RollWorks

A twenty year old advertising business with a substantial and genuinely load bearing model layer sitting on infrastructure that predates it. Three engines are named and each has a real job: a bidding engine that the vendor states makes more than 2.5 trillion predictions a day to decide which impressions to buy, an insight engine that predicts buyer journey stage and surfaces early signals for the account based product, and a set of assistant agents that recommend next steps and execute discrete tasks.

The vendor leans on the model story hard, opening its own page on the subject by observing that plenty of advertising companies claim to use the latest technologies and positioning a seventeen year data legacy as the difference.

Remove all three and the customer still has the product they bought: a demand side platform with its own inventory relationships, target account lists, fit scoring, website visitor identification, bidirectional record system integration, campaign management and attribution reporting. Bidding would get worse rather than stop, since a programmatic buyer can bid on rules.

The company was founded in 2006 as an advertising network and the account based product was assembled on that foundation in 2018, so the models are an improvement to an existing machine rather than the reason it exists. Ask which capabilities cease to function without the bidding and insight engines, as opposed to performing less well.

Two automated systems act without a human in the loop and the controls over both are undescribed. The bidding engine buys impressions continuously at machine speed, factoring what the vendor states are more than eighty intent signals per impression, so budget is committed thousands of times a second without review.

Separately, a signal the vendor calls Spike detects rising engagement at an account and automatically emails the representative who owns it, and pushes the same signal into the record system, so a model decides when a salesperson is prompted to act. The assistant agents are described as working behind the scenes to help align strategy, develop creative and draw insights, which is a description of scope rather than of limits.

A third route was published in beta and it carries the largest consequence of the three. An interface for external assistants lets a connected agent create and manage campaigns and pull reporting from inside an assistant the vendor names, which places campaign creation, and therefore the commitment of media budget, on the far side of a boundary this vendor does not operate. That fact does not lower this grade, and the reason is architectural rather than charitable.

Creative review stands between any agent and an audience: a staffed team inspects every advertisement, its copy, its destination and its landing page before it can run, and edits trigger re review, so an agent may assemble and launch but cannot cause anything to be seen by anyone. What that review does not gate is money.

It inspects creative rather than budget, so an agent working with already approved creative is not evidently prevented from raising a budget, widening targeting or extending a flight, and no per role permission model or administrator control over the interface is described anywhere. Nothing published describes budget pacing controls, bid floors or ceilings, what a customer may override in the bidding engine, what threshold triggers a Spike alert, or how a representative suppresses one.

Ask what bidding parameters a customer can set or override, what threshold and suppression controls exist on automatic sales alerts, and what an administrator can permit or withhold per role on the assistant interface.

Three systems named and trademarked, and not one number that would let a buyer judge whether they work. The vendor publishes a dedicated page for the subject, names each engine, states its function, distinguishes bid optimisation from audience insight from agent assistance, and even discloses the trademark status of each, which is more structure than most vendors offer.

The volume figures are stated with precision, at more than 2.5 trillion predictions a day and more than eighty intent signals per impression. What none of it includes is any measure of quality. No model family, architecture, provider or version appears for any of the three engines.

No accuracy, precision, lift or error figure is published for fit scoring, for journey stage prediction, or for the intent signals that decide which accounts a customer spends money against, and the only quantified outcome offered is a claim that identified accounts are four times more likely to convert, with no baseline stated.

The parent trust centre opens by naming data, privacy, security and artificial intelligence as the four things it covers, and then contains sections on every one of those except the last. A buyer reading the page written to explain the models learns their names and their throughput and nothing about their reliability. Ask for the accuracy of fit scoring and journey stage prediction, measured against what baseline and over what population.

Named customers with specific figures, sitting under three quantified claims that carry no method at all. The credible half is real: a lifecycle marketing manager at a named mortgage technology company states the platform reached 93 percent of its priority addressable accounts and describes it as directly influencing deals, a growth and product marketing executive who identifies himself as a former analyst at a major research firm endorses it by name, and further named customers appear with attributed quotes.

Marketplace recognition is independent and current across several award categories, and the vendor publishes an annual state of the industry report. Against that, the claim printed in the page furniture of every page on the legacy domain is that the platform delivers over five times the return on investment compared to other vendors in its category, and no study, comparator set, period or method is cited anywhere for it.

A second claim states that accounts identified by the platform are four times more likely to convert, again with no baseline. A third states that this is the only platform in the category able to deliver real results in less than thirty days, which is an unqualified superiority claim about every competitor. A buyer therefore meets three numbers before meeting any evidence. Ask for the study behind the five times return figure, including which vendors it compares against and over what period.

