ClickAi Meets GPT: Upgrading an AI Product with LLMs

ClickAi has been an AI-powered website builder since before that phrase became a buzzword. When I first built it, "AI-powered" meant rule-based content generation, template matching, and some clever conditional logic. It worked. Users could describe their business and get a functional website. But there was always a ceiling on how creative and personalized the output could be.

With LLMs, that ceiling has been shattered. Over the past two months, I've rebuilt ClickAi's content generation pipeline around GPT, and the difference isn't incremental. It is a step change. Here is the full story of how I did it and what I learned.

What ClickAi Was Before

To understand the upgrade, you need to understand what ClickAi was. The original system worked like this:

1. User selects an industry (restaurant, law firm, photography studio, etc.)
2. System loads an industry-specific template with placeholder sections
3. Rule-based engine fills placeholders with content pulled from a database of industry-specific copy
4. User customizes the result

This pipeline was fast, reliable, and predictable. Every restaurant website got variations of the same copy. Every law firm website followed the same structure. The output was professional but generic. Users could tell they were getting a template, even if the template was well-designed.

The biggest complaint from users was always the same: "The content doesn't sound like me." And they were right. A pizza shop in Brooklyn and a fine dining restaurant in London have very different voices, but ClickAi treated them identically.

The LLM Upgrade Plan

I identified three areas where LLMs could make the biggest impact:

1. Initial Content Generation. Instead of pulling from a database of pre-written copy, use the LLM to generate unique content based on the user's specific business description.

2. Tone Matching. Let users describe their brand voice (professional, casual, playful, authoritative) and have the LLM adapt all generated content accordingly.

3. Iterative Refinement. Allow users to give natural language feedback ("make this section shorter," "add more emphasis on our eco-friendly practices") and have the LLM revise the content.

I decided to tackle these in order, shipping each one before starting the next. This was a deliberate choice. I've learned from experience that trying to ship everything at once means shipping nothing well.

Building the New Pipeline

The new content generation pipeline looks like this:

1. User describes their business in plain text (even one sentence works)
2. A classification prompt extracts structured data: industry, location, key services, target audience, tone
3. A generation prompt creates section-by-section website content using the extracted data
4. Content is injected into the design template
5. User reviews and can refine with natural language instructions

The key architectural decision was step 2 -- the classification step. I could have sent the user's raw description directly to the content generation prompt, but separating classification from generation gives me several advantages. I can validate and correct the classification before generation. I can cache at the classification level. And I can use different models for each step (GPT-3.5 for classification, GPT-4 for generation when quality matters).

The Prompt That Changed Everything

The content generation prompt went through dozens of iterations. The first version was embarrassingly simple:

"Generate website content for a {industry} business called {name}. Include sections for hero, about, services, and contact."

The output was generic and bloated. It read like every other AI-generated website. Not terrible, but not good enough to justify the API cost over my old template system.

The prompt that finally worked was much more specific. It included:

• The exact JSON structure I needed for each section
• Word count limits per section (hero headline: max 8 words, hero subtitle: max 20 words)
• Three examples of excellent website copy for similar businesses
• Explicit instructions about what to avoid (no cliches, no "Welcome to our website," no generic superlatives)
• The user's specified tone and audience

The output from this prompt was genuinely good. Not perfect, but good enough that users were surprised. Multiple users told me the AI-generated copy was better than what they had written themselves. That was the moment I knew the upgrade was working.

Handling the Edge Cases

The happy path worked beautifully. The edge cases nearly killed me.

Multilingual Users. ClickAi has users who describe their businesses in languages other than English. The original system handled this because the template content was available in multiple languages. With LLMs, I needed to detect the input language and generate content in that same language. GPT handles this well for major languages but degrades for less common ones. I ended up adding explicit language detection and instruction to the prompt.

Inappropriate Requests. Some users try to generate content for businesses that violate OpenAI's terms of service. I added a classification step that checks the business description before generation and routes inappropriate requests to a rejection flow.

Very Long Business Descriptions. Some users write essays about their business. This blows up the token count and slows down generation. I added a summarization step that condenses long descriptions before feeding them to the content generation prompt.

Hallucinated Details. The LLM sometimes invents details about the business that the user never mentioned. If someone says they run a bakery, the LLM might add that they have been "serving the community for over 20 years" -- a detail pulled from nowhere. I added post-processing that flags any specific claims not grounded in the user's input.

The Results

I've been tracking metrics closely since the LLM upgrade launched.

Content Quality. User satisfaction with generated content increased significantly. The most telling metric: the percentage of users who edit the generated content before publishing dropped. They are happier with the first draft.

Time to Publish. The average time from signup to published website decreased. The blank page problem is essentially solved. Users get a complete, personalized first draft instantly.

Retention. Too early to draw conclusions, but early cohort data looks promising. Users who experience the LLM-powered generation seem to be more engaged.

Cost. API costs are real but manageable. The average content generation costs a few cents in API calls. For a product with a monthly subscription, the unit economics work. Caching helps significantly for popular industries.

What I Would Do Differently

If I were starting this integration from scratch, I'd invest more upfront in evaluation infrastructure. I built the feature first and the evaluation system after, which meant I was flying blind during the critical early iterations. Having a proper evaluation framework from day one would have saved me time and improved quality faster.

I'd also start with GPT-3.5 everywhere and only upgrade to GPT-4 where the quality difference is measurable and meaningful. I started by using GPT-4 for everything, which was expensive and slow. Most features don't need the most powerful model.

The Competitive Landscape

Every website builder is adding AI features right now. Wix, Squarespace, WordPress -- they're all integrating LLMs. This doesn't worry me as much as you might think. ClickAi's advantage was never just the AI. It is the specific workflow, the industry specialization, the opinionated design system. The LLM makes all of those better, but it isn't the whole product.

The real competitive threat isn't other website builders adding AI. It is the possibility that LLMs become so good that they can generate entire websites from a single prompt, making website builders unnecessary. I think about this a lot. But for now, the gap between "generate a website" and "generate a website that's actually good enough to publish" is still wide enough to build a business in.

ClickAi with LLMs is the best version of the product I've ever shipped. And I am just getting started.