of the terms requires the customer to warrant that it is the rightful owner or licensee of all rights subject to the services, and section 8.4(b) states that the sufficiency of the customer's own registrations is outside Marq Vision's control. Nothing describes the platform verifying a registration's status before enforcing on it.

Not addressed, and the omission is most consequential on the capability the vendor markets hardest. Nothing published describes what happens when the recognition engine is uncertain: whether a marginal visual similarity is flagged, whether a confidence score accompanies a match, whether a watch alert distinguishes a strong conflict from a weak one, or whether the system will report that it could not reach a conclusion. The brand protection claim sharpens it, since the vendor states the engine finds counterfeits even in blurred or obscured product listing images, which is precisely the condition under which a visual model is least certain and most likely to be wrong in both directions, and no confidence signal is described for output that can support an enforcement action against a third party seller. Checked the brand protection page, the watch page, the search material, the FAQ and independent coverage on 29 Aug 2026.

Confidence is scored and surfaced, and nothing describes the system declining. The impersonation material states that every classification produced by the multi-agent system comes with its reasoning and a risk score, so a user sees both a rationale and a graded confidence rather than a bare determination. A published article on false positives adds that a tiered human-in-the-loop model routes only borderline cases to analysts, which is a confidence threshold operating on the workflow even though the threshold itself is never defined. What is absent is any account of what the system does when it cannot reach a determination at all, whether it abstains, holds an item, or defaults to escalation, and nothing states what confidence level is required before an automated enforcement action proceeds. The customer-maintained whitelist is a suppression mechanism supplied by the customer rather than a behaviour of the model.

None located, with the instrument named. General web searches combining the vendor and product names with court, order, sanction, negligence and clearance terms returned nothing on 29 Aug 2026, and no named docket database, trademark office record system or court record tracker was searched. Recorded as a statement about what this search found, not as a clearance. The exposure shape matches Corsearch rather than the patent drafting tools: no legal authority is generated, so the analogous adverse finding would be a proceeding contesting the adequacy of a clearance search or an opinion letter, or a wrongful enforcement action arising from a false positive brand protection match, neither of which surfaces as a sanctions order.

Searched the AI Hallucination Cases database maintained by Damien Charlotin, and reporting drawing on it, on 2 September 2026 on the product and corporate name MarqVision and Marq Vision Inc. No court order, opinion or disciplinary record naming the product was located. This is a statement about the public record rather than a finding about the product. The signal fits this product class poorly and the reason is worth recording: MarqVision generates takedown notices and enforcement packages rather than legal citations, so its characteristic failure is a wrongful enforcement action against a legitimate seller rather than a fabricated case reference in a filing. The vendor's own published material identifies that risk directly, naming misrepresentation liability under 17 U.S.C. 512(f), tortious interference and platform reporting suspension as the consequences of false positives. Those outcomes surface as civil claims or platform sanctions rather than as entries in a hallucination tracker, and no tracker indexes them.

Not addressed. No named ethics opinion, no USPTO Rules of Professional Conduct reference, no 37 CFR citation, no duty of competence discussion and no bar guidance was located. The absence is notable on this record specifically because the vendor sells exclusively to legal professionals and generates a customised opinion letter, which is a document issued under an attorney's name and professional judgement, and it engages the duty of competence more directly than a search result does. Sixth of seven records in this category at this value, with Patlytics the only exception. Checked the home page, the FAQ, the solutions pages, the published rankings material and the site navigation on 29 Aug 2026.

An authority is named and no instrument is, and the claim sits awkwardly against the agreement. The anti-counterfeit page states that MarqVision's AI-driven brand protection solutions comply fully with the ABA's ethics guidelines, ensuring the highest standards of legal integrity in every action. Naming the American Bar Association puts this above a bare gesture at professional standards, which is why the lowest value does not fit. But no rule, model rule or formal opinion is identified, nothing states which obligations are said to be met or how compliance is assessed, and ABA Formal Opinion 512 on generative AI is not referenced. The claim also runs against section 1.2 of the company's own terms, which states that Marq Vision does not and cannot provide legal guidance or advice and that no attorney-client relationship arises, since the ABA's rules bind lawyers rather than software vendors. A reader is left with an assertion of compliance with a framework the agreement says does not govern the service.