The compliance function here is staffed and operational rather than written down and hoped for, which is rare on this axis. Every advertisement is reviewed before it runs, and the review covers the creative, the copy, the destination address and the landing page, plus an evaluation of the advertiser's entire website against published content requirements.

Edits to existing creative trigger re review, and the vendor reserves the right to re examine approved campaigns at any time and to end campaigns that cease to be eligible, on the stated basis that partner network policies and self regulatory codes change.

Categories are refused outright rather than case by case, with advertising that implies a specific health condition or religious affiliation named as prohibited, and the vendor explains why the same advertisement may run on a billboard and not here, which is that behavioural targeting attracts obligations that untargeted media does not.

Minors are addressed with a stated rule rather than a disclaimer: the platform is not provided to sites intended for audiences the children's privacy statute covers, and the advertising code the vendor operates under prohibits targeted advertising to anyone sixteen or younger. Membership of three industry bodies and a European trust seal linked to an independent audit programme for targeted advertising put external codes behind the position.

Held below the top band because the obligations running the other way are unaddressed: nothing published describes what an advertiser must warrant about the target account lists and contact records it uploads for targeting, or where consent for that data is meant to come from. Ask what the advertiser must warrant about uploaded account and contact lists.

An unusually practical rights apparatus for a business built on behavioural data. A data protection officer is named with a published address, requests run through a dedicated privacy request service rather than a support queue, and the process is set out in four steps including a verification stage.

The part that separates this from a policy page is the instruction: the vendor publishes step by step guidance, with screenshots, for locating your own advertising cookie identifier in three named browsers and for finding mobile advertising identifiers, because without that identifier a request against an advertising network cannot be matched to anything. That is a company making its own erasure process usable rather than merely available.

Around it sit regime specific pages for European and Californian law, a standing opt out of personalised advertising, a notice at collection, and the United States privacy choices mechanism with its standard mark. The children's privacy position is stated as a rule the vendor applies to itself. Held below the top band because the axis has a second half this material does not reach.

Everything located addresses the individual whose browsing is tracked, and nothing reached describes what happens to the customer's own uploaded data, meaning the target account lists, contact records and record system fields that flow into the platform for targeting. No processing addendum, sub processor list or retention schedule for customer data was found on any route, and the security portal that might hold them requires entry. Ask for the processing addendum and sub processor list covering advertiser uploaded account and contact data.

The data asset is the entire proposition and the vendor cannot state its size consistently on a single page. The page devoted to the subject carries a statistics panel claiming more than 2.8 billion distinct user profiles and more than a hundred unique intent signals per profile, while the frequently asked questions further down the same page state 2.6 billion user identities twice and eighty or more intent signals per impression.

Two headline figures about the same asset, contradicted within one document, in the two places a buyer would look. A third figure moves elsewhere, with company coverage given as 25 million on one surface against materially lower counts reported by third parties.

Where provenance is described, it is described honestly and in general terms: the vendor states plainly that it uses data gathered from actual browsing behaviour and acknowledges that this attracts obligations untargeted media does not, and it credits enrichment to seventeen or more commerce, mail and marketing platform integrations together with data partners and identity providers.

Intent is the strongest part, sourced through a named partnership with a specialist provider that is separately recorded in this index, so at least one input has a traceable supplier. What is absent is the licence chain. The identity providers and data partners are not named, no licence basis is described for the contact and company records, and nothing states what rights pass to an advertiser who targets against them. Ask which identity providers and data partners supply the graph, and under what licence the contact and company records are held.

Own infrastructure, official channels, and a conformance position stated against named external codes. The vendor operates its own demand side platform rather than reselling another company's application layer, which is why its inventory reaches display, native, video, connected television, digital out of home and social rather than a single network.

Where it does depend on somebody else's platform, the connection is the official one, including the professional network's own advertising product and advertising placements inside an assistant product. The conformance position is explicit and external rather than self declared: the vendor is a member of the network advertising self regulatory body whose code of conduct it cites as constraining its own behaviour, is registered with the industry accountability programme for advertising supply chains, holds a European trust seal that links to an independent audit scheme for targeted advertising, and states that partner network policies and self regulatory codes are the reason campaigns can be refused or ended.

That is a vendor accepting outside rules it does not write. Held below the top band because the dependency that actually threatens a business built on 2.6 billion browser identities goes unmentioned on every surface reached. Browser and operating system vendors decide whether third party identifiers continue to function, that decision is made without reference to any advertiser, and no page located states a position on identifier deprecation or describes what replaces the graph if it degrades. Ask what the platform's position is on third party identifier deprecation and what identity approach succeeds the current graph.

The governance that exists is data governance, and it constrains the models by accident rather than by design. Real limits do apply to what the engines can be pointed at: targeted advertising to anyone sixteen or younger is prohibited under the code the vendor operates by, services are refused to sites intended for children, advertising that implies a health condition or religious affiliation is refused outright, and a standing opt out removes an individual from personalised advertising altogether.

