# Message test — https://www.resemble.ai/products/detect After reading your page, only 6 of 15 personas could name a reason to pick you over a similar option. - **Page tested:** https://www.resemble.ai/products/detect - **Audience tested against:** Security and fraud leads at telecom carriers, banks and insurers, government agencies, and large platforms who are seeing deepfake fraud in calls, meetings, or media and need to catch it as it happens. They champion the deal but rarely own the budget, so they're building a case with a committee of 4 to 5 people, often including compliance, HR, or a channel partner, and they usually need on-prem or data-residency options to get through procurement. They can put a number on the fraud and have a buying timeline, typically for deals of $50K and up. - **Personas:** 15 simulated - **Report:** https://grader.wynter.com/r/multimodal-real-time-deepfake-detection-at-ent-JubeZuo > These answers are generated by AI, scored on Wynter's B2B Message > Layers framework using behaviorally-diverse simulated personas. The > methodology is real and the critique is directional. What a simulated > persona cannot have is a live budget, a renewal coming up, or a boss > asking about this quarter. --- ## 01 · The scores Every persona answered all four questions. These are four independent proportions of the same panel, not stages of a funnel. | Layer | Question | Cleared the bar | Strength | Of those who passed | | --- | --- | --- | --- | --- | | 1. Clarity | Do they understand what you do? | 15/15 | 88% | 7 without hesitation, 8 with reservations | | 2. Relevance | Can they tell what it solves, and who it's for? | 14/15 | 80% | 4 without hesitation, 10 with reservations | | 3. Value | Do they actually want it? | 14/15 | 74% | all with reservations | | 4. Differentiation | Is there a reason to pick you over the alternatives? | 6/15 | 44% | all with reservations | **Brand alignment** (a side metric, not one of the four layers) — 7/15, 48% strength (all with reservations). Does the page read like the company you actually are? **Fix first: Differentiation.** Earliest failing layer, walking the sequence in order — not simply the lowest score. --- ## 02 · What to change, layer by layer Ordered worst-first. Specific edits, not a restatement of the score. ### Differentiation **Move the named-generator accuracy table above the fold, directly under the hero line.** The per-model coverage list of named generators with accuracy per model is the only claim buyers said they could test, and it sits below general benchmark copy. Put it high on the page so the specific evidence lands before the broad claims. *effort medium · impact high · tested against Specifics beat superlatives* **Add the source and date beside every accuracy figure on the page.** Accuracy percentages appear with no named dataset, test date, or false-positive rate, so buyers discount them and assume a vendor-run test. State what corpus was used, when, and the false-positive rate next to each number. *effort medium · impact high · tested against Proof next to the claim* **Add a line under the hero stating why to pick this over score-only detection tools.** Nothing on the page says what this does that a competing detection vendor does not. Write one line naming the specific difference, verdict with explanation and exportable audit trail, plus on-prem deployment. *effort low · impact high · tested against Give a reason to choose you* ### Relevance **Name the buyer role and team in the first two lines of the hero.** Buyers had to infer the audience from compliance wording rather than read it. State it outright, for example fraud and security teams at banks and contact centers screening inbound voice calls. *effort low · impact high · tested against Name the audience* ### Brand alignment (side metric) **Rewrite the hero subhead in risk terms: fraud loss, regulatory exposure, call volume.** The voice reads as product marketing rather than security, so security and fraud leaders discount it before reaching the evidence. Lead with the loss being prevented and the regulatory exposure, not the capability list. *effort medium · impact high · tested against Problem before solution* **Add a deployment and compliance section near on-prem covering data residency and audit export.** Security buyers are already asking where audio is processed and stored, and the page leaves it to them. Answer it plainly beside the on-prem line: what stays on their infrastructure and what the audit trail exports. *effort medium · impact medium · tested against Answer the live objection* **Standardize on one product name and one word, verdict, across every section.