We Hit $1M MRR With Gen AI Before ChatGPT Went Viral: Here's What Actually Mattered
16 June 2026
When we crossed $1M MRR in August 2023, most people assumed we'd ridden the ChatGPT wave. We hadn't. Our AI product launch strategy was already in motion before the hype cycle existed, and the lessons we learned are more useful precisely because the market wasn't doing half the work for us.
I'm Saad Tariq, a Technical PM who built the analytics infrastructure for Imagine Art from day one and watched every consequential decision from the inside. This is the account I wish had existed when we were building in the dark: specific metrics, real financial decisions, and the product choices that actually moved the needle on our path to eight figures ARR.
This isn't a post-hype retrospective. It's a practitioner's account of building AI products before pre-ChatGPT AI success was a category people celebrated, when you had to manufacture your own momentum and every percentage point of margin was a deliberate choice.
What Actually Matters for AI Startup Success?
Why Pre-ChatGPT AI Success Stories Matter More Than You Think
The companies that built and scaled generative AI products before November 2022 had to solve problems that most AI startups today never encounter. There was no category tailwind. There was no press cycle writing "AI is the future" every week. There was no mainstream user who understood what image generation even was.
While VCs like Sequoia discuss Act Two of generative AI, the companies that survived Act One have different lessons to share. The survival skills from that period, specifically how you validate demand when the market doesn't understand your product yet, how you price when there's no benchmark, and how you build retention when users have no prior mental model, are exactly what matters most when you're building a differentiated AI product today.
The hype cycle has made AI startup growth feel inevitable in retrospect. It wasn't. We made specific, sometimes counterintuitive decisions that compounded into the outcome. Understanding those decisions is the point.
The Three Metrics That Predicted Our $1M MRR (Not DAUs or Downloads)
Early in the build, I made a call that shaped everything downstream: I refused to let the team optimise for downloads or daily active users as primary success signals. Both metrics are gameable and neither tells you whether your product is doing something genuinely valuable.
The three metrics that actually predicted our trajectory were second-session return rate within 48 hours, first generation completion rate, and the time elapsed between a user's first generation and their first social share.
Second-session return rate told us whether the initial experience created a reason to come back. First generation completion rate told us whether users understood the product well enough to get a result. Time to first social share told us whether the output was good enough that users wanted to attach their identity to it publicly.
These were leading indicators. Downloads were a lagging vanity signal. Once we had conviction in this framework, we stopped optimising for the wrong things.
How Did We Build a Successful AI Product Before the Hype Cycle?
December 2022: Launching When Nobody Was Looking
We launched in December 2022. ChatGPT had just gone viral days earlier, but our product was an AI image generation platform, not a chatbot, and the mainstream conflation of "AI" into a single category hadn't happened yet. We were building in a niche that most consumers didn't know existed.
That context matters because it meant we had to earn every user. There was no organic press surge. There was no app store featuring AI tools because it was trendy. There was no analyst report our enterprise prospects could point to as justification for buying.
What we had was a product that produced genuinely impressive visual outputs and a small cohort of early adopters who were already active in creative communities online. We built for them first and let the analytics tell us when we'd found something repeatable.
The Imagine Art product we shipped in that window wasn't perfect. But it was complete enough to generate the behavioural signal we needed to make decisions with confidence.
The Analytics Infrastructure That Made Everything Else Possible
As the product rolled out, so did a basic version of an analytics event tracking schema. This is the decision I'd make again without hesitation.
The schema was designed around user intent, not just user action. We tracked not only what users did, but the sequence in which they did it, the time between steps, and whether they completed or abandoned at each stage. Every generation event carried metadata: model used, prompt length, style selection, output resolution, and whether the user modified the result or accepted it immediately.
At launch, this felt like over-engineering. By May 2023, when we hit our first growth plateau, it was the only reason we could diagnose what was actually happening.
The event schema matured as the product did. We added social share events, community interaction events, and subscription state change events as those features shipped. By the time we needed to understand why a cohort of users was churning, we had six months of clean behavioural data to interrogate.
Most teams instrument analytics reactively, after they notice a problem. We instrumented proactively, before we knew what questions we'd need to answer. The asymmetry in insight quality between those two approaches is enormous.
What's the Right Generative AI Business Model When You're Scaling Fast?
Why We Chose 5-10% Margins Over Profitability
From December 2022 through the end of Q1 2023, we operated at 5 to 10% margins. That was not an accident or a failure of financial discipline. It was a deliberate strategic call, and it was the right one.
