Matt Swulinski on Building and Measuring a Product-Led Growth Engine

Matt Swulinski, head of growth at Viktor and formerly a growth leader at Wispr Flow and Superhuman, joins Harry Stebbings to discuss paid acquisition, measurement, creative production and AI-assisted marketing. His metrics and examples reflect his reported experience rather than independently audited results.

Swulinski thinks product-led growth remains relevant as agents start selecting tools. A company still needs a self-service experience that works for a human, while also understanding how an agent researches and chooses products. He expects strong product fundamentals to help companies adapt, though the future role of agents is uncertain.

At Superhuman, he says, he learned the value of craft: small details in a product can encourage people to keep using it. The company grew through founder-led referrals and word of mouth, but he believes a narrowly defined early audience eventually created a growth ceiling that required new approaches.

His preferred approach borrows from e-commerce: measure whether spending drives a meaningful action, produce varied user-generated creative and coordinate several channels. He says he applied parts of that playbook at Superhuman and Wispr Flow. The contribution of any single tactic to their growth is not independently established in the interview.

Asked when to begin paid acquisition, Swulinski says early testing can provide rapid feedback on messaging, creative and conversion. He does not argue that product and organic work should stop. His view is that paid testing can complement those activities when the product and measurement foundation are ready.

He argues that paid campaigns can test positioning and landing-page conversion faster than waiting only for organic content to spread. The advantage is speed of feedback, in his view, not a substitute for interviewing users or understanding why they stay. A test still needs an explicit budget and interpretable measurement.

He names Meta, Google and lifecycle communication as core elements of an acquisition system: paid social and video, search intent, and follow-up through email, SMS or app notifications. He suggests starting with a small number of channels rather than spreading a limited team across many platforms.

Stebbings proposes a hypothetical $100,000 test budget for a company that raised $3 million. Swulinski says the right amount depends on average revenue per user, unit economics and funnel length, while accepting it as an illustrative starting point. That is not a universal budget recommendation or assurance that paid growth will work.

Before spending, he stresses measurement. Ads on multiple platforms and lifecycle messages can all claim credit for the same customer. He describes analytics and attribution tools used in e-commerce and says software companies often have to connect their own data, product events and business-intelligence layer.

Swulinski estimates many software companies skip important setup, but his percentage is a personal estimate, not a cited survey. He recommends making sure website and product events reach marketing platforms accurately. Without that foundation, a team may misread acquisition costs or attribute the same revenue to multiple channels.

He gives examples such as conversion tracking and match rates. In his account, advertising platforms use the events they receive to find similar customers; incomplete signals can impair optimization. The actual impact depends on the platform, implementation, consent and the quality of the underlying data.

Swulinski says a platform's optimization depends on the events it receives. If only some subscriptions are matched back to the ad system, the platform has an incomplete picture of the users the company wants. He suggests reviewing instrumentation before concluding that an entire channel cannot work.

For an initial test, he recommends choosing an event that matters to the business, such as a subscription, trial or download, and monitoring acquisition cost. The appropriate event may change as the funnel matures. He says the product team must also understand retention and user quality, not just generate cheap traffic.

At Viktor, he describes a more involved conversion: a customer needs to add the product to Slack or Microsoft Teams after onboarding. He says the company can accept a higher acquisition cost because its revenue per customer is higher. These are examples of his company's approach, not audited unit economics.

Asked about lifetime value relative to acquisition cost, he says the acceptable ratio depends on available capital and the business model. He describes one-to-one as a possible early-stage choice for a well-funded company and three-to-one as a common longer-term objective. These are his heuristics, not guarantees of profitability.

He adds that gross profit matters, especially for usage-based AI products. Revenue per customer can vary widely while inference, hosting and other servicing costs grow with usage. A useful model should include those costs rather than counting only the bill from a single model provider.

Two subscribers with the same initial conversion may have very different revenue and inference costs, he notes. Averages can conceal that spread. He recommends working with finance to include hosting, model and other delivery costs so a growth team understands whether additional usage creates value.

On creative spending, he recommends substantial early investment in assets and testing. He says platform targeting increasingly depends on the creative itself, so different messages and visual treatments can help reach different people. His description of platform algorithms reflects his experience and is not a complete specification of their current operation.

