On June 7, I gave a talk at Qishifu AI Salon Demo Day, session 15, in Liangzhu.
The theme of the event was going global and growth in the AI era. My topic was:
“Zero-Budget Overseas: How Mew.Design Got Paying Users in 80+ Countries.”
Honestly, when I first saw that title, I hesitated.
It’s the kind of title that wants to become a growth story: how we did SEO, how we cold-started, how we found overseas paying users with no budget, how we built traffic from zero.
All of that is fair game, and it’s what people most want to hear.
But if that’s all I talked about, I’d be missing the point.
Because the most interesting thing about the Mew Design journey isn’t that we found some magical growth hack. It’s that a very small team, in the AI era, can genuinely build a global product in a way that would have been hard to imagine before.
So this piece is a supplement to that talk. Not a transcript, not an event recap. More like me re-walking the whole path, one more time.
01 - Talk about the product first, otherwise growth means nothing
Mew Design is an AI Design Agent that my co-founder and I built.
In plain terms: like hiring a design agency, a user just tells it what they need, and it handles understanding, generation, revision, and delivery.
People often ask: text-to-image is already this good — why build Mew Design?
I’ve been asked this many times, and my answer keeps getting simpler: because ordinary people’s design problems still aren’t solved.
What many don’t realize is that AI image design is an incredibly crowded space. By all logic, a small team has no business entering a red ocean like this. But our bet back then was: the more crowded the space, the narrower you have to cut.
When we started, text-to-image had three hard, obvious pain points:
- Inaccurate text rendering — the words on a poster would come out garbled;
- A strong “AI look” — generated images clearly weren’t usable;
- No 4K/8K output — couldn’t meet the requirements of production environments like printing.
We didn’t build “yet another general AI design tool” to fight everyone. We nailed our target user down to one type: people who need to actually put the design out into the world — small businesses making a hiring poster for their coffee shop, producing printed materials.
This cut has a real advantage: businesses want something they can ship, not something they can play with. The need is concrete and hard. Once you nail the output in one scenario, word of mouth and repeat usage become very real.
They don’t care which model you called in the background. They don’t care how many tool runs it took. They only care whether the final result is usable.
So from day one, we resisted turning Mew Design into a model showcase.
There’s no model picker in the UI. We never wanted users to learn prompting first.
That looks like it goes against AI-product common sense.
But for our users, that’s exactly the responsibility the product should own — users bring a design need, not a multiple-choice question about models.
02 - Why global, and why SEO
We chose to go global not because foreign markets are romantic, but because it was more like having no choice.
The domestic market isn’t friendly to small teams early on. Traffic is more expensive, everyone is grinding, marketing costs are high, and compliance can eat months. Without a budget, trying to break through with paid acquisition or big campaigns is basically hopeless.
Our budget back then went almost entirely into the product and tokens. Not out of restraint — there genuinely wasn’t much money to burn.
But going global in the AI era is different from before.
Going overseas used to be expensive: understanding user behavior in different countries, localization, competitive research, studying foreign aesthetics and assets, writing landing pages — all of it depended on experience, consultants, or long trial and error. Now AI can fill in that first layer of understanding for you. It can’t guarantee you get it right, but it lets you start faster.
This matters a lot for small teams. Often, small teams don’t die because the direction was wrong. They die because they got dragged under by upfront costs before they could even reach the core question. AI at least gives us a shot at reaching that core question faster.
As for why we picked SEO in the end, the reasons aren’t mystical.
First, we had no money. Ads and influencers are doable, but the cash-flow pressure on a small team is too high, and a lot of that traffic comes fast and leaves fast.
Second, SEO is a long-term game, and it brings inbound traffic. With ads and influencers, you push something in front of users who didn’t ask for it. With SEO, users come to you on their own, with a question already in mind. That traffic carries intent by nature, and it’s more precise.
Third, every conversion can be cleanly attributed. SEO traffic comes from specific, individual keywords — you can see exactly which word brought a visitor, and which visitor paid. This becomes essential later when we start pruning.
Fourth, AI has made data analysis easy. It used to take real effort just to see your data clearly; now the barrier is much lower.
And there’s an underrated bonus: the process of doing SEO itself forces you to understand your brand and your product.
