I

InnovationQ+

The models are the engine of the product's core capability, layered on assets that retain value without them, which is the B band, though this sits at the deeper end of it. Retrieval itself is the model here rather than a feature bolted onto retrieval: Semantic Gist is a probabilistic language model combined with a deep neural network, and it is what turns a query into results. CompassAI sits above it as a distinct reasoning layer assessing novelty, similarity and technical differentiation.

So both halves of what a searcher does, finding and judging, run through named models. What holds this at B rather than higher is that the platform is also a content and analytics business. A buyer is paying for coverage of 154 patent authorities, more than 190 million IP data points, exclusive full-text IEEE access and IP.com's own Prior Art Database, none of which is model-dependent. Keyword, image and document-based search modes are offered alongside concept search, so conventional retrieval remains available.

And the scored reports have their own published logic, the Portfolio Intelligence Report resting on a fourteen-factor framework benchmarked against a peer group rather than on a generative step. Strip out the AI and a buyer still holds a large licensed corpus with keyword search and structured scoring over it, which is the B test. Verified 13 September 2026.

Grounding and retrieval method are documented more thoroughly than by any other vendor in this corpus, and no measured accuracy is published, which is the B band. The method disclosure is a published technical white paper, not a marketing page. It sets out the problem in retrieval terms, polysemy and synonymy in patent vocabulary; surveys the alternatives it rejected, naming latent semantic analysis and its successors and citing the evidence that they degrade at scale; and then describes the architecture actually used, being probabilistic language modelling for high-dimensional document representation combined with a deep neural network that compresses those into low-dimensional semantic signatures, with the two used in concert for coarse filtering and fine ranking, plus a multi-level concept layer using graph clustering and automatic n-gram phrase recognition.

It names the patents covering the work and carries a bibliography of the research it draws on. On the reasoning side, outputs are stated to be grounded in retrieved evidence rather than free-form generation, and every insight remains connected to its source documents for inspection. What is absent is the number. The vendor states that it benchmarks retrieval and reasoning against real-world IP use cases and evaluates on relevance, recall and expert-validated outcomes, so a measurement programme is described with its metrics named, and not one result from it is published. On a prior art tool, recall is the figure that matters, and the A limb asks for it. Verified 13 September 2026.

A written commitment that the models work alongside the user, with real review surfaces, short of the full control structure, which is the B band. The commitment is architectural rather than promissory, and the vendor argues it that way. Retrieval is deliberately separated from reasoning so that the reasoning layer can only work on evidence the retrieval layer has produced, described as reasoning from verified evidence within defined constraints, with safeguards embedded at the architectural level rather than added later.

CompassAI is stated to be designed to support expert judgment, not replace it, and outputs are described as guided rather than generative guesswork, grounded in evidence rather than free-form text. Review surfaces follow from that design: insights remain directly connected to the underlying source documents, users can inspect how conclusions were formed and validate them against the original material, and private interaction histories are retained so an organisation can revisit prior analyses.

One further oversight point is unusual and is graded on the bar guidance signal rather than here, being the vendor's stated care not to overstep USPTO inventorship guidelines. What the A band requires is not published. No threshold is stated at which any step proceeds without review, no confidence signal attaches to a retrieved result or a reasoning output, no mode distinction separates supervised from unsupervised operation, and nothing describes what happens when an output is found wrong. Verified 13 September 2026.

Customer types and platform figures are published without a single named customer, which is the C band. What exists is real but is about the product rather than about deployments. Coverage and performance figures are specific: 154 global patent authorities searchable in one query, more than 190 million IP data points across patent and non-patent literature, and a claim of 30 per cent faster per search. Buyer types are named by function, being engineering and R&D, IP strategy and IP professionals, and the vendor states that patent offices are among its customers and that its Prior Art Database is a required reference for international examiners under Rule 34 of the Patent Cooperation Treaty, which is institutional validation of the corpus if not of the platform.

What is missing is attribution. No customer is named anywhere on the estate. The one case study located, on uncovering critical prior art at a Fortune 500 company, is both anonymised and gated behind a lead-capture form, so a reader can neither identify the customer nor read the account without surrendering contact details. The 30 per cent figure carries no baseline, method or sample. Recorded and expressly not credited under ground rules section 3 and the naming determination for this record: the stronger outcome figures on the IP.com estate, being 90 per cent lower patentability evaluation costs and 80 per cent faster idea-to-report, belong to IQ Ideas+, a different product. Verified 13 September 2026.

Substantive published commitments on the limbs that bite hardest for this product, defeated on the rest, which is the B band. The training commitment is the strongest located in this lane and is drafted to close the obvious gap: customer data is never used to train IP.com's models **or any third-party models**. Extending the prohibition to third parties matters on a platform with a reasoning layer, because it forecloses the route by which customer material would otherwise reach an external provider.

Confidentiality is independently attested rather than merely asserted, the SOC 2 Type 2 examination expressly covering Confidentiality of the Software Services System and, in the vendor's own description, the protection of sensitive information from unauthorised disclosure across collection, creation, use, processing, transmission and storage. Interaction histories are described as private and secure. Two limbs fail and one is sharp.

