Matt Swulinski: SaaS Should Steal Ecommerce’s Paid Playbook — and Rebuild Marketing Around Systems Thinkers

Add to Google Preferred Sources
Matt Swulinski has run growth at three of the most-watched AI software companies — Superhuman, Wispr Flow, and now Victor — and his message will unsettle most B2B marketers: software companies should stop treating paid ads as a later-stage luxury and start running them like ecommerce brands, with hundreds of creators, thousands of ad variations, and measurement wired before the first dollar is spent. In a conversation with host Harry Stebbings, Swulinski describes a playbook built on a “core three” of Meta, Google, and lifecycle channels, validated with roughly $100,000 of a $3–5 million seed round and judged on a three-month go/no-go timeline. The discipline underneath is unglamorous: around 90% of companies skip proper conversion tracking, then blame the channel when costs blow out. After Meta’s Andromeda update, he argues, creative volume is the real targeting signal — roughly 400–500 new assets per month for every $100,000 of Meta spend — and the winning 20% of creators can make $20,000–30,000 a month. The sharpest edge is reserved for talent: fewer than 1% of marketing candidates can describe their own job as a system, and Swulinski says that will flip team structure by 2029, when agents handle 80% of execution and humans keep 20% for strategy. For investors, the implication is that distribution is the only durable moat, and the rare AI-native systems thinker may become as valuable as an AI researcher.

Key Elements

If a startup just raised a $3 million seed round, the least fashionable advice in software is to set $100,000 on fire in paid ads before anyone has heard of the product. Matt Swulinski — who has run growth at Superhuman, Wispr Flow, and now the AI knowledge-worker platform Victor — thinks that is exactly what most founders should do. Speaking on 20VC with host Harry Stebbings, he argues that the playbook that built direct-to-consumer ecommerce brands — hundreds of short videos, thousands of ad variations, paid acquisition from day one — is now the right playbook for software. The people doing nothing but “building brand,” he warns, are conceding the race to a competitor who can clone their website in a week.

The ecommerce playbook comes for software

Swulinski’s core claim is that paid advertising is not a dangerous habit to postpone. It is the fastest instrument a product-led growth — or PLG — company has to find out whether the product can scale at all. He watched the alternative play out at Superhuman, where the legendary founder-to-founder referral motion worked brilliantly until it hit an asymptote — a finite pool of users who looked like the founders — and the company had to build a paid engine late. That experience produced the doctrine he later applied at Wispr Flow, the AI voice note-taking app acquired by Grammarly: treat every ad dollar as if it must equal a purchase, run hundreds of creators, and never mistake organic buzz for a growth strategy.

The reason the ecommerce playbook transfers to business software, in his telling, is that the old B2B/B2C divide is a fiction. “Consumer marketing is not consumer marketing and B2B marketing is not B2B marketing. It’s just marketing. At the end of the day, there’s a consumer who has a buying decision,” he said.

For a $3–5 million seed-stage company, his advice is to avoid scattering spend across ten channels and concentrate on a core three:

The practical sequence: wait until roughly 50 of your chosen conversion events have happened organically, so the ad platforms have a picture of who your best customer is. Then launch, iterate creative for two to three weeks, and hold a full go/no-go verdict at three months. “Paid is the easiest way to validate that you have PLG, that you have a product that can scale in any way, shape or form,” he argued. “Distribution to me is the only moat. And you have to have that strong of a playbook when it comes to marketing, because in today’s world, that’s the only way to succeed.”

The 90% mistake: measurement before the first dollar

Before a cent of that spend moves, Swulinski wants something he says about 90% of companies skip: a tracking stack that proves whether advertising actually becomes revenue. Ecommerce solved this with out-of-the-box tools — he names Triple Whale and Elevar as platforms that ingest ad spend from Meta and Google and report true new-customer revenue in about 15 minutes. Software has no equivalent; every startup hand-builds its own database-to-BI-to-attribution stack, and most never finish.

His first hire in any new role is therefore an analytics leader plus an analytics developer, whose job is to wire website and product signals into the ad platforms before any campaign goes live. The specific failure he targets is the match rate on Meta and the enrichment score on Google — the platforms’ ability to connect a click to a conversion. At a 50% match rate, half of your paying users are ghosts to the algorithm, and the algorithm cannot learn who to target.

