Current state
Designing a Restaurant menu in the Wix editor
These design panels are part of the holistic editor experience across all of the Wix apps.
Users describe the look they want, the AI generates five distinct styled options, and they pick one. The manual controls stay exactly where they were, this was an addition, not a replacement. For some the feature would act as a great starting point to then tweak the design and for some they would use it as is.
We deliberately kept V1 lean, a simple panel with minimal features so we could ship fast, get real usage data, and only invest further if the feature proved its value.
Rather than asking the AI to design from scratch every time, we gave it a head start: a database of 42 high-quality, pre-made menu templates to use as a foundation. The model used these as a starting point and would design options based on them. I designed these templates in a design blitz alongside a small group of designers, then tagged each one with keywords across four categories; Atmosphere, Fonts, Restaurant Type, and Color Scheme.
The model was designed to use semantic search. This means it understands the intent behind a user's words, not just the literal keywords. Even vague or non-technical prompts would reliably land on a relevant starting point.
​
Collaborating with the data science team to train the model was a challenge: giving the model too much structure led to generic results. Too much freedom and they fall apart. Finding the range where output feels both creative and professional took countless iterations.
We knew that this method of a static database isn't scalable but at the time, the models weren't mature enough to create something beautiful without the database and the race in AI was real, we accepted this tradeoff.
Takeaways
Research for V2
Competitor analysis
Across the AI generation tools we looked at, rather than just having a standard prompt box our competitors organized the creation flow into clear, distinct steps was a consistent pattern. Providing prompt examples upfront, showing users what the AI can achieve before they type anything, also set clearer expectations and reduced frustration.
​
Reading what users inputted:
Analyzing what users typed into the V1 prompt box was revealing. Many requests had nothing to do with design, users were asking for general Wix editor capabilities, unrelated design changes, and things entirely outside the scope of Style AI. While these inputs helped shape other areas of our roadmap, they made one thing clear: users weren't sure what the feature was for or what it could do.
​
Insights
The research pointed to two problems to solve: discoverability and expectation setting. Users who found the feature converted well, but too few found it, and those who did often didn't know what to ask for or what to expect. The competitor analysis gave us a clear direction: structure the input, lead with visuals, and give the option to show an example prompt.
Approach
Training the model with the use of a database
An example of a tagged menu from the database
Color scheme:
Soft/Monochrome
Atmosphere:
Elegant/Modern
Restaurant type:
French / Bakery
Font:
Handwritten / Vintage / Modern
Tradeoff
Brand alignment vs. the wow impression
The menus app sits inside a user's Wix website, which already has an established color palette and typography. When users type a prompt, they're expecting a wow effect. But if the AI ignores their existing site styles, the result can look completely out of place from the rest of their site leading to potential frustration.
​
After debating it heavily, we defaulted to unrestricted, allowing the AI to use any color and font available. Ultimately the wow effect is what makes users stay and continue designing and publishing their menu and site. We gave users the option to toggle on brand restrictions if they wanted alignment. This way we acheived maximum impact by default, control available for those who need it.
Impact
After one month, we analyzed ~2,500 Wix sites that interacted with the Menu Settings panel (where the feature is nested).
+28%
"Ready Premium"* conversion rate among 'Style AI' users
The gap between those two numbers is the clearest signal from V1: the value is real, but discoverability is the next problem to solve. Users who find the feature convert significantly better — we just need more of them to find it.
* A ready premium user is a user that published his site and purchased a premium plan
Next steps
V1 was intentionally lean — a way to validate the concept before investing further. Once the data confirmed the value, I led the design direction for V2.
Plans for V2
V1
V2
Takeaways
V2 decisions
Boosting discoverability through the entry point
Previous entry point
New entry point
In V1, getting to Style AI required clicking "Settings" first — an extra step that buried the feature. In V2, "Settings" becomes an icon button and a dedicated textual "Style AI" button takes its place. One less click, a much clearer path in to boost discoverability. We planned to roll this out in an A/B test to see that we don't harm other intents.
Guiding the user's input
The new panel targets the friction points V1 exposed: the free-text field is shorter and optional, removing the pressure to write a perfect prompt. Guided dropdowns for atmosphere and style set clearer expectations and produce more reliable results. "Try an example" shortcuts tackle blank canvas syndrome head-on, giving users a concrete starting point when they're not sure what to ask for.
​
V2 designs were completed and ready to ship, but weren't rolled out due to team capacity and competing priorities at the time. The groundwork is there for whenever the team picks it up.
Final designs
V1 (MVP)
V2
Solution
Final designs
V1
V2
Solution
All screens
V1
V2
Takeaways
​
​
2. Training a model is a design problem.
Too much structure and the results feel generic. Too much freedom and they fall apart. Finding the range where output feels both creative and professional took constant iteration — and that iteration never really ends.
​
3. Influencing a company takes longer than shipping a feature.
We set out to establish a reusable AI design pattern across Wix products — and it happened, just not on sprint timelines. Teams behind Wix Vibe, Base44, and Wix Nano all adopted a similar model.
Note that the users are free to do any manual iterations after their AI result, not all results here are one-shot
Real users
Gallery of real users that generated a menu with 'Style AI'
Overview
How we added an AI shortcut to a design process that was losing users.
Role
Lead Product Designer
Platform
Wix Editor
Year
2024
During a hackathon, our PM proposed an AI-powered shortcut for one of our biggest pain points: styling a restaurant menu. We ran a proof of concept, liked what we saw, and took it all the way to production.
The result was Style AI - a feature that lets users describe the look they want in plain language and get five fully-styled menu options in seconds, while keeping all existing manual controls in place.
Contributions
UX, Visual design, Graphic design, Model training & tagging
Problem
The Wix Restaurant Menus app gave users full control over their design but that control came at a cost. Every element had to be adjusted individually, and for users without a design background, the sheer number of decisions was overwhelming. Those that do have a design background still suffered from this as a time consuming and tedious task.
Research
Data
41% of users who started customizing their menu gave up and didn't end up publishing. Of those, 17% deleted the app within 3 days of installation. The data also revealed a high number of users who published with default design settings. Digging into this showed that many had actually started designing and then reset everything back to default. Users weren't just giving up, they were actively undoing their own work, which pointed to something deeper than confusion.
​
Recorded user sessions
Reviewing sessions revealed users scrolling through the panels and more often than not abandoning the design without making any meaningful changes to their menu.
​
User voice
Support tickets and user feedback forms consistently raised questions about how to style individual elements. The capability existed, but users were struggling to find it. Our support team was spending time writing how-to guides instead of providing actual support.
​
Main insights
The pattern was consistent across all three sources: we were losing valuable users. They had high intent- they'd bothered to search for and download the app, and we weren't delivering on our end. Users weren't overwhelmed by one hard decision; they were worn down by the accumulation of small ones. They didn't lack taste. They lacked the time and confidence to execute it.
MVP- V1
'Style AI' a basic AI shortcut nested in the existing design panel.
Full flow in context- MVP (V1)