Privilege and work product are not addressed by name, which under R33 forecloses A on its own. And retention is stated only as a fact, that histories are retained, with no period, no deletion right and no end-of-subscription position. That gap has particular weight here because the query the vendor invites is a draft invention disclosure or a full claim set, submitted before filing, which is novelty-destroying if it escapes and is frequently privileged in an attorney's hands. Verified 13 September 2026.

A real published position on where the tool stops and professional judgement begins, short of the supervision and competence dimension, which is the B band. The position is stated repeatedly and in the vendor's own voice. CompassAI is described as designed to support expert judgment, not replace it; the platform is said to keep experts in control and to provide transparency at every step; and outputs are framed as guided rather than generative, grounded in evidence so a professional can check them.

One sentence goes further than any comparable record in this lane by naming the specific professional boundary at risk: AI accelerates analysis without obscuring judgment or overstepping the inventorship guidelines established by the USPTO. Inventorship is precisely the question AI-assisted invention raises for patent practice, and the vendor engages it by name rather than gesturing at responsibility in general. The audience is named plainly too, covering R&D professionals, IP managers and legal counsel, so it is explicit that non-lawyers are intended users.

What the A band requires is absent. Nothing addresses supervision or competence: no statement of who within a customer should operate which capability, what review a supervising practitioner should apply to a prior art result, or what training is expected. No jurisdiction limit is stated. And the duty of disclosure of material prior art, the obligation this product most directly touches, is nowhere addressed. Verified 13 September 2026.

A published governance position with mechanisms behind it, short of testing results and a named owner, which is the B band. What lifts this above a principles page is that the safeguards are described as properties of the architecture rather than policies laid over it. Separating retrieval from reasoning is presented as the governance mechanism itself, constraining the reasoning layer to evidence already retrieved and thereby reducing hallucinations and unintended outputs by design, with the vendor stating expressly that safeguards are embedded at the architectural level, not added later.

Around that sit a Responsible AI page, a Responsible AI white paper, and a stated continuous improvement loop in which models are refined against performance data and expert feedback. Data governance is addressed concretely, with training exclusion extending to third-party models and organisational isolation on the platform. A measurement regime is described with its metrics named, being relevance, recall and expert-validated outcomes.

What the A band asks for is missing. No individual, committee or function is named as accountable for model behaviour. No external framework is adopted or cited, which is a real gap when comparable vendors in this lane name one. No pre-release testing regime is described beyond the general benchmarking statement, and no evaluation result is published. Bias is not addressed in any form, which matters on a retrieval engine whose synonym handling is learned from historical patent drafting conventions. Verified 13 September 2026.

Substantive published policy covering most of the ground, short of the full set, which is the B band. The strongest limb is the training and governance position, stated without qualification: customer data is never used to train IP.com's models or any third-party models. Security is independently attested rather than asserted, a SOC 2 Type 2 examination covering Security, Availability and Confidentiality of the Software Services System, described as testing both the design and the operating effectiveness of controls, with the examination scope stated to cover protection of information through collection, creation, use, processing, transmission and storage.

A public trust centre carries that summary. Design intent is documented: safeguards embedded architecturally, the reasoning layer constrained to retrieved evidence, and private interaction histories retained under access control. Three limbs fail. No subprocessor list is published and no processor is named, so a buyer cannot establish who else touches the data, and that gap sits awkwardly beside a training commitment that expressly contemplates third-party models.

No retention period or deletion commitment is published for queries or interaction histories. And no incident or breach notification commitment to customers was located, notwithstanding that the SOC 2 scope implies an incident programme exists. Verified 13 September 2026.

Integration is claimed within the vendor's own suite and no external connection is documented, which is the C band. What is established is internal. The Innovation Power Suite is presented as a set of products sharing the Semantic Gist retrieval layer, so InnovationQ+ sits alongside InnovationQ, IQ Ideas+ and the report products with common retrieval underneath, and the Patent Vitality Report is stated to run self-serve inside InnovationQ+ with results in minutes.

Content integration is real and is a genuine differentiator rather than a connector: full-text IEEE is delivered in workflow, which the vendor describes as unavailable on other platforms, alongside the Prior Art Database and patent data from 154 authorities. Output leaves the platform as structured reports intended to be shared with stakeholders. What the higher bands require was not established. No connector catalogue is published, no external system is named as supported, no API or developer documentation was located for this product, and nothing describes how results, saved work or portfolio data would move into a docketing, IP management or document management system.

For a corporate IP department that already runs such a system, that is the integration question, and the estate does not answer it. The Architecture and Security page was located and not read, and under R25 is named as a surface that could move this row. Verified 13 September 2026.

Tenancy separation is stated and region is not, which under R38 is the B band because tenancy and region are co-equal limbs and publishing either one clears C. The tenancy statement is explicit and is made as a governance commitment rather than as an architectural aside: customer data remains fully isolated from other organisations on the platform. On a shared prior art search service where the input is an unfiled invention disclosure and the other tenants include competitors in the same technology field, isolation between customers is the deployment property that matters most, and the vendor commits to it in terms.