“If you have poor conversion tracking, Meta doesn’t know who those people are, and it just randomly targets people — you’ll have a super high CAC, and you’ll say, ‘paid doesn’t work for me.’ I’d say most of the time people haven’t done the actual setup correctly before they can say paid doesn’t work for me,” he said. CAC is the cost of acquiring a customer. “I’d say 90% of companies don’t do that as a first step. Before you spend your first cent, have everything set up.”

The conversion event itself is a deliberate choice that can sit anywhere on the funnel, depending on unit economics. At Wispr Flow it was the app download — a deliberately early event because tracking a desktop client plus an iOS app was technically hard. At Victor, where customers pay much more, the event is the down-funnel moment when a workspace actually adds Victor into Slack or Microsoft Teams — an event so expensive that a couple thousand dollars in acquisition cost is acceptable. The principle: let the platform optimize for the event that matters to your economics, and change it as you learn.

The numbers he works with, using LTV:CAC as the ratio of a customer’s lifetime value to what it cost to win them:

AI-software companies add a cost layer founders routinely miss: token cost. Meta and Google’s algorithms, he notes, go “absolutely nuts” over usage-based B2B products because one user can become a $50,000 annual contract while another is worth $50. He includes inference costs — the bill from Anthropic, Modal, or any model-hosting service — in the cost analysis, and wants a finance person who treats free trial credits as marketing spend. Excluding them flatters your numbers until cash flow catches up.

His method for finding the ceiling is deliberately aggressive: at Wispr Flow in 2025, he scaled the monthly budget fivefold in one step, watched the pieces break, pulled back, and rebuilt with a map of which channels were truly incremental. “You don’t have six months to say we need to be growing 30–40% month over month… go hard to understand what doesn’t work, bring things down, then scale back up with those learnings,” he said. Above roughly $1 million a month, he wants a marketing-mix model and holdout testing, accepting that the model will be wrong at first.

Creative is the new targeting

The most consequential change in paid social, in Swulinski’s telling, is Meta’s Andromeda update: Meta stopped using audience settings to find people and began analyzing the creative itself to infer who you want to reach. The strategic consequence is that the media buyer’s job — tinkering with campaign settings — died, and creative strategy became the entire job. The tactical consequence is that creative volume is now the binding constraint on scale. At a $100,000 monthly Meta budget he wants 400 to 500 genuinely new creatives per month; anything less, he says, and the account plateaus.

He runs a true ecommerce-style assembly line at Victor: a creator program of a couple hundred user-generated content creators paid a percentage of ad spend, five external agencies, and an in-house creative team. Each onboarded creator produces three to four videos a week. The economics concentrate brutally — roughly 80% of results come from 20% of creators, because the platform finds the best ad and floods spend into it.

“We have kids that are like 17, 18, 19 that are making 20, 30k a month just making a couple ads for us,” he said.

What separates the winning 20% is not polish but pattern disruption. In a feed drowning in sameness, the best ad often looks like a mistake — a rough camera shake, a messy start — because it stops the thumb. But he insists on variety across the entire asset package: different ages, genders, settings, hooks, and deliberately strange ideas. “If you only focus on ‘this is what works’ and you pump just that, performance will crater,” he cautioned.

On AI-generated creative he is precise: the superpower is variation at scale, not authorship. He cites Speechify founder Cliff Weitzman’s approach — film one human video, then use AI to change the background and clothing and turn a single asset into thousands of variants. Fully AI-generated videos, by contrast, are what he calls “slop” that audiences and algorithms can smell; he caps them around 5% of an account. Asked whether the platforms actively denigrate AI content, his answer is immediate: “You’re 100% right.”

Google prints, TikTok and X don’t

Across three companies, Swulinski’s channel verdicts are specific and occasionally contrarian. Google was the single best-performing channel at Wispr Flow — the main driver, spanning non-branded search, Performance Max, and YouTube — because it concentrates education and intent in one place. Wispr’s product, which turns voice notes into formatted text, needed explanation, so 30-second to one-minute YouTube videos ran for months, generated hundreds of millions of impressions, and fed the rest of the Google suite. Google also offered what Meta no longer does after Andromeda: fine-tuned control over acquisition cost and geography, including spending differently in India versus the United States for a premium product.

YouTube is the channel most software teams fear because its best ads are landscape 16:9 while their UGC library is all vertical stories. His fix at Victor was practical: an internal app that takes any story video, composites it onto a static branded template with customer logos, a G2 rating, and a call to action, and exports it as a YouTube ad. But the creative fails differently there — YouTube rewards an organic, video-essay feel — so Victor runs a separate team writing and filming YouTube-native content rather than recycling Meta assets.