03 - How we actually cold-started SEO
After going global, the biggest question was how to cold-start.
Influencers, paid ads, Reddit, communities, SEO — we considered everything we could think of, and settled on SEO.
But the hard part of SEO isn’t the method. It’s the pacing. The feedback loop is long.
We launched in June 2025. Stable Google traffic didn’t show up until August. The first real organic visitor appeared about two months after launch; the first paying conversion came four to five months in. Only at year-end, riding Christmas and New Year content, did we get our first real breakout — and that first breakout already covered our costs.
Those middle months are the hardest. Feedback is sparse, and self-doubt creeps in easily.
So our cold start was essentially three steps:
Step one: build up site authority (about one to two months).
First, understand one thing: a brand-new site has zero credibility in a search engine’s eyes. Why would it hand a searcher’s query over to you? Largely, it depends on how many other sites — especially old, established ones — are willing to link to you.
In SEO this is called a backlink. Think of it as a “vote of trust” another site casts for you. The older and more authoritative the voting site, the more that vote is worth, and search authority flows along the link.
So for the first two months after launch, we did unglamorous things: bought backlinks from high-authority aged domains, found free but high-value directory listings, set up social accounts and linked back, and did link exchanges with other sites. None of it is sexy, but every new site has to go through it. We were essentially collecting trust votes.
And the process of building authority is itself a way of spreading the brand outward.
Step two: expand keywords — but prune first.
This is the most valuable and most overlooked part of SEO for beginners.
A lot of people think SEO means finding a pile of high-search-volume keywords and writing furiously. But behind every keyword is intent. Search terms that all look design-related can be very different:
- Someone searching “poster design inspiration” wants references — that’s informational;
- Someone searching “AI poster generator comparison” is choosing a tool — that’s commercial;
- Someone searching “how to make a coffee shop hiring poster” actually needs to make one — that’s transactional, and that’s exactly what we want.
We don’t go straight after big terms like “AI design.” That’s a battlefield where big companies win with money and authority; a small team there is cannon fodder. We look for longer, more specific terms — a particular printed material, a particular design style, a particular use case.
These are called long-tail keywords. Individually the volume is small, but they have three advantages: very clear intent, very low competition, very high conversion. And added up, long-tail traffic is often more stable than a few big keywords.
But the more important move is pruning. We ruthlessly cut keywords that bring high traffic but low conversion — the “just looking,” “just playing” terms. They bring lots of visitors who never pay, and they just burn your server and token costs while breeding a crowd of freebie-seekers.
The logic here comes back over and over: for an AI product, not all traffic is worth having.
Step three: use trending content to get early positive feedback.
Because SEO feedback is so slow, you have to weave in trending content along the way to prove the direction isn’t wrong.
There’s a useful pattern here: a lot of design demand follows the holidays, and it’s predictable. Before Christmas, New Year, Valentine’s Day, Mother’s Day — search volume for related design content has a reliable surge.
We’d build content around these dates in advance. On this kind of content, we saw small breakouts several times.
These small breakouts may not bring many direct conversions, but they let you confirm, during the long wait, that you haven’t gone the wrong way — for an early team, that psychological reassurance matters more than the numbers.
I don’t want to over-mystify SEO. Getting traffic from SEO isn’t actually hard. The hard part is getting traffic that’s valuable to your product.
04 - What AI products fear most is vanity traffic
This section, I want to seriously make a counterintuitive point: for AI products, more traffic is not always better.
When I worked on traditional products, my sense of traffic was different. Traditional products are mostly a customer-acquisition-cost problem — once users are in, the marginal cost is roughly under control. More people, just add more servers.
AI products are different. Every user who comes in can bring token cost, image-generation cost, retry cost, failure cost, and even a pile of feedback that drags the product off course. The more users, the more these costs scale linearly.
So early on, an AI product can’t just chase more traffic.
We learned this the hard way.
For a while, because of some borderline content, we attracted a huge wave of low-quality users. It was lively, sure — but zero conversion, and it pushed our server costs way up.
This is where you see the value of SEO attribution: we could see clearly from the data which keyword, which article, brought this crowd in, and what their conversion rate was. The answer was clear: this traffic was a liability.
Then we used an algorithm to clean out that low-quality traffic. The result was interesting — traffic dropped, but paid conversion was completely unaffected.
