Customer Acquisition Strategy for AI Startups

A customer acquisition strategy is the written answer to three questions: which channels you buy attention on, which single event you pay for, and how you will know the money worked. Most early teams have a strong opinion on the first one and nothing at all on the other two. That is why their conclusion, six months and a few hundred thousand dollars later, is that paid does not work for them.

Matt Swulinski has built that engine three times. He is head of growth at viktor.com, the AI coworker that installs into Slack and Microsoft Teams and raised a $75 million Series A led by Accel in May 2026. Before that he ran growth at Wispr Flow, the voice dictation app that raised $280 million at a $2 billion valuation in August 2026, and earlier he ran affiliate and sponsorship marketing at Superhuman, the email client Grammarly acquired in July 2025 and the name Grammarly later took for the whole company. On 20VC he laid out the playbook in full, and his framing is blunt: "distribution to me is the only moat."

His thesis is that software companies should stop inventing their own growth model and run the ecommerce one, where every cent of spend is supposed to map to a purchase or an add to cart, creative is produced in the hundreds, and the channel mix is managed as a portfolio. What follows is that playbook, with the parts worth arguing about marked as such.

Three channels, not ten

Swulinski calls it the core three: Meta, Google, and lifecycle. Meta and YouTube cover high-intent video, Google covers people searching for the problem without knowing your product exists, and lifecycle (email, SMS, push) is the net that catches everyone the first two touched. His claim is that those three alone can carry a company somewhere between its first one and ten million in ARR.

Everything else is a later decision. "Everyone's like, 'Oh, we have to do TikTok and Reddit,'" he says, and the comparison he reaches for is agents: "a lot of people get lost with agents is they like start doing a million things with agents and they do them all poorly. It's the same thing here, right? If you do a million channels and you do them all poorly, it's not going to help you out."

Having raised a $3 million to $5 million seed is his marker for being ready to spend, and the figure on the table in that conversation was roughly $100,000. Treat it as a number big enough to learn from rather than a rule. What it has to buy is a verdict on whether the funnel converts strangers at all.

Measurement is the first channel, and most founders skip it

This is the part of the playbook that costs nothing and gets cut anyway. Ecommerce solved attribution years ago with off-the-shelf products: Triple Whale for deduplicated attribution across channels, Elevar for server-side conversion tracking into Meta, Google, and the rest. Both are Shopify-native, both install in an afternoon.

Software has no equivalent. Every startup builds its own analytics layer, its own warehouse, its own first-click and last-click logic, and most of them stop at product analytics without ever checking whether the ad platforms receive the same events. Swulinski's number for how many teams get this wrong before spending: "I'd say 90% of companies don't do that as a first step."

The failure is mechanical, not philosophical. Meta scores how well your conversion events match real people (match rate); Google does something similar with its own enrichment signals. If half your subscribers arrive as unmatched ghosts, the algorithm optimizing for subscriptions has no idea what a subscriber looks like, so it spends your budget on near-random people and your cost per acquisition goes vertical. "The platforms won't tell you because you have the wrong setup for conversion tracking you don't have measurement and you're spending into the air." His verdict on the industry's favorite conclusion: "most of the time people haven't done the actual setup correctly before they can say paid doesn't work for me."

So the first hire for a paid engine is not a media buyer. It is someone who can instrument the product and the site so that one event, chosen deliberately, arrives complete at every platform you intend to spend on.

Start paid early, but only after 50 conversions

Swulinski disagrees with the standard advice to earn organic traction first. "Paid is the easiest way to validate that you have PLG," he argues, because paid compresses the feedback loop: messaging, positioning, funnel, and landing experience can all be tested inside a week instead of across a quarter of content and user interviews. Do the organic work too, he says, but do not wait for it.

There is still a floor, and it is a technical one. The algorithm needs enough examples to model your buyer, so his threshold is 50 conversions of whatever event you picked before you turn on real spend. Below that you are paying to teach the platform nothing.

Once it is running, he gives two clocks. Within two to three weeks, after a couple of creative iteration cycles, you will know whether you can print money or need to go back to the drawing board. Give the full test three months before you decide. And treat the landing experience as part of the ad, not downstream of it: "If I read nothing else but your headline, do I know what you do?" His most common complaint about startup sites is embarrassingly basic, that "the main CTA is like a scroll and a half away."