On the channels that don’t work he is blunt. TikTok “has not worked yet” at any of his three companies, and he has never met a SaaS growth leader who cracked it. X ads are worse, despite — or because of — the platform’s popularity among founders. Stebbings calls X the most polluted growth channel, saturated with formulaic shock-and-awe launch videos that spend more time on theatrics than product. “I have yet to meet a SaaS head of growth or performance marketer that says that X ads print,” Swulinski said.

The most underappreciated channel in his stack is affiliate. At Victor, external partners earn 10–15% of recurring revenue — a partner who lands a $10,000-a-month company makes $1,500 a month from a single sale — and affiliates now produce 10–15% of monthly acquisition at what he calls the highest ROI of any channel. “If you start doing it in the beginning, it ramps,” he said.

What happens between the click and the credit card

Paid can only do so much if the middle of the funnel leaks. Swulinski’s landing-page rule is disarmingly simple: if a visitor reads nothing but the headline, they should know what the product does and want to try it; the main call to action must not be a scroll and a half away; mobile must work. He concedes that deliberately weird sites can succeed — Stebbings points to PostTalk’s chaotic UI, which Swulinski calls calibrated pattern disruption for its ideal customer — but the default failure is vagueness rather than audacity. His own test: “If I read nothing else but your headline, do I know what you do?”

Paywall placement, in his phrase, is “a big game” of sequencing the aha moment, the honeymoon phase, and the ask. The ideal is to make the magic happen as early as possible, let the user sit in the honeymoon, and place the paywall inside that phase at the first genuinely magical discovery. Wispr timed this well: the “rocket-ship moment” when a user’s voice notes convert into polished text, followed immediately by the reality of a word limit. At Victor the live experiment is how many free credits to grant and how to nudge the AI to surface a workflow so useful that the user hits the limit while “this is game-changing” is still fresh. Those trial credits, he repeats, are marketing cost — “fully loaded CAC” — or the real economics are fiction.

Referral design, he argues, is the difference between a growth engine and a dead feature. His three companies show the range:

The two failure modes he warns against are intangibility — “refer your friends” with no clear benefit — and over-engineering — twenty gamified tiers nobody will read. The winning mechanic is always the same: make it tangible, make it easy, and surface it at the moment of felt value. If a product is genuinely loved — Superhuman’s every element hyper-refined, Victor’s eight-person teams paying $15,000–20,000 a month because the AI replaced hiring — then the customer doubles as the sales force. The mature end state he targets: 35–45% of acquisition flowing from organic, word of mouth, and referral, so that switching off paid does not collapse the company.

The quiet rise of answer-engine marketing

The newest layer of the stack is answer-engine optimization — making sure AI answer engines describe your product favorably when someone types “what is the best tool for…”. Swulinski notes that this discipline has reverted to classic SEO’s volume logic: he audits competitors’ sitemaps and sees AI-native companies publishing 100–200 pages a week, most of it what he calls generated AI slop that signals what not to do. The content that gets cited must be genuinely valuable, because the AI crawler ingests it and your phrasing becomes what gets written about you.

The highest-leverage answer-engine asset for a young founder, counterintuitively, is YouTube reviews — long-form, long-tail-ranking videos that are “really high citation on ChatGPT.” Next come Reddit and the wider social narrative, followed by your own site. Traditional PR has not died but has changed jobs: the TechCrunch feature and the Product Hunt launch are now founder initiation — table stakes — while the real purpose of PR is the citation trail it creates. “The more of a non-owned narrative is being talked about positively about you, the better off the product is,” he said.

Page speed feeds directly into this: it is a core ranking signal for both SEO and AI crawler visits, and every 100-millisecond improvement is a measurable conversion gain.

The sub-1% hire and the team that runs itself

The sharpest edge of the conversation is reserved for marketing talent. Swulinski’s hot take — delivered with full seriousness — is to fire most of your marketing team and rebuild around systems thinkers. In his experience the gap between AI-native workers and everyone else is widening so fast that former B players are becoming D players, and the delta shows within weeks. His estimate of how many candidates clear the bar: less than 1%.

“If they’re AI native or a systems thinker and can deconstruct what makes their job hum, that person plus experience will out compete someone that just has experience and is not an AI native,” he said. “Fire most of your marketing team that is not a systems thinker, stop brute forcing people in, hire the right people into the roles, because the role has changed — the JD is no longer the same as it was a year ago.”

