# Message test — https://www.resemble.ai/ After reading your page, only 3 of 15 personas could name what kind of product this is, unprompted. - **Page tested:** https://www.resemble.ai/ - **Audience tested against:** The central security and risk teams - CISOs, VPs of CyberSecurity, Heads of Fraud or Risk, Heads of Compliance. Trust and Safety Teams. But always ideally the leader of those groups. Industries or segments - regulated first and everything else second. Telco is a stronghold for us now. Contact center owners or those with a large CC are huge buyers. Insurance and Finance as well. Any industry where the ROI is very obvious and conetnt or communications are used as proof to giving out money. AKA, you upload a claims photo of your vehicle, insurance reimburses you $800 for the damage based on the photo adn the estimate from the local repair shop. But you used AI to generate both the estimate and the photo. Company size - upper mid market to enterprise. It's not for individuals, consumers, pay as you go monthly people. This is an API / models copmany that needs a surface (integration / widely used enterprise product / SaaS offering) or a tech team to implement. - **Personas:** 15 simulated - **Report:** https://grader.wynter.com/r/multimodal-deepfake-detection-and-watermarking-Carinu0 > 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? | 3/15 | 90% | 8 without hesitation, 7 with reservations | | 2. Relevance | Can they tell what it solves, and who it's for? | 15/15 | 90% | 8 without hesitation, 7 with reservations | | 3. Value | Do they actually want it? | 15/15 | 78% | all with reservations | | 4. Differentiation | Is there a reason to pick you over the alternatives? | 12/15 | 67% | all with reservations | **Brand alignment** (a side metric, not one of the four layers) — 13/15, 70% strength (all with reservations). Does the page read like the company you actually are? **Fix first: Clarity.** Earliest failing layer, walking the sequence in order — not simply the lowest score. ### What they thought you sell 12 of the personas who named a category got it wrong: - 4× “Deepfake / synthetic media detection API” - 3× “Deepfake / synthetic media detection” - 3× “Deepfake / synthetic media detection software” - 2× “Deepfake detection software” --- ## 02 · What to change, layer by layer Ordered worst-first. Specific edits, not a restatement of the score. ### Clarity **Replace "world model" in the hero with a plain description of what the detector inspects.** "World model" tells a fraud lead nothing about how a call is judged fake. Write what the system actually listens for in audio and video, in words a security buyer already uses. *effort medium · impact high · tested against Plain language* **Cut "deterministic" or define it beside the accuracy claim in one sentence.** Detection models are probabilistic, so "deterministic" reads as a contradiction and costs trust. Either drop the word or say plainly what is repeatable: same input, same verdict. *effort low · impact high · tested against Plain language* **Rename the model list entries by fraud use case instead of internal model names.** The model names read like an internal catalogue, leaving buyers unsure which one handles mid-call voice fraud versus video KYC. Label each by the job it does, with the model name secondary. *effort medium · impact medium · tested against Lead with the use case* ### Differentiation **Add methodology, dataset and false-positive rate directly under the 99.5% and 98% figures.** The accuracy numbers stand alone with no test set, false-positive rate or latency, so they read as self-reported. Put one line of test conditions right beneath the numbers. *effort medium · impact high · tested against Proof next to the claim* **Fix or remove the broken benchmark CSV component on the accuracy section.** A broken benchmark element beside accuracy claims makes the evidence look unfinished. Either render the benchmark table properly or link a static results page. *effort low · impact medium · tested against Proof next to the claim* ### Brand alignment (side metric) **Replace the tagline "Deepfakes are everywhere. So are we" with a specific outcome line.** The tagline reads flippant for a fraud-prevention buyer signing off on security infrastructure. State what the product stops and when, such as detecting a cloned voice before a call completes. *effort low · impact medium · tested against Specifics beat superlatives* --- ## 03 · What is working ### The hero line and use-case cards identify the problem and buyer within the first screen Six respondents said the audience and problem were obvious without job titles being named, crediting the hero line and the use-case grid specifically. > the hero line "Deepfakes are everywhere. So are we" plus the subhead "Detect AI-generated audio, video, and images in real time with explainable results enterprises can trust" told me the problem in about five seconds > > — Head of Risk, Insurance, 1001-5000 > That use-case grid is doing the real work of audience targeting; I didn't have to hunt. > > — Chief Information Security Officer, Contact Centers, 501-1000 > the row of use-case cards right below it — "Contact center fraud," "KYC + account onboarding," "Law enforcement forensics," "Social trust and safety" — made the target reader clear > > — VP of Cybersecurity, Financial Services, 501-1000 > The line "Deepfakes are everywhere. So are we" plus "Detect AI-generated audio, video, and images in real time with explainable results enterprises can trust" told me the problem and the pitch within the first screen. > > — Head of Fraud, Contact Centers, 1001-5000 ### Real-time