One caution worth holding onto, from the other side of the table. a16z general partner Anish Acharya, in the firm's State of AI session, points out that consumer AI has no cheap front door: "we've never had an AI native distribution channel. There's no app store for AI." He also notes the marginal cost problem that makes paid acquisition for free products brutal, using an app he built himself to browse his X timeline, which he says "costs $250 to onboard a new user." His conclusion runs the other way from Swulinski's, that because the big networks now guard against anyone building a channel on top of them, "what founders have to do is actually build a product that has the original network effect which is word of mouth." Both things are true at once. Paid is the fastest validation loop available; it is also the most expensive way to discover your product is not worth telling a friend about.

Creative volume is the targeting now

Meta's Andromeda update changed what a media buyer does. Swulinski's version is that Meta decided to "stop telling me in audience settings in the campaigns who people are" and to read the ad itself instead, analyzing the creative and finding the matching people for you. In his words, "the creative is the targeting," and the media buyers who built careers on campaign structure had to become creative strategists overnight.

Worth being precise here, because his phrasing is stronger than the facts. Meta did not delete targeting. Under Advantage+ audience, custom audiences, lookalikes, age range, gender and detailed targeting became audience suggestions, which Meta prioritizes and then expands past, while locations, minimum age, languages and excluded custom audiences stay audience controls the system must respect. Jon Loomer documents that split from Meta's own product documentation. The operational conclusion survives the correction: creative is now your highest-leverage input, and distinct creative concepts beat narrow audience settings.

Which sets the volume bar, and the bar is high. On a $100,000 Meta budget, he says, "you probably need at least 400 to 500 new creatives a month. Otherwise, you're going to plateau." That is not 500 edits of one video. It is different ages, genders, settings and hooks, because a homogeneous set exhausts itself: pump only the thing that worked yesterday, he warns, and performance craters. His standing requirement is that "you need to constantly be creating net new completely wild weird ideas otherwise your acquisition costs will start to rise."

Nobody staffs that in house. The supply chain he describes is a creator program that pays a percentage of ad spend, five agencies, and an in-house creative team: a couple of hundred creators each shipping three to four videos a week. Because that creative never runs on the creator's own page, one person can make ads for you and for your competitor in the same week, which is why the supply exists at all. Partnership ads are the separate case, since those do post from the creator's account, and he puts roughly 30 percent of spend behind them. He says some of his creators are teenagers earning twenty to thirty thousand dollars a month.

On whether AI replaces that: variations yes, originals no. His view of fully synthetic ads is short, that "full videos that are AI, they're slop", and he gives them maybe 5 percent of an account as deliberate weirdness. On the rest: "in no way, shape, or form is anything in creative to the point where you can replace real people."

Fully loaded CAC, or you are lying to yourself

Here is the line most AI startups get wrong, and it is an accounting choice rather than a marketing one. Free trial credits are marketing spend. "We call this fully loaded CAC. So, our free trial credits get summed with marketing spend. If you're not doing that, then you're not really calculating your acquisition cost." Anything free you give away "is cost and it goes under it is a marketing cost item."

The cost side needs the same honesty. Founders quote their model provider bill and forget orchestration, hosting and the rest of the inference stack, which is why Swulinski argues for measuring against LTV on gross profit rather than LTV on revenue once usage costs are real. If you are still deciding how to charge in the first place, that question belongs with AI pricing before it belongs with your ad account.

On ratios he is deliberately unromantic. Early on, roughly one to one is acceptable if you have raised enough to fund it, because the point of the first phase is users and belief rather than margin. Below one to one you are simply burning money and need to fix the funnel. Three to one is the number most teams aim for at scale, reached as the channels that do not bill you per click start contributing: sponsorships, the other channels you add with size, and the answer-engine work he calls AEO.

Which event you optimize for also decides who shows up. At Wispr Flow the conversion action was the download, deliberately upper funnel because a desktop client and an iOS app are hard to track. At viktor.com it is the Slack or Teams install, far down the funnel, affordable because average revenue per user is high enough to carry a four-figure acquisition cost. The trap in usage-based AI products is variance: one signup becomes a $50 account and the next becomes a $50,000 one, and the bidding algorithm cannot tell them apart from the signup event alone.