That’s when a sentence finally clicked for me: for an AI product, a non-converting user isn’t an asset — it’s a liability.
So we stopped chasing “making the numbers bigger” and started caring more about whether the traffic was precise. Users who convert, who give feedback, who help the product get better — that’s good traffic. Traffic that only makes the dashboard look bigger often just burns money faster for an AI product.
This is also why I believe AI-native growth can’t just copy the traditional internet playbook. An AI product has to run two sets of books at once: one for acquisition, one for delivery. Every time traffic comes in, you have to ask — can the revenue it brings cover the cost it consumes? If you only run the first set of books, you can calculate yourself to death.
Grow volume on one hand, control it on the other. It sounds contradictory, but for an AI product it’s necessary.
05 - AI executes, humans judge
The Mew Design team has always been very small.
At the start it was just my co-founder and me — she owns product tone, design direction, and aesthetic judgment; I own the AI agent, technical architecture, and product development. We shipped the first demo in three months, then brought in a third partner for growth. Today the team is five people, all part-time, with a lot of the work handed to AI.
AI has changed how I build more than anything.
Early on, Cursor and I were about 50/50 — a lot of detail still needed my own hands. After Cici came out, I basically stopped writing much code by hand.
That sounds great, but it’s easy to misread. Not writing code doesn’t mean having nothing to do. Quite the opposite — I spend more time on these things: product architecture, technical boundaries, agent rules, how to make tradeoffs in each iteration, what to refactor, what counts as a good result.
AI is great at execution. It’s not great at taking responsibility. Product direction, system boundaries, aesthetic standards, long-term tradeoffs — those still have to be set by a person.
We felt this deeply when hiring. I originally wanted to bring on a full-time technical co-founder, but after talking to dozens of candidates, I gave up. Not because the people weren’t good — but because their old technical mental model didn’t fit a team mode where “AI does most of the work.”
System architecture follows the same logic. We don’t over-build ahead of time. We only refactor a module after a model’s capability genuinely takes a leap. Pace has to follow model iteration; running too far ahead is just waste.
Growth is the same. A lot of people think: since AI is this strong now, can SEO content be fully automated? Let AI write post after post, and just flood the volume?
In practice, for us, no.
There’s actually a basic SEO fact here that many don’t know: Google has spent years cracking down on low-quality, mass-produced, search-born content, and it’s gotten better and better at telling whether a page is actually useful to a user. AI can produce content for you, but it can’t make that content useful on your behalf.
AI is genuinely useful in growth — especially for data analysis, insight synthesis, competitive research, keyword work. But the real content creativity, the marketing expression, the judgment of user need — that still has to be owned by a person.
Google has no shortage of auto-generated pages, and neither do users. What’s missing is content that solves problems. It sounds plain, but after you’ve taken the hits, you believe it more.
06 - Finally
After the talk in Liangzhu, I came away with one strong feeling:
In the AI era, small teams really do have a chance.
Because a lot of what used to require resources, headcount, and accumulated experience is being massively cheapened by AI. Research, translation, development, asset understanding, data analysis, execution speed — all of it is getting cheaper.
But don’t misread this. AI just gets you to the battlefield faster. It doesn’t win the war for you.
What the product does, who the user is, why they’d pay, which traffic is worth having, which feedback to listen to, which aesthetic to hold — you still have to figure these out yourself.
The thesis of that whole event was really this one line: going global isn’t translating a Chinese product into English. It’s re-understanding demand, re-organizing distribution, re-building trust.
One last thing — a direction we’re watching. There’s been an obvious shift over the past couple of years: more and more people are getting into the habit of asking an AI directly instead of searching Google. This has produced something new, called GEO (Generative Engine Optimization) — studying how to get AI to mention you and recommend you when it answers a user’s question. The traffic is smaller than SEO for now, but the conversion rate is often higher. For a team that grew up on SEO, it’s the next station we have to keep our eyes on.
If I had to compress this talk into one sentence, it’d be:
AI gives small teams speed; user demand decides the direction.
The rest is just: don’t be afraid of being slow, and don’t get pulled off course by the hype. Build the product well, keep the costs clear, and do the things that compound, one by one.
That’s everything we have to share for now.