The logic was straightforward: momentum is the hardest thing to manufacture in consumer AI. You can buy users, but you can't buy the compounding effect of users who return, generate, share, and bring other users with them. To get that flywheel moving, you need volume. Volume in the early days required aggressive spend on acquisition, infrastructure, and model quality that we weren't going to recoup immediately.
Profit was not part of the conversation in that window. This made some stakeholders uncomfortable. The conventional instinct, especially coming from a SaaS background, is to demonstrate unit economics early and expand from a position of margin health. We made the opposite call.
The reasoning was specific to the moment. We were in a narrow window where generative AI image tools were genuinely novel, where early adopters were actively seeking the best product in the category, and where the cost of not capturing them was permanent. Users who found a competitor first and built habits around that product were not coming back. The window to be the default choice for creative AI users was measured in months, not years.
So we bought the window. We spent aggressively on paid acquisition, kept our credit pricing competitive to the point of thin margins, and prioritised volume over profitability. We had the conviction that the behavioural data we were accumulating, the cohort quality we were building, and the community momentum we were generating would compound into something that justified the margin sacrifice.
It did.
The Unit Economics Decision That Bought Us Momentum
The specific mechanism was credit pricing. Generative AI image products sell on a credit model: users purchase credits and spend them on generations. Your unit economics depend entirely on the relationship between credit cost, compute cost per generation, and the conversion rate from free to paid.
We priced credits aggressively low relative to compute cost in the early months. This meant we were subsidising user behaviour to drive volume. The bet was that volume would produce the behavioural data and community content that would make the product defensible.
By the time we approached $700K MRR in Q1 2023, the credit pricing was starting to create margin pressure. But the community flywheel was also starting to spin. Users were sharing outputs, which was driving organic acquisition, which was reducing our paid acquisition dependency. The margin sacrifice in month one was buying organic growth in month four.
This is the unit economics logic that most generative AI business model frameworks miss: in a consumer product with strong social sharing mechanics, early margin compression can be a growth investment, not a failure mode. The precondition is that you have the analytics to know whether the flywheel is actually spinning.
We did. That's what made the decision defensible rather than reckless.
How Do You Actually Validate AI Product-Market Fit?
To validate AI product-market fit, track three behavioural signals before any revenue metric: second-session return rate within 48 hours of first use, first generation completion rate, and time between a user's first generation and their first social share. If users return within 48 hours, complete their first generation, and share the output publicly, you have genuine product-market fit. Optimise for these before DAUs or downloads.
Beyond Traditional SaaS Metrics: The Behavioural Fingerprint
Traditional SaaS product-market fit validation relies on retention curves, NPS, and churn rate. These are lagging indicators. By the time your retention curve tells you something is wrong, you've already lost the cohort.
AI products have a different validation problem. Users often don't know what they want until they see what the product can produce. The first session is frequently exploratory and inconclusive. The signal that matters is what happens after that first session.
What I discovered by interrogating our Mixpanel data was that paid converters had a distinct behavioural fingerprint from users who churned. It wasn't just that they used the product more. It was the sequence and timing of specific events that differentiated them.
Converters consistently completed their first generation within the first session, returned within 48 hours for a second session, and shared at least one output within their first week. Churners typically abandoned their first generation attempt, or completed it but never returned for a second session.
This fingerprint became our north star metric cluster. We optimised the onboarding flow to maximise first generation completion. We designed re-engagement notifications around the 24-hour mark to pull users back before the 48-hour window closed. We made social sharing frictionless because time to first share was a leading indicator of long-term retention.
The Event Sequence That Predicted Paid Conversion
The specific event sequence that predicted paid conversion with the highest confidence was: account creation, first prompt submission, first generation completion, output download or save, return session within 48 hours, second generation, social share.
Users who completed this sequence within their first week converted to paid at a rate significantly higher than the baseline. Users who dropped out at any step before social share converted at rates that fell off sharply.
This told us something important about the product's value proposition. The value wasn't in generating images. The value was in generating images worth sharing. Our product had to produce outputs that users were proud to attach their name to publicly. That reframed how we thought about model quality, style options, and output resolution.
Every product decision from that point was evaluated against a single question: does this increase the probability that a user generates something they want to share?