What he tests for is not tool familiarity but systems thinking: can you step one degree of separation from your own tasks, map every input and output of your job, and design agents around the high-leverage parts? Agents, in his usage, are AI systems that do not just answer questions but execute multi-step work on their own. His interview question is deceptively personal: how do you use AI workflows in your personal life? A bad answer is “I have a context-rich chat thread in ChatGPT” — a context layer, not a workflow, because it lacks a feedback loop. A good answer is a self-improving loop: when the AI returns slop, the people who get frustrated and write the document themselves have failed; the people who feed the critique back into the system pass.

He demonstrates the bar himself. At Wispr Flow he built a marketing operating system without formal coding — “I don’t code; I learned how to do this by asking how to build an OS,” he said — using Claude Code on a laptop, a folder-and-skill-file structure, and scheduled checks that read his email at 9, noon, and 6. The system automated the entire newsletter-sponsorship function: an inbound sponsorship email triggers research on the partner’s audience and rates, a first-pass negotiation, contract ingestion, copywriting, link creation, conversion tracking, and performance analysis — with the agent making the renewal decision based on whether cost-per-thousand and conversion targets were hit. At peak, he personally managed 70 to 120 unique newsletter partners and ran a budget of $3–5 million a year as the only person doing execution, until December 2025. He estimates 90–95% of his daily work now flows through that AI pipeline.

His compounding trick for anyone, regardless of tool: a “session-end skill” that analyzes each work session, distills decisions and open tasks, and writes them into an Obsidian vault. Because Claude sessions do not persist memory, the vault becomes an external memory that connects today’s problem to last week’s dead end. “If you don’t know how to do something, goddamn ask it. Have it tell you. Give me everything step by step that I need to do and go and do it,” he said.

The structural prediction is the investment thesis behind Victor: today the split is roughly 80% manual work and 20% agents; in three years it inverts, and companies become “like a board of directors where 20% of the work is the strategy and the thinking and then agents just do 80% of the execution.” The sub-1% scarcity is not a hiring crisis, he argues, but a restructuring, because one genuinely systems-thinking person plus a suite of agents replaces a five-person influencer marketing team — his example is a Grammarly-style function that shrinks from three to five people down to one person plus an agentic suite capable of scaling to 100,000 influencers a month.

The stakes mirror the broader AI labor market. Business Insider recently described how Cursor’s recruiters build a private Slack channel for every target candidate, and how Meta’s chief executive once personally sent soup to an OpenAI employee he was courting. Swulinski’s version of that scarcity is specific to marketing, and Stebbings matches it with a proof point from his own fund: every deal call is now graded by an AI grader that stack-ranks companies across five variables, and in 12 weeks and more than a thousand companies, the system’s prioritization has never been wrong. Stebbings predicts marketing roles will have their own “ML crazy moment” in the next year — millions of dollars for top talent — once the market realizes marketing unicorns are as rare and valuable as AI researchers.

The through-line of Swulinski’s argument is that distribution is the only durable moat in a market where a rival can clone a product in a week, and the scarce resource is the systems thinker who can run a self-improving machine rather than a task list. For investors, the implications cut two ways: companies that wire measurement before spend and treat creative as the targeting signal may out-execute better-funded rivals, while the talent market is about to get expensive in ways the industry has not priced in. The unresolved experiment is his own company’s positioning. Stebbings pushes back hard on Victor’s “AI employee for everyone” tagline — “I don’t know what the f*** that means” — and Swulinski concedes the framing works only once concrete use cases invert it. Whether a horizontal AI-knowledge-worker story can convert into the customer-specific funnels he describes, whether the 35–45% organic mix materializes, and whether agents really absorb 80% of knowledge work by 2029 are the live questions. If the last one hits, the marketing team of the future is a board of directors, and the sub-1% systems thinkers just became the most valuable hires in software.

Full content available at:How to Build a $100M Growth Engine: Lessons from Wispr Flow & Superhuman | Matt Swulinski

• Whale Transactions Surge Across Altcoins, Led by 500% Jump in Humanity Protocol
• S&P 500 Profit Surge Hits 50% as Broad Earnings Strength Eases AI Reliance Fears
• WhiteFiber Analyst Upgrades Highlight Shifting Views Across Wall Street
• Guitars and Custom Slack Channels: Silicon Valley’s AI Talent War Has Moved Beyond Money

Once added, BigGo Finance appears first in Google Search Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global financial news first.