Selling shopkeepers a picture of themselves: a budget paid-acquisition program for Fairmart

Published: 2025-03-27 · #Paid Acquisition

TL;DR: A three-month paid-acquisition program for Fairmart, a seed-stage Singapore commerce platform for SMB retailers, that used AI-generated anime creative to turn cold Meta ads into SGD 46k of pipeline in a single month on SGD 1.8k of spend, by showing shopkeepers an image of themselves instead of a list of features.

Problem

Fairmart is a commerce platform for modern SMB retailers: marketplace eCommerce, in-store POS and payments, and supplier ordering in one system. The customers are minimart owners, kopitiam operators, and similar small retailers running physical stores in Singapore.

Until this program, every customer came through organic channels or the sales team. That works at small scale, but it does not tell you whether you have a repeatable growth channel, which is the thing a seed-stage company needs to know before it spends real money finding out. Paid acquisition was the test.

The constraints were tight, and they shaped every decision that followed:

  • Budget was small. The whole program ran on a few thousand SGD, with roughly SGD 1.8k in the strongest month. There was no room for waste and no room for an agency.
  • No budget for a creative shoot. Photographing real shopkeepers who matched the ICP was out of reach.
  • The team was thin. I built the first version and ran it, then trained one internal team member to help iterate creative under my review. The sales lead assessed lead quality.
  • The market is small. Singapore is a single small country, and the target segments inside it are smaller still. Thin audiences push cost per lead up, and that ceiling sat over the whole program.

Approach

The buyer is a shop owner who does not want software. They want to stop losing sales when the store is shut, and to grow without taking on more work. Fairmart is a stack of features, but features are not what the owner is buying.

So the program was built around one bet: sell the customer an image of themselves and a desire, not a feature list. Concretely that meant:

  • Scoped in: Meta lead generation, broad ICP targeting, retargeting on people who engaged, weekly creative rotation, and creative that named the store type directly.
  • Scoped out: feature-led creative, photorealistic creative, interest-based targeting, agency production, and any second channel.

Everything below follows from that one decision about what the ad should say and who it should picture.

Key decisions and tradeoffs

AI-generated anime creative over realistic or feature creative. This was the central bet, and it started as a hunch around the time of the Ghibli image meme. With no budget to shoot real people, I generated creative with Nano Banana, Google’s image model. The first version showed the product and the setting with no person in it. Adding a person, rendered in an anime style the ICP could see themselves in, was the change that moved conversion. Realistic AI creative never converted. Feature-led creative never converted. The relatable anime version of the customer beat product-specific creative by roughly ten times.

The second-order effect mattered more than the first. Because generation was nearly free, creative stopped being the expensive part of the funnel. That unlocked the next decision.

Store-type-specific creative over one generic ICP ad. Once creative was cheap, I could make a separate ad for each kind of store and name it in the image: a banner reading “Minimart”, a kopitiam scene. Owners self-identify. A minimart owner who sees a minimart converts more than one who sees a generic shop. This worked, and the data shows it clearly: the Minimart and Kopitiam creatives were the top performers, with the Kopitiam variants landing under SGD 5.50 per lead.

The sharper finding came from where it stopped working. Segment-specific creative only pays off when the segment is large enough to fill a campaign. Singapore has more than five thousand kopitiams and around two thousand minimarts, so those worked. Pet shops and bike stores number in the hundreds, even though the audience interested in pets is far larger than the number of pet shops. For those niches the creative was fine, but the in-market buyer pool was too thin to find efficiently. The cheap-creative capability turned into a market-sizing probe: it told me, segment by segment, where the real addressable demand was and where it ran out.

Broad ICP targeting with retargeting over interest targeting. I tried interest targeting first, and the cost per lead was prohibitive: niche interests in a small market produce thin, expensive audiences. Broad targeting, with retargeting on the people who engaged, worked far better. The tradeoff is that you give up manual control over who sees the ad. But that control moves into the creative: naming the store type in the image is the targeting mechanism. The picture filters the audience, so the algorithm does not have to.

Architecture

The system is simple, and that is the point. Generate creative with Nano Banana, rotate five to eight new concepts a week through a broad prospecting campaign, retarget engaged users, capture leads through a Meta form, and hand them to the sales lead to qualify.

The leverage point is the cost of creative. When generation is near zero, the binding constraint stops being “can we afford to test this” and becomes “can the market absorb the test.” The image does double duty as both the message and the audience filter.

The failure modes are structural, not operational. In a small market, audiences saturate fast, so creative fatigues and cost per lead climbs as a niche is exhausted. Segment-specific creative caps out at the size of the segment. None of this is fixed by running the campaign better. It is fixed by a bigger market.

A note on rigor: on a budget this small, most concepts were paused after well under SGD 100 of spend. Those were directional calls, not statistically clean ones. The strong winners earned more spend and more confidence; the rest were cut before they could prove much either way.

Results

Leading with the commercial number: in its strongest month, the program generated about SGD 46k in pipeline opportunities from SGD 1.8k in spend, roughly 38 percent of that month’s total pipeline.

Underneath that, the operational picture across the program: about SGD 3.6k spent, 388 leads, and a blended cost per lead near SGD 9. The blended figure understates the winners. The top creative produced 123 leads at SGD 8 each, and the best Kopitiam variant came in under SGD 4. Of roughly eighteen concepts tested, a small handful (the store-type ads and the “sell while you sleep” hook) carried almost all the volume. Over the three months, those leads converted into around SGD 18k of closed deals.

The one-line version: relatable AI creative made cold acquisition profitable on a seed-stage budget, and it did so by picturing the customer rather than the product.

What I’d do differently

Start earlier, and run it somewhere bigger. Singapore is a small market, and targeting niche retailers inside it makes the audiences thinner still, which kept cost per lead higher than the creative deserved. The same engine pointed at a larger market like Australia or the US would have more room to run before hitting the ceiling that capped this one. The constraint on the result was never the creative or the method. It was the size of the pond.

Tech stack: Meta Ads (broad targeting, Advantage+, lead forms) · Nano Banana / Gemini image generation · retargeting · weekly creative testing