It is reinforced by the SOC 2 Type 2 examination covering Confidentiality of the Software Services System, and by the description of interaction histories as private and secure. Delivery is a hosted browser-based platform, with the Innovation Power Suite accessed at ip.com and enterprise deployment handled through the account team. What is not published is region. No data centre location, country of processing or regional election is stated anywhere located, which is a real gap for European and Asian customers with residency obligations, and no cloud provider is named.

Nor is the tenancy model itself characterised: the vendor states the effect, that organisations are isolated, without saying whether that is achieved by single-tenant deployment or by logical separation within a shared environment. Verified 13 September 2026.

Certification is real, current in type and stated with scope, short of accessible evidence, which is the B band and near the top of it. The claim is specific in the two ways that usually go missing. The type is named: IP.com has undergone a SOC 2 Type 2 examination, which the vendor correctly describes as confirming both the suitability of design and the operating effectiveness of controls, rather than the point-in-time design test a Type 1 provides.

And the scope is stated: the controls covered are those relevant to Security, Availability and Confidentiality of the Software Services System, with the examination described as covering protection of information during collection, creation, use, processing, transmission and storage, the accessibility of information and services, and the protection of sensitive information from unauthorised disclosure. A public trust centre carries this, hosted on a third-party trust platform and reachable without a sales conversation, and the vendor states that it maintains certifications and undergoes regular third-party audits.

Two things hold it at B. No report, examination period, date or auditor is published, so a buyer can read the scope and not the evidence or its currency. And the access flow for the underlying documentation was not established: trust portals of this kind commonly gate reports behind an NDA click-through, and under R5 where the portal does not state which tier applies the lower tier is graded and the reason given. Verified 13 September 2026.

The architecture and the ownership of the retrieval engine are disclosed in more depth than anywhere else in this corpus, while the reasoning layer's provenance is not, and no change notification exists, which lands in the B band. On Semantic Gist the disclosure is close to complete. The model is named. Its architecture is described in technical terms rather than by adjective: probabilistic language modelling providing high-dimensional document representations, a deep neural network compressing those into low-dimensional semantic signatures, both used in concert so that signatures drive coarse filtering and the language model drives precision ranking, with a further multi-level concept layer built on language modelling and graph clustering and automatic n-gram phrase recognition.

Provenance is stated as a negative that is worth more than most positives: the engine was developed entirely in-house and is expressly not built on open-source search frameworks or licensed third-party components. Ownership is evidenced rather than asserted, the technology being covered by United States Patents 8,539,000 and 8,548,951 with applications pending. Under ground rules section 2 that is the best-evidenced form of a supplier absence, because the absence is stated rather than inferred.

Two gaps hold it at B. **CompassAI, the reasoning layer, is not covered**: nothing states whether it is built in-house or sits on a third-party foundation model, which is the question a buyer would now ask first. And under R34 no version is identified and no change notification is offered, the vendor instead describing continuous improvement. Verified 13 September 2026.

No pricing information is published at any level, including the unit of charge, which is the D band. The page inventory was taken under R20 from the site's own navigation and footer, covering the four Why IP.com pages, the IP Suite products, the professional services lines, the industry and use-case pages, resources, company pages, the trust centre and the legal pages. There is no pricing page. Every commercial route resolves to Book a Demo, Request a Demo, Contact Sales or a lead-capture form, and even the white papers and sample reports are gated behind contact details.

Nothing published states the charging model, so a buyer cannot establish whether InnovationQ+ is licensed per named user, per concurrent seat, by search volume, by module or by corpus entitlement, and nothing indicates what the IEEE full-text access costs relative to the base platform or whether the Patent Vitality Report is included or metered. Under R10's closing discipline no structure means no row, so no VendorPricing row is written.

Recorded and expressly not credited under ground rules section 3 and the naming determination for this record: the estate does publish one relative price point, that defensive publishing costs as little as two per cent of a patent filing, but that concerns InnovationQ, the free publishing interface, and not this product. Verified 13 September 2026.

Coverage is described with real substance and quantified, with the boundaries left open, which is the B band. Data coverage is stated precisely: 154 global patent authorities searchable in a single query and more than 190 million IP data points spanning patent and non-patent literature. One element of it is genuinely distinctive rather than comparative, being full-text IEEE access, which the vendor describes as the most-cited technical source in patent prosecution and as unavailable on other platforms, alongside IP.com's own Prior Art Database of technical disclosures.

Search modality is itself a form of coverage here and is enumerated: concept, keyword, hybrid, image and whole-document search, so image-based prior art and search-by-disclosure are both supported. Buyer coverage is named across engineering and R&D, IP strategy and IP professionals, with patent offices stated as institutional customers. Use cases are enumerated concretely, covering prior art and patentability, freedom to operate, invalidity, landscape mapping, competitive and competitor analysis, portfolio analysis, licensing opportunities and merger and acquisition diligence.

What is left open is every limit. No technology area is named as better or worse covered, no jurisdiction is identified as out of scope among the 154, nothing states which languages the retrieval engine handles well, and nothing distinguishes what InnovationQ+ covers from the adjacent suite products. Verified 13 September 2026.