Second-Session Return Rate vs. Retention Curves
Retention curves are measured in weeks and months. Second-session return rate within 48 hours is measured in hours. For a consumer AI product, the 48-hour window is where you win or lose the user.
The reason is habit formation. If a user returns within 48 hours, they're beginning to build a behavioural pattern around your product. If they don't return within 48 hours, the probability of a meaningful long-term relationship drops sharply. We saw this in the data clearly: cohorts with high 48-hour return rates had retention curves that were structurally superior at 30 and 90 days.
This is the metric I'd tell every technical founder to instrument before they launch. Not because it's exotic, but because it's early. It gives you signal you can act on in week one rather than week eight.
What Does the Path to $1M MRR Look Like for AI SaaS Companies?
Q1 2023: Approaching $700K Through Aggressive Spend
By the end of Q1 2023, we were approaching $700K MRR. The growth was real but it was expensive. We were spending heavily on paid acquisition across iOS and Android, operating at thin margins, and the credit pricing model was creating pressure.
The AI revenue growth in this period was driven primarily by paid acquisition efficiency. We had gotten good at identifying the channels and creative formats that produced high-quality users, meaning users who completed the behavioural fingerprint sequence. We were not optimising for lowest cost per install. We were optimising for lowest cost per user who completed a second session within 48 hours.
That distinction in optimisation target made our acquisition unit economics look worse on a surface level but significantly better on a 30-day LTV basis. It also meant we were building a higher-quality cohort that would compound through community behaviour.
May 2023: The Plateau That Forced Us to Understand Our Product
In May 2023, growth stalled. We hit a plateau that lasted several weeks and forced a genuine reckoning with what we actually knew about our product.
The Mixpanel infrastructure I'd built became critical here. We were able to segment the May cohort against earlier cohorts and identify the behavioural differences. What we found was that acquisition quality had drifted. We'd expanded to channels that were producing users who looked fine on surface metrics but were not completing the behavioural fingerprint sequence at the same rate as earlier cohorts.
The plateau was not a product problem. It was an acquisition quality problem. We tightened our channel mix, cut spend on underperforming sources, and accepted a short-term volume reduction to restore cohort quality.
This is why analytics infrastructure matters before you need it. In May 2023, we needed it urgently. Because we'd built it in December 2022, we had the data to diagnose the problem in days rather than months.
August 2023: Crossing $1M MRR
We crossed $1M MRR in August 2023. The scaling AI SaaS milestone felt significant, but the more important observation was how we got there.
The community flywheel was spinning by this point. Organic acquisition was contributing meaningfully to growth, which meant our effective acquisition cost was declining even as absolute spend remained stable. The margin compression from the early months was unwinding as the organic mix improved.
The product had also matured. We had a clearer understanding of our core user, their workflow, and the features that drove retention. The community infrastructure we'd invested in was generating the social proof and shared content that made the product sticky in ways that model quality alone never could have.
Eight figures ARR wasn't the result of a single good decision. It was the compound result of a consistent analytical framework, a clear conviction about where to spend in the momentum window, and a product strategy that prioritised community over technology.
How Do You Build Defensible Moats When Foundation Models Are Commoditizing?
Why We Invested in Community Infrastructure Over Model Quality
Early in the product's life, there was an internal debate about where to invest engineering resources. The case for model quality investment was intuitive: better outputs mean happier users mean better retention. It's a clean causal chain.
I argued against it as the primary investment vector, and here's why: every competitor we had access to the same foundation models we did. Stability AI, Midjourney, and the other players in the space were all building on similar underlying infrastructure. Model quality was going to converge. Whatever edge we built through model investment would be temporary.
Network effects frameworks are useful, but the moat we built looked different in practice. The moat that was actually defensible was community: the accumulation of user-generated content, social behaviour, shared workflows, and community identity that competitors couldn't replicate by switching models.
The specific product decision that embodied this thinking was the Community Feed. As PM, I shipped this feature with a clear strategic intent: give users a reason to engage with each other's outputs, not just their own. The feed surfaced high-quality generations from across the user base, allowed users to like and comment, and created a social layer on top of what had been a purely individual creative tool.
The Community Feed changed the product's retention mechanics. Users were no longer just returning to generate. They were returning to see what others had created, to get inspired, to participate in a creative community. The product became a place, not just a tool.
This is the distinction that VC frameworks often miss when they talk about AI moats. Technology moats in foundation model products are temporary. Workflow and community moats are structural. When your users have invested their creative identity in your platform, when they have followers, when they have a portfolio of shared work, the switching cost is not technical. It's social and psychological.