He describes a creator program at Viktor, alongside agencies and an internal team, producing many demonstrations and use cases. For a hypothetical $100,000 Meta budget he suggests hundreds of new assets per month. That number is his operating heuristic; creative needs and budgets vary by product and audience.

Creators can make ads for a brand without posting them to their own account, he explains, while partnership ads can also use a creator's distribution. He suggests asking peers which agencies have delivered reliable work. He reports that some young creators earn substantial amounts from performance-based arrangements; those figures are anecdotal.

Swulinski says creator-made ads can be published through the company's ad account rather than the creator's own feed. A creator can therefore work for multiple brands, subject to the terms agreed with each. He distinguishes those assets from collaborations that also use the creator's audience.

Swulinski says a small fraction of creative often attracts a large share of spend. Hooks can disrupt a viewer's pattern, but he cautions against copying only one winning format. A portfolio with different people, settings, openings and use cases can help the team test more audiences and reduce fatigue.

He recommends looking at the overall set of ads instead of declaring one format the permanent winner. Variety can include a different speaker, hook or demonstration, with results evaluated over time. His examples are creative tactics, not evidence that a particular demographic always performs best.

He sees AI as useful for creating variations of existing assets, while expressing dissatisfaction with some fully AI-generated video. Stebbings speculates that platforms may treat AI content differently; the conversation does not provide evidence of a general platform penalty. Their creative judgments are subjective and may change with the tools.

Before scaling a paid campaign, Swulinski wants a product customers already value and enough correctly tracked conversions for the platform to learn. He suggests reviewing early signals after a few weeks and allowing a longer test to refine creative and the funnel. The timing is a rule of thumb, not a controlled guarantee.

If acquisition underperforms, he recommends examining the whole path: the ad, landing-page message, page speed, app-store presentation and conversion tracking. A weak result may arise at several points, and changing only the ad may not identify the cause. Teams should use experiments to distinguish these possibilities.

A landing page should make its purpose and next action clear, especially on mobile. Swulinski asks whether a visitor can understand the product from the headline without scrolling. He also notes that an unusual design can fit a particular audience, so the right experience depends on whom the product serves.

Swulinski suggests reviewing the landing page from the perspective of someone arriving from an ad: can they tell what the product does, see the next step and use the page on their device? He treats the ad and landing experience as one connected path rather than separate projects.

He considers page speed important for user experience and discoverability, including search and answer engines. He predicts slow pages can hurt conversion, but does not present a controlled measurement for a particular site or establish how every search system ranks content.

When a campaign begins to work, he charts spending against acquisition and recurring revenue while accounting for conversion delays. The key question is whether additional spend produces additional customers and revenue. If spending rises without a corresponding lift, the apparent growth may not be incremental.

A growth leader must also understand interactions among channels, he says: paid social, search, newsletters, public relations and product discovery can influence one another. A channel specialist may optimize one account while missing the combined effect. His desired outcome is a coordinated mix rather than a single dashboard's attribution claim.

Asked why Superhuman's growth slowed, Swulinski points to a premium product and an early focus on founders, a finite audience. He says reaching further segments required adapting the product and message, including use cases for sales teams. His explanation is a former employee's interpretation, not a complete causal analysis.

The pair compare another email product's growth. Swulinski points to paid acquisition, a different customer proposition and enterprise sales as possible contributors, while acknowledging that the products serve different markets. He also says Wispr Flow had enterprise adoption through self-service before building a conventional sales team.

He describes enterprise adoption as an additional route for a product-led company: individual or team usage can lead to larger contracts. He says Wispr Flow saw business customers before it had a dedicated account executive. The anecdote does not establish that a sales team is unnecessary in every market.

YouTube requires a different creative treatment from short vertical social ads, he says. At Viktor, the team placed vertical videos inside a landscape template with branding and a call to action, then edited the results for the platform. Some assets worked better than others; he recommends a distinct storytelling program rather than assuming direct reuse always works.

YouTube audiences may be willing to watch a longer demonstration than someone scrolling a short social feed, he says. His team adapts scripts and edits to the placement rather than merely changing dimensions. The appropriate length and opening still need testing for each audience.