Those are meaningful constraints on the population the models act upon, and they are enforced rather than aspirational. What has no coverage at all is the models themselves. Nothing published describes what data trains the bidding, insight or agent systems, whether advertiser data or individual browsing histories are used to improve models across customers, what retention applies to model inputs, or what evaluation or review any of the three engines undergoes.

No governance document, model risk assessment or management standard for artificial intelligence was located. The gap is made conspicuous by the vendor's own framing, since the parent trust centre names artificial intelligence in the sentence describing what it covers and then contains no section on it, while the assistant agents are described as executing tasks in a marketer's workflow without any statement of what they may not do. Ask what data trains the three engines, whether one advertiser's data informs models serving another, and what evaluation the agent system undergoes before it acts.

The strongest position on this axis that this project has recorded, and it comes from the advertising industry's own machinery rather than from anything the vendor invented. A person targeted by this platform has a working route out and a working route in.

A standing opt out from personalised advertising is published at a stable address, the United States privacy choices mechanism is implemented with its standard mark, a notice at collection is published, and an access and deletion process runs through a dedicated request service with identity verification.

The vendor then does the thing that makes those rights usable rather than nominal, publishing illustrated instructions for finding your own advertising identifier in three named browsers and on mobile devices. It is a member of the body whose code requires transparency to the people being targeted, and it carries a European trust seal that links to an independent audit programme for targeted advertising specifically.

Held below the top band on the half of the product that is not advertising. The account based side identifies anonymous companies and contacts visiting a customer's website and pushes those contacts into a sales record system to be acted on, and no disclosure reaches the individual at that point: nothing describes what a person is told when their employer is identified from their visit, or how they would know that had happened in order to exercise any of the rights above. The advertising is disclosed and the identification is not. Ask what disclosure reaches an individual whose visit is resolved to a company and passed to a sales team.

Broad, specific and named across several categories rather than gestured at. Bidirectional synchronisation is published for three major record and marketing automation systems, alongside analytics, a connector for the professional network, a site visitor interface and sales automation.

Partner integrations are named rather than implied, covering three web personalisation vendors and three gifting and direct mail vendors, two of which are separately recorded in this index so a buyer can evaluate the components independently. The vendor states seventeen or more integrations that activate first party data.

A connector for assistant tooling is published in beta with a help centre article, allowing campaign creation, account intelligence retrieval and reporting from inside an external assistant, and there is a partner portal and an agency partner programme behind it. Two limits keep it below the top band.

Connectivity is gated by package, with the bidirectional record system integrations and the site visitor interface appearing in tiers whose prices are not published, so a buyer cannot tell what the integration they need actually costs. And no developer documentation or public interface reference was located on any route, with the site visitor interface appearing as a feature name on a comparison table rather than as a documented endpoint. Ask which package includes the bidirectional integration and the site visitor interface, and where the interface documentation is published.

Recorded as gated rather than absent, because the material that would answer this sits behind a portal. The vendor operates a structured trust programme and routes all security information to a dedicated security portal on a recognised trust portal service, which requires entry rather than publishing to the open web, so hosting provider, regions, tenancy model, isolation and recovery objectives were not readable on this pass and are not established as missing.

What is readable on the open surfaces gives the shape without the substance. The product is delivered as a hosted service with no deployment options offered or discussed, the operating company is registered in the United States, and a European data protection page exists alongside the general privacy notice.

Residency has weight for this particular product beyond the usual reasons, because the platform holds browsing derived identifiers for billions of individuals including European ones, and because European advertising and privacy obligations bear directly on where that processing happens. Nothing on the open surfaces states where it happens. Ask for the hosting regions, the tenancy and isolation model and the recovery objectives, and whether any regional processing option exists for European advertisers.

A real trust programme with the security half behind a gate. The parent operates a trust centre at its own address covering data controls, privacy rights, advertising guidelines, invalid traffic handling and intellectual property, and it is substantive on all of those rather than decorative.

Security is handled by routing the reader to a separate security portal on a recognised trust portal platform, which is a legitimate arrangement and also means that no certification name, audit report, audit period, auditor or control enumeration was visible without requesting entry. Nothing on the open surface asserts a certification either, so this record neither credits nor discounts one.

External accountability of a different kind is verifiable without entering anything: registration with the industry accountability programme for advertising supply chains, membership of two industry bodies, and a European trust seal that links out to an independent audit scheme for targeted advertising, all displayed with links to the issuing organisations rather than as unlinked badges.

Those are audits of advertising practice rather than of information security, and a buyer should not read them as the latter. The position matches the pattern this index records repeatedly, where a buyer cannot verify the security posture before a sales conversation. Ask for the current audit report with its period and auditor, and whether the security portal can be opened without a mutual non disclosure agreement.