** The page switches between verdict and score and varies the product name, which blurs the exact distinction it is selling on. Pick one label for the output, define it once, and use it everywhere. *effort low · impact medium · tested against Plain language* --- ## 03 · What is working ### The multimodal detection promise comes through clearly Three respondents could state the product back: multimodal deepfake detection returning a verdict and explanation rather than a bare score, clear from the hero line. > takes in audio, video, or images and returns a verdict plus an explanation of why something's flagged as AI-generated or manipulated, rather than just a bare score > > — Fraud Prevention Manager, Financial Services, 5000+ ### On-prem deployment reads as a concrete procurement advantage Three respondents said the on-prem option resolves data residency and procurement constraints that rule out cloud-only competitors, and fits existing stacks without rip-and-replace. > real-time voice clone and live-call-interception detection inside our existing contact center stack means fewer successful vishing/social-engineering fraud losses I have to explain after the fact > > — Head of Security, Insurance, 1001-5000 > the on-prem/data residency mention buried in the header ("Available on-prem or in the cloud via API") matters a lot given my procurement constraints > > — Director of Fraud Operations, Technology Platforms, 1001-5000 ### Explainability and exportable audit trails are the value respondents actually want Five respondents said verdict-plus-explanation and audit trail close a real compliance gap and matter more than raw accuracy percentages, versus score-only tools and reactive monitoring they run today. > the explainability angle — audit trails, artifact-level reasoning, "why it was flagged not just that it was" — would matter for our compliance and trust & safety workflows, since our current tools mostly just give a score with no rationale to hand to legal or regulators > > — Security Lead, Telecommunications, 1001-5000 > a real-time flag with an audit trail ("exportable for legal, compliance, and regulatory review") could close a gap our current monitoring doesn't touch > > — Fraud Prevention Manager, Financial Services, 5000+ > The "audit trail: exportable for legal, compliance, and regulatory review" line is the part that actually matters to me, not the accuracy percentages. > > — Chief Information Security Officer, Government, 5000+ ### The named per-model coverage table is the most credible differentiator on the page Four respondents singled out the list of named generators and per-model accuracy as specific enough to test and more believable than unsourced benchmark claims. One still wants independent verification of it. > The "250+ generative AI models tested against in real time" with named coverage — ElevenLabs, Midjourney, Kling, Sora, etc. — with specific per-model accuracy numbers (99% Flux, 98% Midjourney, 94% Stable Diffusion) is the one thing that could tip a shortlist decision, because most competitors just say "detects deepfakes" without naming what they actually beat > > — Security Lead, Telecommunications, 1001-5000 > the "attack coverage" list — voice clone, replay attack, face swap, partial edit, synthetic image, live call interception, video conferencing, reverse image search all marked "DETECTED" — because that's specific enough to test against > > — Fraud Prevention Manager, Financial Services, 5000+ > The "250+ generative AI models tested against in real time" table with specific accuracy numbers per model — DALL-E 3 at 98%, Flux at 99%, Veo at 99% — is more concrete than most vendors give me, and would make me pick this over a competitor that just claims "industry-leading accuracy" with no breakdown. > > — Chief Information Security Officer, Government, 5000+ > The 250+ model coverage table with named generators — ElevenLabs, Sora, Kling, Veo, Suno, StyleGAN — broken out by modality is the one thing that would put this ahead of a vaguer competitor, because it's checkable > > — Senior Security Officer, Telecommunications, 5000+ --- ## 04 · What the personas said ### Accuracy and model-count claims are unsourced, so respondents discount them Five respondents flagged the accuracy figures as lacking a named source, independent benchmark, false-positive rate, or third-party validation against their own attack types. Two said this puts it behind Pindrop and Actimize. > the 99.5% accuracy and "250+ models" figures need a named source or independent benchmark before I'd put any weight on them > > — Fraud Prevention Manager, Financial