He says the team at Wispr Flow once increased a budget roughly fivefold from one month to the next to discover where the system broke. The exercise exposed limits in audiences, creative and channel mix, after which the team pulled back and adjusted. That is his account of a high-risk experiment, not a recommendation to spend beyond a company's means.

When an ad is weak, he distinguishes results that are clearly uneconomic from cases with promising signals that merit more testing. A more tailored landing page or message could change conversion. He recommends setting a boundary, testing alternatives and stopping when the economics do not justify further spend.

Swulinski says Google ads were the strongest paid channel at Wispr Flow in his experience. He describes longer YouTube demonstrations, search and other placements working together to educate potential users before conversion. The interview does not independently measure each placement's incremental contribution.

Audience fatigue calls for finding the next group of users rather than only repeating the same message. At Viktor, he says the team examined early adopters such as agencies, e-commerce businesses and small companies, then tested segments one at a time. He also discusses using AI to support audience research, which still needs validation against real customers.

At Viktor, he says the team kept broad evergreen material running while developing specific messaging for one customer segment at a time. Once it understood a segment's funnel and economics, it could repeat the process for another. He treats audience research as a starting hypothesis, not proof of customer behavior.

Stebbings, who discloses an investment in Viktor, challenges the broad phrase 'AI employee for everyone' as unclear. Swulinski says the initial positioning was meant to introduce a broad category, and that specific customer stories can later explain what the product actually does. Their disagreement illustrates the tradeoff between a broad claim and concrete use cases.

For referrals, he recommends making the program visible, easy and tied to a benefit the user values. He recalls Superhuman's offer of a free month for both referrer and new user, and a Wispr Flow prompt near a usage limit. The useful moment and incentive depend on the product and its economics.

He says Viktor offers credits for referrals or creator activity and describes a share of referred-company revenue returned as credits. He reports examples of customer spending and acquisition costs, but the interview does not independently verify those figures or establish that every customer achieved a particular labor saving.

A credit-based referral reward can reduce a customer's own product bill, he says, while providing the company with another customer. The arrangement only makes sense if the reward and service cost fit within acquisition economics. He offers it as one example, not a suitable structure for every company.

Referral programs fail, in his view, when the benefit is vague or the rules are so complicated that people abandon them. He prefers a clear exchange and a low-friction action. Teams still need to check whether the reward is sustainable and whether referred users become valuable customers.

Paywall timing should follow a moment when a user has experienced meaningful value, Swulinski argues. A limited trial can show a useful workflow, then ask for payment while the value is fresh. He frames this as a design hypothesis to refine, not a fixed rule that every product should use the same limit.

He says free credits should be included in fully loaded acquisition cost alongside marketing spending. Otherwise the apparent price of acquiring a customer excludes the cost of serving the trial. That matters particularly when usage has a material inference cost.

On answer-engine optimization, he says teams should publish useful content and understand where AI systems cite information. He warns against automatically producing large amounts of low-value material. His account of how particular answer engines select citations is an observation and hypothesis, not a verified ranking formula.

He considers YouTube reviews, external coverage and other sources useful because they can explain a product in detail and may be cited by answer engines. Stebbings discloses an investment in a company in this area. Neither speaker demonstrates that a particular review or media placement guarantees a citation or ranking.

Traditional public relations can still build credibility, Swulinski says, even if the aim is not an immediate traffic spike. He views independent coverage as one input into a broader discovery strategy. His description of the relationship between coverage and AI-generated answers remains a hypothesis to test.

TikTok has not yet worked as a paid growth channel for the three companies he discusses, he says, though he leaves open trying again for a suitable audience. That experience contrasts with other founders' reported results and should not be generalized into a claim that the platform cannot work.

What happens if paid spending stops? Swulinski says a mature business should also have a meaningful organic contribution from product referrals, content and other discovery. He proposes periodically reducing spend or using holdouts to test incrementality, while recognizing that a marketing-mix model begins imperfectly and needs refinement.

A marketing-mix model can be improved by deliberately changing spend and comparing the observed result with the model's expectation. Swulinski mentions holdouts and budget reductions as possible tests. External changes and lagged conversions can complicate interpretation, so a single test may not settle attribution.

He describes learning marketing by running an agency across e-commerce, software and enterprise clients. That experience made him comfortable executing as well as directing strategy. He says he personally handled significant paid execution at Wispr Flow for a period; the budget figures are his recollection, not an independently audited account.