Savings claims only, and unusually the savings are stated in money against a named market baseline rather than as a percentage of time. The vendor publishes that clearance searches which would otherwise cost hundreds of dollars each are included in a $299 per month unlimited subscription, and independent coverage corroborates the baseline, noting that competing tools commonly charge $100 to $800 per search. That makes the claim testable in a way a percentage never is. Its relevance to this signal is also unusual: search costs are typically disbursed to the client in trademark practice, so a shift from per search charging to a flat subscription changes what a firm can pass through and what it must absorb. Nothing published addresses that, or how AI assisted clearance time should be billed, and no exportable record shows what portion of a search or opinion letter was machine generated.

Time and cost savings are published with figures and nothing addresses the bill. The company states that Marq AI saves its expert teams 180 times the time previously required to enforce infringements, that enforcement runs 50 per cent faster than traditional brand protection companies, and that mean time to resolution is fifteen minutes, and it publishes a Saturation Rate metric and revenue recovery framing intended to show enforcement paying for itself. Nothing addresses how AI-assisted work is recorded, billed or disclosed, and no per matter record of AI-assisted work was located. The signal's usual direction is partly inverted here, since the buyer is the brand owner rather than a firm billing a client, so compressed enforcement time accrues to the customer rather than changing an invoice. It is not fully inverted: MarqLaw engages partner law firms who bill for litigation, customs and field investigation work supported by platform output, and nothing published describes how AI-assisted preparation is reflected in those engagements.

Not addressed. No trust centre, security page, named certification, subprocessor list, named model provider, data processing agreement or documentation request route was located, so a firm has nothing it could forward to a client and no destination to point one toward. Second record in ip-and-patents at this value alongside Corsearch, and both are the trademark vendors, while all five patent vendors reach at least on request. Checked the home page, the FAQ, the solutions pages, the pricing material and the site navigation on 29 Aug 2026.

Checked the MARQ AI terms in full, the privacy policy in full, the home page and the anti-counterfeit, brand protection, content protection, impersonation and MarqLaw pages on 2 September 2026. No subprocessor list exists in any form. The privacy policy describes third parties only by category, covering agents performing tasks on Marq Vision's behalf, affiliated businesses and third-party law firms, advertisers and partners receiving de-identified data, without naming a single entity. No model or AI provider is named anywhere on any surface. No data processing agreement is published, and no forwardable client-facing disclosure material was located. A Vanta-hosted trust centre is linked from the footer and would be the natural home for a subprocessor list; it renders client-side and returned no readable content, which is recorded as a retrieval limit rather than an absence. A brand asked by a partner or regulator which entities process its catalogue and seller data could not answer from anything published.

Partial record, and this product is built around producing an evidentiary artifact, which is uncommon. On the enforcement side the vendor states that evidence of infringement can be compiled into a report in a single click for use if litigation is pursued, and the founding motivation described in independent coverage was the manual burden of collecting listing screenshots for enforcement cases, so evidence capture is the problem the company was built to solve. On the clearance side a comprehensive report and a customised opinion letter are generated, and the opinion letter is precisely the document produced years later in infringement litigation on the question of what was known before adoption. The familiar two limbs are missing and one is sharpened. Nothing indicates that output records which model produced a similarity match or when, and no human verification record is captured showing the attorney reviewed it, which matters more here because the generated artifact is an opinion letter issued in the attorney's name.

This is the closest any record in this pull comes to the signal without meeting it. Evidence capture is a published product function rather than a by-product: the anti-counterfeit page states that the platform automatically extracts and stores the evidence needed to shut down illicit activity and generates full self documentation reports, MarqLaw is marketed on courtroom-ready evidence packages, the impersonation material describes a report-ready enforcement package prepared for each channel, and a published article states that consistent enforcement backed by a documented record of every action is what redirects counterfeiters elsewhere. The anti-counterfeit page also carries an AI transparency commitment, stating that a customer will always have full visibility into when and how AI is used to enforce and protect their brand. What is not described is the record this signal asks for: nothing states that the visibility is exportable, and nothing captures which model or agent made a given determination, on what confidence, or which human reviewer confirmed it, which is precisely what a party challenging a takedown or a 512(f) claim would seek.