Model quality gets you to the door. Community keeps users in the building.
What Go-to-Market Strategy Works for B2B AI Product Launches Post-ChatGPT?
What Early AI Adopters Got Right (And How to Catch Up)
The early adopters in the AI image generation space, the users who found products like ours in late 2022 and early 2023, shared a common characteristic: they were already active in creative communities. They were on Discord servers for digital art, on Reddit communities for prompt engineering, on Twitter sharing AI outputs before most people knew what those were.
These users were not waiting for a product to be perfect. They were willing to invest effort in learning a tool if the ceiling of what it could produce was high enough. They valued creative control and output quality over polish and ease of use.
The AI product launch strategy that worked for reaching this audience was distribution through the communities they already inhabited. We didn't try to create a new community from scratch. We participated in existing ones, built credibility through genuine product quality, and let users advocate for the product in their own language.
For B2B AI product development teams trying to reach this audience now, the dynamic has shifted but the underlying logic hasn't. Your early adopters are still community-native. They're in Slack groups, on LinkedIn, in practitioner forums. The channel is different; the approach is the same.
Channel Optimisation Across iOS and Android Privacy Constraints
The post-iOS 14.5 privacy environment made mobile acquisition significantly harder for consumer AI products. Attribution became noisier, ROAS calculations became less reliable, and the ability to optimise for downstream behavioural events rather than installs required more sophisticated measurement infrastructure.
Our approach was to use our Mixpanel data to build probabilistic cohort models that estimated LTV from early behavioural signals. Rather than waiting for 30-day retention data to evaluate a channel, we used 48-hour second-session return rate as a proxy for cohort quality. This let us make channel allocation decisions in near real-time rather than waiting for lagging retention signals.
On Android, attribution was cleaner and we were able to run more aggressive optimisation toward downstream events. On iOS, we relied more heavily on creative quality and landing page optimisation to attract the right user profile organically, then used behavioural signals to validate whether the cohort was performing as expected.
The practical lesson for B2B AI product development teams is that mobile attribution in the current privacy environment requires you to have strong first-party behavioural data. If you don't have the analytics infrastructure to measure cohort quality from early signals, you're flying blind on channel allocation.
What Should Technical Founders Do Differently When Building AI Products?
Identify Your Momentum Window and Have Conviction to Buy It
Every AI product category has a momentum window: a period where the combination of novelty, early adopter enthusiasm, and competitive landscape creates an outsized opportunity to establish a dominant position. This window is always shorter than it looks from the outside and always more expensive to exploit than you expect.
The technical founder AI strategy mistake I see most often is optimising for the wrong things during this window. Teams spend the momentum window tuning model quality, refining UX details, or waiting for better unit economics before scaling acquisition. By the time they're ready to move, the window has closed.
The correct call is to identify the window, have the conviction that it's real, and spend aggressively to capture it. This requires accepting margin compression and acquisition costs that look irrational by traditional SaaS standards. It requires trusting your leading indicators over your lagging financials.
We made this call in December 2022. It was uncomfortable. It was the right decision.
Let Product Analytics Mature Before You Optimise
There's a temptation to start optimising the moment you have any data. Resist it. The first four to six weeks of data from a new product are noisy, incomplete, and often misleading. Your event schema has gaps you haven't discovered yet. Your user segments haven't separated into meaningful clusters. Your cohort sizes are too small for statistical confidence.
The better approach is to instrument comprehensively from day one, let the data accumulate, and begin optimisation once you have enough signal to distinguish patterns from noise. This requires patience that most teams don't have, but it produces decisions that are actually grounded in reality rather than early noise.
I built the Mixpanel schema before launch and spent the first two months watching rather than reacting. By month three, I had enough data to make decisions with genuine confidence. That patience paid off at the May 2023 plateau, when we needed clean historical data urgently.
Focus on Leading Indicators, Not Vanity Metrics
Downloads, DAUs, and total registered users are the metrics that look good in board decks. They are not the metrics that predict whether your product is working.
The leading indicators that matter for AI products are behavioural and early: second-session return rate within 48 hours, first generation completion rate, time to first social share. These tell you within the first week of a user's life whether they're on a trajectory toward long-term retention and paid conversion.
Build your reporting around these. Optimise your product against these. When you hit a growth problem, interrogate these first.