mid-call detection is the value respondents could restate Five respondents named real-time detection before call completion as the concrete payoff — cost reduction, KYC-stage voice fraud, and a compliance gap in their current call platform. > If it worked as promised, I'd get real-time flagging on live calls before the fraudster completes account takeover or wire fraud — that's the "alert your team before the call ends" line, and that's the actual money-saver for a telecom fraud desk. Today we catch this stuff downstream, after the loss; catching it mid-call changes the economics of our fraud ops entirely > > — Head of Fraud, Telecommunications, 501-1000 > catching voice-clone fraud in the contact center before the call ends, which is a live problem for us with IVR and agent-assist social engineering > > — CISO, Telecommunications, 1001-5000 > Fewer voice-fraud calls getting through KYC. Telnyx and Okta quotes are decent proof. Worth a short call, not a project yet. > > — Vice President of Cybersecurity, Insurance, 5000+ > I get a real-time detection layer sitting in front of contact center calls and onboarding flows that catches synthetic voice/video fraud before it costs us money > > — VP of Cybersecurity, Financial Services, 501-1000 > the "Contact center fraud" card specifically calling out "Integrate directly into carrier call infrastructure. Alert your team before the call ends" — that's a concrete workflow claim, not a feature list, and it maps exactly to a gap I have today > > — Head of Risk, Insurance, 1001-5000 ### Named customers and the Deutsche Telekom challenge win carry the credibility Four respondents cited the Okta VP and Telnyx CEO quotes, the Deutsche Telekom/T-Mobile 2025 challenge win, and carrier infrastructure integration as what separates the page from generic detection APIs. > Okta's VP saying it's "critical to strengthening the identity security fabric," Telnyx's CEO saying "compliance and security will be on by default across our network...thanks to Resemble," and the Deutsche Telekom/T-Mobile challenge win in 2025 > > — Head of Compliance, Financial Services, 5000+ > The third-party validated benchmarks section — "We don't grade our own homework," with the 99.5% audio / 98.2% video accuracy numbers and the RTF-vs-accuracy chart — is the thing that would tip me toward this one over a competitor, because it's a specific, checkable claim rather than marketing fluff. > > — CISO, Telecommunications, 1001-5000 > Okta and Telnyx named-customer quotes tip it in — real enterprises vouching, not just claims. > > — Vice President of Cybersecurity, Insurance, 5000+ > "Integrate directly into carrier call infrastructure. Alert your team before the call ends" — is the one concrete differentiator > > — VP of Cybersecurity, Financial Services, 501-1000 --- ## 04 · What the personas said ### Marketing language — 'world model', 'deterministic' — obscures how detection actually… Four respondents said the framing hid the mechanism, and one flagged that 'deterministic' contradicts the statistical nature of detection models. Model naming also read as an internal catalogue rather than mapping to fraud use cases. > the one phrase that got in the way was "1st world model for detection," which is jargon-y positioning language that sounds like it's borrowed from the LLM world and doesn't tell me anything about accuracy, latency, or fraud outcomes > > — Head of Fraud, Telecommunications, 501-1000 > the model-line naming — DETECT-World, DETECT-3B Omni, PerTh Multimodal, "Signal," "Identity" — reads like an internal product catalogue, not something I can map cleanly onto "this stops fraud X." > > — Head of Risk, Insurance, 1001-5000 > "World model architecture that recognizes deviations from physical reality" is the phrase that stopped me — I don't know what a 'world model' is in this context, how it differs from a standard classifier, or what a 'deviation from physical reality' actually means as a detection signal. > > — Chief Information Security Officer, Contact Centers, 501-1000 > deterministic usually means reproducible/non-probabilistic, but a detection model spitting out an accuracy percentage is inherently statistical > > — VP of Cybersecurity, Financial Services, 501-1000 ### The vertical examples stop at telecom and security, leaving insurance and banking buyers… Respondents said the page lacks an insurance case study and bank-specific compliance and regulatory language; the multi-vertical pitch left their own sector unaddressed. > I'd want an insurance-specific line or case study — something like a named insurer or financial services firm using this for claims-document forgery or call-center verification, not just Telnyx and Okta > > — Head of Risk, Insurance, 1001-5000 > I'd need an insurance-specific use case on the page — claims fraud, forged medical imagery or synthetic voice on a claims call — named the way KYC and contact-centre fraud are named now > > — Head of Compliance, Insurance, 501-1000 > Nothing here mentions SOC2, regulatory frameworks, or bank-specific compliance language, which is what would tell me they understand my actual buying process. > > — Chief Information Security Officer, Financial Services, 1001-5000 ### The accuracy figures are self-reported with no published methodology Five respondents accepted the 99.5% audio / 98% video numbers as stated but said no methodology, false-positive rate, or latency data was shown, and a broken benchmark CSV component made it worse. > Deepfake detection — audio, video, image verification for fraud and identity checks. > > — Vice President of Cybersecurity, Insurance, 5000+ > What would rule it out, or