Wix’s 2025 LLM Strategy: 40% Content Cut

Wix is a huge name in website building, and their work with LLM automation offers a real playbook for other companies. They’re not just talking about it. By 2025 they’d cut the time it takes to write marketing copy for certain campaigns by a full 40%, and they credit their LLM integration for it. Their story, especially around content generation, has some practical lessons for anyone trying to get real work done with these models.

1. Define Clear Automation Targets and Content Types

Wix didn’t start by asking “what can the AI do?”. They started by looking at their own workflows to find the boring, repetitive parts. Not all content is the same, right? Some of it needs a person’s touch, but a lot of it is just formulaic. They zeroed in on product descriptions, basic marketing emails, drafts for social media posts, and articles for their internal knowledge base. The trick was targeting content that always followed a pattern and just needed to pull from existing data. For example, writing a description for a new website template meant grabbing a list of features from a database and spinning it into marketing-speak, a perfect job for a machine.

Pro Tip: Start small. Seriously. If you try to automate your most creative or high-stakes content first, you’re just setting yourself up for failure and a whole lot of frustration. Nail the high-volume, low-creativity tasks first to get a win.

Common Mistake: Boiling the ocean by trying to automate every piece of content at once. It spreads your team too thin and makes it impossible to figure out what’s working and what’s breaking in your process.

2. Choose the Right LLM Architecture and Model

Forget just grabbing the biggest, most-hyped LLM off the shelf. Wix ran a bake-off between commercial APIs and open-source models to see what fit. When they needed deep customization and had to teach the AI their own internal data, they put in the work to fine-tune models like Llama 3 8B. That was the only way to get the model to adopt Wix’s specific brand voice, use the right internal terms, and understand product details. For other, more generic writing tasks, they looked at commercial APIs that delivered good results without all the engineering overhead. The decision was a constant trade-off between paying API fees versus paying for your own engineers’ time to fine-tune and host a model.

Screenshot description: A simplified diagram showing Wix’s LLM architecture. On the left, “Data Sources” (Product Database, Marketing Guidelines) feeding into “Data Preprocessing.” In the center, a “Fine-tuning Module” for Llama 3 8B. On the right, “API Gateway” connecting to “Content Generation Services” (Product Descriptions, Email Drafts).

3. Curate and Prepare High-Quality Training Data

This is the part everyone underestimates: the quality of your LLM’s output is a direct reflection of the data you feed it. Garbage in, garbage out. Wix spent a ton of time cleaning up thousands of their existing product descriptions, marketing emails, and blog posts to create a high-quality dataset that showed exactly the tone and accuracy they wanted. The data was cleaned, de-duped, and annotated. They even built a dataset of “negative examples” (content that was off-brand or just plain wrong) to explicitly teach the model what not to do. This data prep work is a grind, but skipping it guarantees you’ll get useless results.

Pro Tip: Don’t just dump raw data into the fine-tuning process. You have to annotate it. Tagging your examples with labels like “good,” “bad,” or “needs review” gives the model much clearer signals to learn from.

Common Mistake: Grabbing some generic, uncurated dataset off the internet. You’ll end up with generic, off-brand text that’s probably full of factual errors and needs a human to rewrite it from scratch anyway.

4. Develop Strong API Integrations and Workflows

The AI model is the shiny part, but the real work is in the plumbing. You have to integrate it into how people actually do their jobs. Wix had to build custom APIs to hook their LLMs into their internal content system (Contentful), their marketing tools, and even their project management software. That meant building a system with clear inputs (like a product ID and target audience) and predictable outputs (like a JSON object or simple text). They also had to implement version control for all their prompts and model settings, so they could actually trace where a piece of content came from and reproduce the results. This is complex dev work that requires your AI nerds, software engineers, and content people to all be in the same room.

Screenshot description: A screenshot of an internal Wix dashboard. On the left, a “Content Request Form” with fields for “Product Name,” “Key Features,” and “Target Audience.” On the right, a “Generated Content Preview” pane showing a draft product description, with options to “Approve,” “Edit,” or “Regenerate.”

5. Implement Iterative Prompt Engineering

You don’t just “write a prompt” and call it a day. Prompting is a constant process of refinement, a craft you get better at over time. The Wix team was always tweaking the instructions they gave the LLM to steer its output. They learned that giving very specific examples, defining the tone of voice (e.g., “be professional but approachable”), and spelling out constraints (like “don’t exceed 150 words” or “make sure you include a call to action”) made a huge difference in the quality of the first draft. They also used techniques like chain-of-thought prompting, which involves telling the LLM to “think out loud” before writing the final text, resulting in more logical content. This job is never finished. Prompts have to change as the model gets updated and as the business needs evolve.

Pro Tip: Build a prompt library. Documenting the prompts that work, and the output they create, is a lifesaver for standardizing content and getting new people up to speed.

Common Mistake: Writing lazy, vague prompts. A prompt like “Write marketing copy” will give you useless, generic mush. A prompt like “Write a concise, benefit-driven product description for a new e-commerce template, highlighting its mobile responsiveness and integrated payment gateways, targeting small business owners, under 100 words” will actually give you something you can work with.

6. Establish Human Review and Feedback Loops

Automation doesn’t mean you can fire your editors. In fact, they become more important. Wix built a mandatory, multi-stage human review step into their workflow where content specialists would edit and fact-check every AI-generated draft. But here’s the important part: the system was built so reviewers could explain *why* they made a change. That feedback didn’t just go into a black hole, it was fed back into the system to either fine-tune the model or improve the prompts for the next time. This continuous feedback loop is the real engine for improving the AI’s performance over the long haul. Without it, your model just keeps making the same mistakes forever.

Screenshot description: A snippet of a Wix internal review interface. A generated paragraph is highlighted, with a comment bubble next to it stating, “Tone too formal for our brand. Suggest ‘friendly’ instead of ‘professional’.” Below, a button labeled “Submit Feedback to LLM Training.”

7. Monitor Performance Metrics and Iterate

You can’t manage what you don’t measure. Wix was obsessive about tracking key performance indicators (KPIs) to see if this whole effort was actually working. They watched:

• Time to Content Creation: The raw reduction in hours from the initial request to the final, approved content.
• Editing Time: How long did a human have to spend fixing the AI’s draft? If this number went down, the model was getting better.
• Content Quality Scores: An internal score based on how well the content followed brand rules, its factual accuracy, and how engaging it was.
• Cost Savings: A direct comparison of the cost of a person doing the work versus the cost of the automated system (including compute and review time).

This data dictated what they did next. If editing time on product descriptions was still too high, for instance, that was a clear signal to go back and either retrain the model or rework the prompts. A data-first approach is the only way to get lasting gains.

The big lesson from Wix is that LLM automation isn’t a magic wand. It takes a ton of planning, hard work, and constant tweaking. You have to treat it like a real business project, not just a cool new toy. The payoff comes from weaving these powerful models into your actual day-to-day operations with a clear eye on the metrics. For anyone going down this path, it’s worth thinking about how LLMs redefine workforce automation and the new skills it requires. And as you handle more and more data, paying attention to LLM data privacy becomes non-negotiable. Finally, every business needs to get smart on the LLM ethics and compliance risks to deploy this stuff responsibly.