The teams that scale AI SaaS products successfully are the ones who build a culture of leading indicator obsession early. It's harder to do than tracking downloads. It's also the only approach that gives you enough time to course correct before problems compound.
This journey from zero to $1M MRR taught me that AI startup growth is not primarily a technology problem. It's an analytical problem, a conviction problem, and a product strategy problem. The technology is the entry ticket. Everything else is the game.
If you're building in the current AI environment and want to discuss the specifics of analytics infrastructure, unit economics, or product strategy, my broader portfolio including enterprise data infrastructure work and other case studies is available for context. The frameworks that worked in 2022 and 2023 are still applicable. The window is different; the fundamentals aren't.
Written by Saad Tariq, Technical PM specialising in AI products and data infrastructure.
Frequently Asked Questions
How long does it take to reach $1M MRR with an AI product?
We reached $1M MRR in approximately eight months, from December 2022 to August 2023. The timeline depends heavily on category timing, acquisition spend, and cohort quality. Products launching into an established category with strong distribution can move faster. Products launching into new categories, as we did, typically need three to four months to validate their behavioural signals before scaling acquisition aggressively. The honest answer is that timeline is less important than the quality of the cohort you're building. A lower-quality $1M MRR is structurally weaker than a higher-quality $700K MRR.
What metrics matter most for AI startup success?
The metrics that predicted our success were second-session return rate within 48 hours, first generation completion rate, and time between first generation and first social share. These are leading indicators that tell you within a user's first week whether they're on a trajectory toward long-term retention and paid conversion. Downloads, DAUs, and total registered users are lagging vanity metrics. They tell you what happened; they don't tell you what's about to happen.
How do you validate product-market fit for AI products?
Validate AI product-market fit by tracking three behavioural signals before any revenue metric: second-session return rate within 48 hours, first generation completion rate, and time to first social share. If users return within 48 hours, complete their first generation, and share the output publicly, you have genuine product-market fit. These signals appear within the first week of a user's life, giving you enough time to act before problems compound. Traditional SaaS metrics like NPS and retention curves are too slow for AI product validation.
Should AI startups prioritise profitability or growth in early stages?
In a momentum window, prioritise growth. We operated at 5 to 10% margins from December 2022 through Q1 2023 because momentum is the hardest thing to manufacture in consumer AI. Profit was not part of the conversation. The precondition for this decision is having the analytics infrastructure to know whether the growth you're buying is building a high-quality cohort. Aggressive spend without cohort quality measurement is reckless. Aggressive spend with clear leading indicator validation is a strategic investment in a time-limited window.
How do you build a defensible moat in AI when models are commoditising?
Build community and workflow moats, not technology moats. Every competitor has access to the same foundation models. Whatever model quality edge you build is temporary. The moats that are structurally defensible are community identity, user-generated content accumulation, social behaviour, and workflow integration. We built the Community Feed specifically to create social switching costs that model improvement could never replicate. When users have followers, shared work, and creative identity invested in your platform, the switching cost is psychological and social, not technical.
What's different about building AI products versus traditional SaaS?
Three things are fundamentally different. First, the validation timeline is compressed: AI product-market fit signals appear in days, not months, if you're measuring the right things. Second, the margin structure in the growth phase looks wrong by traditional SaaS standards because you're often subsidising behaviour to drive the community flywheel. Third, the moat logic is inverted: in traditional SaaS, the product is the moat; in AI SaaS, the community and workflow around the product is the moat because the product layer is increasingly commoditised.
How do early AI adopters maintain their advantage post-ChatGPT?
Early adopters maintain their advantage through accumulated community capital, not technology. The users who built creative workflows, community followings, and content portfolios on early AI platforms have switching costs that late entrants can't easily overcome. For product teams, the implication is that the community infrastructure you invest in now compounds over time. The technology advantage erodes; the community advantage grows. Teams that understand this shift their investment from model quality to community features, workflow depth, and social mechanics.
What unit economics should AI SaaS companies target?
In the momentum window, target cohort quality over margin percentage. We ran at 5 to 10% margins deliberately to drive volume and community flywheel effects. The unit economics that matter are 30-day LTV relative to acquisition cost, measured by cohort quality signals rather than surface install metrics. Once the organic acquisition mix improves as a result of community flywheel effects, margin naturally expands. The mistake is targeting healthy margins before the flywheel is spinning, which usually means sacrificing the volume needed to spin it in the first place.