at least stall it, is that the RTF/accuracy chart and CSV data source are literally broken on the page — "Select a CSV file in the component properties" — which is a bad look for a company selling detection accuracy as its whole pitch > > — Head of Fraud, Telecommunications, 501-1000 > the accuracy numbers (99.5% audio, 98% video) are self-reported on their own page with no visible methodology or third-party leaderboard link right there > > — Head of Risk, Insurance, 1001-5000 > the page gives me a headline accuracy number and a pip install snippet, not a false-positive rate, latency under real call volumes, or how it holds up against adversarial/newer generation models > > — Head of Compliance, Financial Services, 5000+ > I've been burned before on a detection tool that didn't survive contact with production traffic and it cost me credibility internally, so a meeting is cheap — I'll take it — but I'm going in to interrogate the 99.5% number > > — Chief Information Security Officer, Contact Centers, 501-1000 > they mention "Podonos benchmark" for audio but nothing concrete for video/image > > — VP of Cybersecurity, Financial Services, 501-1000 > They do deepfake detection — real-time AI to flag synthetic audio, video, and images, with claimed accuracy like "99.5% audio, 98% video," > > — Head of Compliance, Financial Services, 5000+ ### The startup-breezy tagline and unpolished page elements clash with fraud-prevention… Respondents flagged 'Deepfakes are everywhere. So are we' as too flippant for security infrastructure, and said unpolished page elements undercut the enterprise positioning. > the "Deepfakes are everywhere. So are we" tagline is punchy but a little glib for a category I'd be betting fraud-prevention infrastructure on > > — CISO, Telecommunications, 1001-5000 ### Buyers say they cannot act without their own pilot and named platform integrations Respondents wanted a pilot on their own contact-center data to check false-positive rates, and integration proof with platforms like Genesys or Avaya before believing deployment claims. > A pilot on our own call data showing the false-positive rate stays low enough that my fraud team isn't drowning in alerts — if it catches synthetic-voice fraud without doubling their review queue, that's the number that makes this worth displacing anything we currently run > > — Head of Risk, Insurance, 1001-5000 > the page doesn't show me a real deployment story or integration proof with our actual stack (Genesys, Avaya, etc. are just logos) > > — CISO, Telecommunications, 1001-5000 --- ## 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 central proof point is unusable in a procurement conversation.** *(high)* Six respondents took the 99.5% audio / 98% video figures at face value but found no methodology, false-positive rate, or latency, and hit a broken benchmark CSV. Two more demanded their own pilot to check false positives — the numbers buy nothing. - **Clarity about the problem is not clarity about the product.** *(high)* Six respondents grasped the audience and problem from the hero line, yet four said 'world model' and 'deterministic' hid the actual mechanism, with one noting 'deterministic' contradicts how detection models work. The page explains who it is for and not what… - **Credibility rests entirely on borrowed names, not on anything the product demonstrates.** *(high)* The four respondents who found differentiation cited the Okta VP, Telnyx CEO, and the Deutsche Telekom challenge win — logos and awards. Six simultaneously found the performance claims unverifiable, so remove the names and nothing separates this from a… - **The multi-vertical pitch converts breadth into irrelevance for the sectors with the most fraud budget.** *(medium)* Three respondents in insurance and banking found no case study and no compliance or regulatory language for their sector; telecom and security examples left them unaddressed. Breadth signalled the page was written for someone else. - **The one value respondents can restate has no deployment path attached.** *(medium)* Five respondents named real-time mid-call detection as the payoff, but two said they cannot act without integration proof for Genesys or Avaya and a pilot on their own data. The proposition is memorable and unbuyable. - **Unpolished execution actively contradicts the security-infrastructure claim.** *(medium)* Two respondents called 'Deepfakes are everywhere. So are we' flippant for fraud prevention, and a broken benchmark CSV component appeared in the accuracy critique. A vendor selling deterministic detection shipped a page that does not work. --- ## 06 · Who answered | # | Role | Industry | Company size | | --- | --- | --- | --- | | 1 | Head of Fraud | Telecommunications | 501-1000 | | 2 | Head of Risk | Insurance | 1001-5000 | | 3 | Head of Compliance | Financial Services | 5000+ | | 4 | Chief Information Security Officer | Contact Centers | 501-1000 | | 5 | CISO | Telecommunications | 1001-5000 | | 6 | Vice President of Cybersecurity | Insurance | 5000+ | | 7 | VP of Cybersecurity | Financial Services | 501-1000 | | 8 | Head of Fraud | Contact Centers | 1001-5000 | | 9 | Head of Risk | Telecommunications | 5000+ | | 10 | Head of Compliance | Insurance | 501-1000 | | 11 | Chief Information Security Officer | Financial Services | 1001-5000 | | 12 | CISO | Contact Centers | 5000+ | | 13 | Vice President of Cybersecurity | Telecommunications | 501-1000 | | 14 | VP of Cybersecurity | Insurance | 1001-5000 | | 15 | Head of Fraud | Financial Services | 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 2026-09-22, then deleted along with the personas and their answers.