Services, 5000+ > An independent third-party test showing the on-prem build hits comparable accuracy against attack types we actually see — voice clone and replay attacks on live telco call traffic — not just their own benchmark page; without that, it doesn't get past the first meeting > > — Security Lead, Telecommunications, 1001-5000 > I'd want their false-positive rate on our actual call volume and a couple of reference customers in insurance or BFSI before I take this past a first call — those benchmark numbers mean nothing until I see them against my traffic, not theirs > > — Head of Security, Insurance, 1001-5000 > If a competitor on my shortlist shows me a false-positive rate and one client in financial services, they win the meeting over this page as written > > — Head of Security, Insurance, 1001-5000 > I'd need the accuracy numbers sourced against a named benchmark before I trust the 99.5% figure > > — Director of Fraud Operations, Technology Platforms, 1001-5000 ### Inconsistent naming muddies the verdict-versus-score distinction Two respondents said the page uses verdict and score interchangeably, undermining its own differentiator, and that the product name varies across the page. > "verdict" and "score" get used almost interchangeably later in the page (accuracy percentages everywhere) even though the whole pitch is verdict-not-score, so the copy slightly undercuts its own distinction > > — Head of Security, Insurance, 1001-5000 > it's called "deepfake detection" in the hero, then "authenticity scoring," "explainable detection," and "Resemble Intelligence" as if those are separate things, so I'm left doing the work of collapsing four marketing names into one product in my head. > > — Senior Security Officer, Telecommunications, 5000+ ### The intended buyer is never named on the page Six respondents said they had to infer the target persona from compliance language and scattered context clues rather than reading it stated. One countered that problem and audience were clear within the first two lines. > Who it's for is less explicit — it's not stated as "built for CISOs" or "for government," but the mentions of "compliance teams, legal review, and trust & safety workflows" and later "Built for regulated industries" let me infer the buyer without much digging. > > — Chief Information Security Officer, Government, 5000+ > The who is inferred, not spelled out: it's built for "compliance teams, legal review, and trust & safety workflows that need evidence, not just a score," plus the industry callout and the Zoom/Teams/Meet/Webex angle points at contact-center and fraud/security roles like mine. So problem: clear and fast. Intended reader: I had to piece it together > > — Senior Security Officer, Telecommunications, 5000+ > A line naming my industry and my exact pain point directly — something like "built for insurance and financial-services fraud teams fighting synthetic voice claims and impersonation in claims calls" — rather than making me infer it from a generic Zoom/Teams/compliance list > > — Fraud Prevention Manager, Insurance, 5000+ > the intended reader is never stated outright — there's no "for security teams at banks" or "for contact center fraud analysts" line anywhere > > — Senior Security Officer, Financial Services, 5000+ > Pretty much instant — the hero line "Know what's real with multimodal deepfake detection" plus "Most tools return a score. Resemble Detect returns a verdict and an explanation" told me the problem (score-only tools aren't enough, you need evidence) within the first two lines. > > — Head of Security, Technology Platforms, 1001-5000 > the intended reader is never explicitly named upfront — I had to infer it from scattered mentions > > — Fraud Prevention Manager, Financial Services, 5000+ > the intended reader is never explicitly named; I had to infer it from scattered mentions > > — Director of Fraud Operations, Technology Platforms, 1001-5000 ### The tone reads startup product-marketing, not enterprise security Five respondents described the voice as confident and feature-forward rather than analytical, missing board-level risk framing, telecom-specific credibility, and case studies. One noted it casts a wide net instead of owning one lane. > it's confident and feature-forward ("Up to 99.5% accuracy," "Zero day coverage") but skips the thing I'd actually want first — a named financial-services logo or a board-level risk framing > > — Fraud Prevention Manager, Financial Services, 5000+ > the total absence of company facts — no "founded in," no customer logos I recognize, no team size — so I'm inferring their maturity from confidence and