I Built a Week of Instagram Posts for a Gelato Shop With AI (It Took 40 Minutes)
A real case study: 40 minutes, AI image generation, one gelato shop, a week of on-brand Instagram posts. What worked, what failed, and what to copy.
A small gelato shop in San Diego was posting once every 10 days. Stock-looking iPhone photos under bad fluorescent light, captions typed in 5 seconds, three emojis, done. I sat down with the owner, opened my laptop, and 40 minutes later she had a full week of Instagram posts ready to schedule. This is what actually happened.
The before#
Her Instagram was a graveyard. One post per week on a good week. The photos were her phone camera pointed at the display case with glare bouncing off the glass. Captions were 12 words and always ended with the same three emojis. She told me she knew it wasn't working, but writing one post took her 20 minutes because she never knew what to say, and by the time she figured it out the dinner rush was starting.
This is the real problem with social media for restaurants. It's not that owners don't care. They run a kitchen, a front of house, payroll, ordering, and a family. A post that takes 20 minutes is a post that doesn't get made.
The 40-minute session#
Here's exactly what we did.
- Uploaded 12 product photos. Her 8 best-selling flavors plus 4 storefront shots she already had on her camera roll. No staging. No studio.
- Added the Instagram URL. The AI pulled her brand voice, her colors, and the vibe of her feed.
- Picked 3 content types. Product Showcase, Behind the Counter, and Lifestyle. I let the tool auto-decide the mix.
- Hit bulk generate for 7 posts. Went and made coffee.
- Came back, reviewed 7 drafts. Kept 5 as-is. Regenerated 2 that felt too generic.
- Scheduled the week. All 7 posts queued before we finished our second espresso.
Total time including the small talk and the regenerates: 40 minutes.
What worked#
The biggest win was the reference photos. When I uploaded her real pistachio, stracciatella, and mango sorbet, the AI copied the actual flavor colors. Not a generic green scoop, her specific oxidized-pistachio green. That's the detail that stops the scroll.
The Surprise Me feature broke the predictable pattern too. Left to its own logic, any AI tool defaults to the same three shots: hero on white, flat lay, happy person holding cup. Surprise Me pulled ideas I wouldn't have picked, like a shot of the display case at opening time with condensation on the glass. That post ended up being the top performer of the week.
Captions landed in her voice, not AI-speak. No "indulge in our decadent" nonsense. They read like a friend texting you about gelato. The hashtag mix was smart: local tags (San Diego neighborhoods, not just #sandiego), flavor-specific tags (#pistacchio, #stracciatella), and a few small-community tags under 50k posts. No dead top-100 hashtags.
The same logic applies to any food business, and it's the same pattern I wrote about in my post on building AI business systems: find the 20-minute task that repeats weekly and collapse it to 5.
What didn't work#
Not everything landed. Three things failed hard.
Logos on the actual gelato. I tried it. It always looks stamped on, like a sticker someone slapped on the scoop. The AI can put a logo on a cup, a bag, or an apron beautifully, but not on food. Food should look like food.
Generated customer faces. Even on the best model, synthetic humans holding gelato fall into uncanny valley. You can feel it's fake even if you can't say why. We skipped every post that leaned on a fake face and instead used shots that showed hands, cups, and the shop interior. Those felt real.
The word "authentic." Every time the tool added it to a caption, the post got worse. Real places don't need to tell you they're authentic. We cut it every time and let the actual photos do the work.
The actual engagement change#
Week one after she posted the AI content, her impressions went 2.3x compared to her previous 30-day average. Profile visits 4x. Saves jumped from 2 per post to 11. She got 3 DMs asking about catering, which she had never gotten before.
One week doesn't prove a long-term outcome. It could regress, it could keep climbing, we don't know yet. But the before-and-after on week one is real, and it cost her 40 minutes instead of 3.5 hours.
Frequently asked questions#
Can AI generate good Instagram posts for restaurants? Yes, if you feed it your real product photos as references. Without real references, you get generic food stock. With 10-12 actual shots of your dishes, the output looks like your place, not a catalog. The quality delta between "no references" and "good references" is the whole game.
What's the best way to write restaurant social media captions? Write like you talk to a regular at the counter. Short, specific, one detail, one call to action. Skip the adjectives. "New mango sorbet, made with Alphonso mangoes from the farmers market, only until Sunday" beats "Indulge in our luscious new tropical creation" every single time.
How many times a week should a restaurant post on Instagram? 3 to 5 posts a week is the sweet spot. Posting daily burns content fast and most owners can't sustain it. Posting weekly is too quiet to grow. Three to five keeps you in the algorithm without breaking your operation. Batch them on Monday, schedule the whole week, forget it until next Monday.
The real lesson from 40 minutes with that gelato shop: the bottleneck on restaurant Instagram was never creativity or caring. It was time. AI just bought her back 3 hours a week, and her feed finally looks like the place feels.