That is an argument for choosing your acquisition target by lifetime behavior, not by conversion rate. Jean-Denis Greze, co-founder and CEO of Town and formerly CTO at Plaid, made the same point from the product side on 20VC: "you want to acquire the users in a paying motion because you want them to be for work use cases because you will keep finding more ways for them to use AI to generate business value for themselves." Personal use runs out of jobs to do. Work does not. If your acquisition engine is buying the cheaper signup rather than the expandable one, it is optimizing you into a ceiling. Deciding that in advance is what the ideal customer profile exercise is for.

Find the ceiling on purpose

The thing Swulinski does that most teams will not is deliberately overspend to locate the ceiling. At Wispr Flow they multiplied the budget five times from one month to the next, watched where it broke, pulled back, and kept the map. His rule for that exercise: "go as hard as you absolutely can and then see it blow up and then understand where and how it blew up."

The logic is that time is the scarce input. If you need 30 to 40 percent month over month growth, you cannot spend six months discovering your channel mix politely. Break it fast, learn where audience saturation and creative supply actually bind, then scale back up with the answer.

The metric underneath this is elasticity: chart total spend against acquisition and revenue, and watch whether revenue moves when spend moves, allowing for the lag between a click and a paid conversion. Without that check, spending more feels like working. Once a channel is performing, he says, it is easy to "dial it up to 13 on a scale of 10", and the result if you are careless is that "all of that additive spend could have not been there and you would have gotten the same results and then you're just giving money to Meta and Google."

The same test runs in reverse. Past a million dollars a month, he wants a media mix model, incrementality tests and holdouts, and he treats spending down as a legitimate experiment: "spending down to understand how much drop off you have is actually a good very strong learning moment." His target for a healthy mature mix is roughly 35 to 45 percent of acquisition coming from word of mouth and organic discovery. If turning paid off collapses growth entirely at that stage, the problem is not the paid engine. It is that nobody did the other half of the job.

The channels founders underprice

Three of them, in his ranking.

Affiliate. His pick for most underappreciated, and at both Wispr Flow and viktor.com "it is the highest ROI channel." The mechanism is that you pay a revenue share, in his case 10 to 15 percent, to people who are not your customers and who build content and reviews on their own initiative. At viktor.com he says a good 10 to 15 percent of monthly acquisition arrives this way. The compounding part: "if you set up a strong affiliate program and you bring people into the door and they feel like they can earn money even before they're your customers, they will become your customers."

Referral tied to the limit. Not a generic invite box. The offer appears at the moment the product is about to stop working for the user: at Superhuman, give a month and get a month; at Wispr Flow, the word limit; at viktor.com, credits. The two failure modes he names are intangible rewards (swag, vague goodwill) and over-gamified twenty-tier programs nobody wants to decode.

YouTube reviews, for the answer engines. Long-form video reviews rank on long-tail queries and get cited by answer engines, which is why he tells founders to invest in them early. The three inputs he ranks highest for answer-engine visibility are YouTube, Reddit, and the social narrative written about you by people you do not control. He also reframes press: traditional coverage matters less for its traffic spike than as a credible external citation, so "PR is made for citations." The mirror image is the anti-pattern he watches competitors run, generating 100 to 200 pages a week, much of it machine-written filler, which is exactly what crawlers learn to discount. If you want the detail on that channel, it is its own discipline: start with LLM SEO.

YouTube as a paid channel has a cheaper entry than most teams assume, by the way. His team built a static template with logos, a rating and a call to action, dropped existing vertical story videos into the frame, and shipped the whole UGC library as 16:9 YouTube ads.

Where a paid engine becomes a security problem

This part is ours, not his, and it matters because the fix for a poor match rate is, in plain terms, sending the ad platforms more identifiable data about your customers.

Treat that as a data transfer, because it is one. Four things to get right before you chase match rate:

1. Route conversions server side, and make that the control point. A server-side pipeline is the only place you get to decide field by field what leaves your systems. A browser pixel sends whatever the page happens to contain, including query strings you did not audit.
2. Keep ad pixels out of authenticated product surfaces. Marketing pages are one risk class. Pages behind login, where URLs carry record identifiers and session state, are another. The convenience of one tag everywhere is not worth it.
3. Have a lawful basis per region, and let consent state actually gate the send. Consent that only hides a cookie banner while events still fire is a finding waiting to happen. Each platform also needs its own data processing agreement on file.
4. Treat the creator and affiliate program as third-party access. You are paying strangers a percentage for attributed conversions, which is a fraud incentive: self-referral, coupon farming, cookie stuffing, brand impersonation in ad copy you never approved. Pay on validated and retained revenue rather than on signups, keep the trademark rules in writing, and audit the top earners.