The clearest example this project has found of a vendor publishing the entire structure and withholding only the number. A page titled pricing sets out four packages in a full comparison matrix, and the detail is real: target account list allowances rising one, two, four and eight across the tiers, account intent breadth increasing by named degree, free static creative capped at two, three, five and five advertisement sets a quarter, support escalating from a named point of contact to an account manager, programme review every six months then every three, and execution services allowanced at two, four and six hours a month with design services at the top.

Feature gates are itemised down to which integrations each tier unlocks. Every one of those columns is headed by a description and none by a price. Not one currency figure appears anywhere on the page. The withholding is sharpened by the vendor's own positioning, which criticises competitors for charging extra for audiences, integrations and measurement and states that its packages include everything needed for one price, under a heading promising full funnel account based marketing without the fine print.

The one price is the single fact not published. A second cost compounds it: media is bought at a variable rate that sits outside the subscription entirely, so even a quoted platform figure would not be the bill. A no fee entry does exist and is advertised on the same page, but it buys retargeting on a pay for media basis rather than the account based product.

Third party aggregations of buyer reports place the platform in a wide band with media spend of a similar order on top, and they disagree with each other. A figure of nine hundred and seventy five dollars a month appeared in the vendor's own announcement of this packaging in 2020 and appears nowhere on the current page. Ask for the price of each of the four published packages and the expected media spend alongside it.

Deletion is well built for individuals and undescribed for customers, and the distinction is the whole finding. A person tracked by the advertising network can request access or deletion through a dedicated request service with a published four step process, verification, and illustrated help for supplying the identifier that makes the request matchable. That is a real erasure capability. The advertiser who pays for the platform has nothing comparable published.

No export scope, format or route was located for the assets that accumulate in an account based programme, which are the target account lists, the fit scores, the journey stage history, the audience definitions and the attribution record that justifies the spend. Nothing describes post termination access, retention or deletion timing for that material, and nothing states what happens to an audience built inside the platform when a customer leaves.

Recorded as a retrieval limit as well as a gap: the master terms of service were not retrieved on this pass, so a contractual export or return provision may exist there, and the surfaces that were read do not mention one. The commercial shape raises the stakes, since packages are annual and the audiences a customer builds are the accumulated product of media spend rather than of a data upload. Ask what exports at termination, in what format, whether audiences and attribution history are included, and what is deleted and when.

The advertising equivalent of this axis is whether the impressions a customer buys reach real people on acceptable inventory, and this vendor addresses it with named machinery rather than assurances. Invalid traffic is defined openly as covering both clicks and impressions that do not reflect genuine interest, including mechanically generated click and bot fraud, and the response is described in three layers. A dedicated trust and safety team performs both manual and automated monitoring.

Block lists exclude domains associated with high fraud, maintained by several monitoring programmes that update them continuously, with manual checks alongside. And a named third party fraud and viewability provider is engaged in addition to the vendor's own methodology, which puts an outside party between the vendor's interest in billing impressions and the measurement of whether those impressions were real.

Registration with the industry accountability programme for advertising supply chains adds an external standard covering exactly this problem. On the creative side, review before serving prevents a customer's advertisement from running anywhere it is not eligible, which protects delivery as well as compliance.

Held below the top band because none of it is quantified for the buyer: no invalid traffic rate, no viewability rate, no filtered impression percentage and no make good or credit policy is published, so a customer cannot see what share of their spend was screened out or what happens commercially when fraud is identified after billing. Ask for the invalid traffic and viewability rates reported to customers, and for the credit policy when fraudulent impressions are billed.

A capacity ladder standing in for a market statement. The four published packages do describe who each suits, but they do it in units of platform capacity rather than of customer, scaling target account lists from one to eight and services from a periodic review to six hours a month, so a buyer infers their own fit from how many account lists they think they need.

The vendor's own answer to who the product is for runs to a single sentence about revenue focused business to business marketers seeking omnichannel account based strategies, which describes an approach rather than a segment. Industry pages do exist and are specific, covering retail, business to business, hospitality, finance, direct to consumer, nonprofit and agencies, but they belong to the parent advertising brand and most of them describe businesses that would not buy an account based product at all, so they do not narrow this record's market.

Positioning language points at smaller teams, with the product described as designed for lean teams that aim high and sold on delivering results inside thirty days. No employee count, revenue band, region or geography statement for the account based product was located on any surface reached, and no minimum spend is stated despite media being a separate and substantial budget line.

A buyer outside the United States has nothing telling them whether the inventory, the identity data or the support cover their market. Ask what company size and annual media budget each package assumes, and which countries the account and identity data covers.