The two advocate understanding more than one part of the funnel. Swulinski calls himself a specialized generalist: develop a particular strength, but learn enough about creative, analytics and implementation to understand how the system works. He sees AI tools as useful learning aids, while still requiring a person to test and understand the result.

His hiring priorities have shifted toward systems thinking and the ability to use AI alongside relevant experience. He defines systems thinking as mapping a role's inputs, outputs, administrative work and dependencies, then identifying where automation might help. This is his hiring preference, not proof that experience or specialization no longer matters.

In interviews, he asks candidates how they use AI in workflows, including outside work. He distinguishes a single chat from a process with feedback: if a draft is poor, can the person identify the problem, provide feedback and improve the next result? He looks for evidence of a repeatable loop rather than a list of tools.

He describes a newsletter-sponsorship workflow built at Wispr Flow. In his account, agents help research incoming requests, ask for rates, prepare copy and track performance, while he approves the decision to work with a partner. The workflow illustrates his reported setup; the interview does not independently audit its performance or automation boundaries.

His sponsorship example includes several steps beyond writing copy: gathering rates, reviewing audience fit, preparing links, sending materials and comparing later performance with the cost. He retains an approval point before committing to a partner. The claimed automation reflects his own setup and is not an independent verification of every step.

His earlier system used a local folder structure, written instructions and scheduled checks to surface requests in Slack. He says he learned by asking a coding assistant how to build the workflow, and that early results often broke or produced poor copy. Repeated feedback and human decisions were part of making it more useful.

He reports handling many newsletter partners and other campaign tasks with a lean team, and says much of his daily work passed through the system. Those figures are his account of his own workload. The key lesson he draws is to decompose a difficult process and improve it rather than assume a tool will work perfectly on the first attempt.

Swulinski estimates that very few candidates he meets demonstrate the depth he seeks. He expects smaller marketing teams with capable people and agents to manage more execution in the future. His claim that AI could perform most execution within a few years is a forecast, not a measured result for every team.

He argues that leadership must understand enough of the new workflows to decide where automation is useful, and provocatively questions whether some current teams fit changing roles. That is his management opinion. The conversation does not provide a validated method for evaluating or dismissing individual employees.

Both speakers question a strict separation between business-to-business and consumer marketing: in either case, a person makes or influences a buying decision. Swulinski says Viktor approaches a high-value business product with some consumer-style creative. The appropriate channel and messaging still depend on the buyer and product.

In a quick-fire round, he names affiliate marketing as an underappreciated channel. Unlike an in-product referral, an affiliate may have an audience before becoming a customer and earn a share of revenue for bringing in buyers. He says the program's economics and incentives need to be clear.

He reports that Viktor receives roughly 10%–15% of monthly acquisition through affiliates and offers revenue shares in a similar range. He gives examples of affiliate earnings. These figures are company-reported and do not imply that a new affiliate or another company will achieve the same outcome.

He calls X crowded with similar launch videos and says he has not yet seen software marketers consistently make its paid ads work. He invites counterexamples. His view is based on his experience and conversations, not a platform-wide study.

The speakers discuss the need for more practical training in AI workflows and growth. Swulinski says he learned marketing by doing rather than through a dedicated college curriculum. Their proposed training format and judgments about formal education are opinions, not an evaluation of every course or learner.

He describes a 'session end' workflow he used with a coding assistant: summarize work, decisions and open tasks into an Obsidian knowledge base so later sessions can retrieve context. He sees the accumulated notes as a way to make work compound. The example depends on the user's setup and access permissions.

The session summary he describes records decisions, unresolved tasks and learnings, then connects them with notes from earlier work. He values being able to return to a previous idea without relying on memory alone. Any use of such a system also depends on which files the user chooses to connect.

Stebbings describes his fund using an AI tool to rank calls for an investment committee and says it had performed well in a limited period. Swulinski argues the framework behind an automated ranking matters. Their account does not independently validate the system or demonstrate that its rankings will remain correct.

For a closing lesson, Swulinski returns to product craft, experimentation and understanding users. He says he admires teams that connect thoughtful product work with a clear launch, citing Wispr Flow while acknowledging his own role. Stebbings thanks him for sharing the methods and the interview ends.