None of this slows a campaign down. Retrofitting it after a customer's security questionnaire arrives does.

The team this playbook needs

Worth saying because the playbook implies a hire, and the hire has changed. For about a year and a half Swulinski was the only person doing execution at Wispr Flow, covering Google, Meta, hundreds of thousands of keywords, ads, experiments, landing pages and between 70 and 120 newsletter partners, by building his own agentic workflows for the repetitive parts.

So what he screens for is no longer channel pedigree. In his reading of the market, "hyper specialists are dying a slow death," and in his interviews "systems thinking is the bar": can you step back from your own job, map its inputs, outputs and boring admin, and automate a layer of it. He puts the share of candidates who clear that bar at under one percent. His hot take about everyone else, and he labels it as one, is "probably fire most of your marketing team that is not a systems thinker", which is a provocation rather than advice, but the underlying shift is real and it is the same one behind revenue per employee as a growth metric. A lean team running systems beats a large team running tasks, and the acquisition function is where that gap shows up first.

What to do this week

1. Write down the single event you are willing to pay for, then verify it arrives complete in Meta and Google. Check match rate and enrichment coverage. If either is weak, fix it before you raise a budget.
2. Count your conversions of that event over the last 30 days. Under 50, spend the week on the funnel instead of on ads.
3. Rebuild your CAC on a fully loaded basis: ad spend plus agency fees plus creator payouts plus the cost of free trial credits and the inference behind them. Divide by new paying customers. Keep the number where your team can see it.
4. Audit your own landing experience on a phone. If the headline does not explain the product and the main call to action is below the fold, that is this week's highest-return fix.
5. Pick one of the three underpriced channels and start it: an affiliate program with a real revenue share, a referral offer attached to your usage limit, or two commissioned YouTube reviews.
6. Before you improve match rate, write down what customer data will leave your systems, through which path, under what consent state, and who signed the processing agreement.

The founders who win distribution over the next two years will not be the ones who spent the most. They will be the ones who could tell, within two weeks and to a dollar, which spending was incremental and which was a donation. That is a systems problem before it is a marketing problem, which is the same muscle every other part of an AI-native company now depends on. The broader version of that discipline, across product, engineering, pricing and go to market, is the AI for startups pillar, and the operating system for running it is the AI Operating System for Startups.

Sources

• How to Build a $100M Growth Engine: Lessons from Wispr Flow & Superhuman (20VC, hosted by Harry Stebbings), the interview this article distills, for the core three, the measurement gap, the creative volume numbers, fully loaded CAC, the elasticity testing, and the affiliate and referral mechanics. Profile: Matt Swulinski.
• Company context for the three growth engines named: viktor.com's $75 million Series A led by Accel (Fortune, Accel), Wispr Flow's $280 million round at a $2 billion valuation (TechCrunch), and Grammarly's acquisition of Superhuman (TechCrunch), after which Grammarly renamed the parent company Superhuman and the mail product became Superhuman Mail.
• How Advantage+ Audience Works, by Meta ads specialist Jon Loomer, for the correction to "creative is the targeting": under Advantage+ audience, custom audiences, lookalikes, age range, gender and detailed targeting are audience suggestions, while locations, minimum age, languages and excluded custom audiences are audience controls Meta must respect. Meta's own Advantage+ audience documentation is the primary source for that split.
• The State of AI: Models, Moats, and the Consumer Renaissance (a16z), for the counterargument that there is no AI-native distribution channel and that word of mouth is the network effect founders have to build. Profile: Anish Acharya.
• Town vs Instinct vs GrokBot: Why the AI Assistant Market Is Not a Bubble (20VC), for the argument that work use cases are the users worth acquiring because the jobs keep expanding. Profiles: Jean-Denis Greze and Town.
• Triple Whale and Elevar are credited by name as the ecommerce attribution and server-side tracking tools Swulinski contrasts with the homegrown stacks software startups build.