polish, not evidence, and that's a gap a page this specific about accuracy percentages should have closed > > — Head of Security, Insurance, 1001-5000 > there's no telecom-specific case study, no logo I recognize from telco fraud, and the "trusted partners" section is just a placeholder line with no names shown to me > > — Director of Fraud Operations, Telecommunications, 1001-5000 > the tone is still more product-marketing than analyst-brief: punchy claims, demo videos, benchmark tables with no methodology > > — Senior Security Officer, Financial Services, 5000+ > the copy has that slightly breathless "we test 250+ models" startup energy rather than the buttoned-up tone of an established enterprise security player > > — Fraud Prevention Manager, Financial Services, 5000+ --- ## 05 · The hardest read An adversarial pass over the findings. Every claim below was checked against the panel's own answers; unsupported ones were dropped. - **The page's own numbers are its weakest asset, and the differentiator only survives where the numbers are specific enough to be checked.** *(high)* Five respondents discounted accuracy and model-count claims as unsourced, two ranking the page behind Pindrop and Actimize; the per-model table earned credit from four precisely because it can be tested, and one still wants independent verification. - **The core differentiator is undermined by the page's own copy before a buyer can evaluate it.** *(high)* Two respondents said verdict and score are used interchangeably and the product name varies across the page, while five others cite verdict-plus-explanation as the value that closes their compliance gap. The page blurs the exact distinction it is selling. - **Buyers have to assemble the audience themselves, which turns qualification into guesswork.** *(high)* Six respondents inferred the target persona from compliance language and scattered context clues rather than reading it stated; only one found problem and audience clear in the first two lines. - **The voice disqualifies the page from the enterprise security deals its own strongest proof points are built for.** *(high)* Five respondents read the tone as startup product marketing missing board-level risk framing, telecom credibility and case studies, while on-prem deployment and exportable audit trails are procurement- and compliance-grade assets being sold in the wrong… - **Clarity about what the product does is not translating into belief that it works.** *(medium)* Only three respondents restated the multimodal verdict promise, while five discounted the accuracy claims as unvalidated and five found the voice feature-forward rather than analytical. Comprehension is shallower than the doubt. - **Spreading across every use case costs the page the one lane where it is most credible.** *(medium)* One respondent said the page casts a wide net instead of owning a lane, and the assets that land — on-prem procurement fit and audit-trail compliance — are narrow, specific claims cited by three and five respondents respectively. --- ## 06 · Who answered | # | Role | Industry | Company size | | --- | --- | --- | --- | | 1 | Security Lead | Telecommunications | 1001-5000 | | 2 | Fraud Prevention Manager | Financial Services | 5000+ | | 3 | Head of Security | Insurance | 1001-5000 | | 4 | Chief Information Security Officer | Government | 5000+ | | 5 | Director of Fraud Operations | Technology Platforms | 1001-5000 | | 6 | Senior Security Officer | Telecommunications | 5000+ | | 7 | Security Lead | Financial Services | 1001-5000 | | 8 | Fraud Prevention Manager | Insurance | 5000+ | | 9 | Head of Security | Government | 1001-5000 | | 10 | Chief Information Security Officer | Technology Platforms | 5000+ | | 11 | Director of Fraud Operations | Telecommunications | 1001-5000 | | 12 | Senior Security Officer | Financial Services | 5000+ | | 13 | Security Lead | Insurance | 1001-5000 | | 14 | Fraud Prevention Manager | Government | 5000+ | | 15 | Head of Security | Technology Platforms | 1001-5000 | --- ## 07 · Before you act on this The methodology is real, and the critique is directional. What a simulated persona cannot have is a live budget, a renewal coming up, or a boss asking about this quarter. **Validate anything you're betting on with real ICPs who are actually in-market.** Being wrong is more expensive than you think. Finding out is cheaper than you'd guess. Wynter runs message testing with verified B2B professionals — trusted by HubSpot, RingCentral, Shopify, Cognism, Paddle, Veeam, Rippling and Miro. This report is kept for 60 days from 2026-09-29, then deleted along with the personas and their answers.