Why AI Content Usually Sucks—And How We're Fixing It

Most of the visual content that AI generates sucks. As good as Gemini 2.5 and ChatGPT 4o have become, making professional-looking visual designs with AI is still a mostly unsolved problem at scale.

If you look at the market today, there are very few products succeeding in this space with any kind of retentive customer base. On the visual side, Canva and PhotoRoom stand out, alongside Blaze. For video, you've got InVideo and a few others. But the list falls off pretty quickly after that.

Today, I want to show you how we're tackling this problem (and opportunity) at Blaze, where creating high-quality AI-generated visual content has been one of our greatest technical priorities for literally years.

I personally spend 2-3 hours every day focusing on what we call "content output quality." We have our best engineers, designers, and product minds working on it.

We've made an insane amount of progress, especially compared to other tools, but if I'm being honest, I don't think we've completely cracked it. So here's an inside look at our progress, our challenges, and the future opportunities in the space.

Why Quality Visual Content Is So Hard

Every day, Blaze creates hundreds of thousands of pieces of visual content for our customers. Every month, that number goes into the millions. For our users, "high quality" means two key things:

1. Personalization: The content looks, sounds, and feels like YOU—your brand, style, and voice, along with specifics about your products, services, customers, and market

2. Professional execution: Fonts are properly sized, images are relevant, colors cohere, spacing is consistent—it looks professional, not amateur

These factors directly correlate with our retention and churn. When content feels authentically yours, you keep using Blaze. When it doesn't, you don't.

How We Personalize

When we first launched Blaze as a writer product, our "Brand Voice" feature set us apart. Just from your website or a paragraph of your writing, we could analyze how you write across seven different attributes: syntax, style, tone, intended audience, purpose, and more.

We made these attributes modifiable so you could tweak our analysis to make it even more accurate. This became the first thing users did when onboarding into the product, and the fact that it felt like magic is why Blaze took off initially.

Generating text was actually the easy part. Copy is less dimensional than visual content—it exists in sentences and paragraphs, with relatively straightforward formatting rules.

Visual content is much more complex. You need to nail not just the tone of the copy, but also the feel of the images, videos, and overall style. When we expanded into visual content, we created "Brand Styles"—analyzing your colors, fonts, logos, and all the images or videos from your website and social accounts.

We've since added RAG (Retrieval-Augmented Generation) into what we now call "BrandKit." When you link your website or social accounts to Blaze, we continually recrawl them to see how your content evolves over time. As you add new products or your tone or style shift, Blaze picks them up automatically.

Even better, Blaze remembers all the content you create with us and incorporates any edits you make into future generations. This wider context and deepening personalization is key to quality.

When Personalization Isn't Enough

But here's the problem: personalization works great if you already have a robust brand, media library, or social media presence. But about a third of our customers are doing nothing on social media when they come to us. Another 50% are doing some posting, but don't like what they're producing.

What happens when we try to serve someone with no established brand? Or worse, when they have some content but it's objectively bad? Personalization based on those inputs just replicates something they don't like to begin with.

This is where output quality becomes really challenging. We need to go beyond understanding and replicating how users currently look—we need to make them better.

That's the hard part with AI, because large language models take existing data to predict what something should look like. They're not naturally creative and can't have a completely original, forward-looking point of view.

Building AI With Taste

So the question becomes: how do we build AI that has taste? It's not purely a technical problem—it's also a design problem, a product problem, and a problem of discretion.

Our first approach was to use templates to direct quality. We used best practices from design to create templates for hundreds of different industries and segments to serve as the basis for personalization.

We have templates for dental offices, clothing retailers, nonprofits, farmers market sellers, restaurants, fitness trainers, financial advisors, podcast hosts, and countless others. These give us a polished, appropriate foundation when we need to start from scratch or create a clean break with old content.

But creating high-quality content for a dental office looks very different than for a Shopify store selling necklaces to goth teenagers—which is also different from a Shopify store selling candles to moms in the Midwest. Even within segments, what looks "good" varies dramatically.

Traditional design tools like Canva or Adobe Express have accumulated huge, professionally designed template libraries that can handle almost any purpose, no matter how niche. But they still force you to do much of the design work yourself.

We started by building a template gallery and using AI to modify templates with original content based on what you want in the design, what's in your BrandKit, and what key media you want to feature.

This approach created several challenges:

1. Template volume: You need a massive library to cover all use cases—we currently produce about 500 templates a month, and it still feels insufficient.

2. User flow complexity: Do users want to start with a photo, a prompt, or a template? All three approaches have pros and cons when combined with AI, and building a product that handles all three is really complex for our team and users.

3. Edge cases upon edge cases: The more complex the template, the more things can go wrong. Multiple images, text boxes, overlays, videos—each element creates new opportunities for mistakes, and therefore add complexity to our codebase and product.

Combining Generation, Templates, and Editability

We've come up with several creative solutions to these problems:

1. Multiple options: Instead of producing just one design, we now generate eight options, increasing the chance of hitting something you'll love.

2. One-click adjustments: You can now quickly test different colors or fonts on the same template without ever entering our full designer.

3. Smarter AI checks: We now have the AI evaluate its own work before showing it to you: "Does this fit the user's brand? Are the text boxes properly aligned? Is the logo displayed correctly?"

4. Advanced image generation: With the latest models from OpenAI and Google, we can now produce photorealistic visuals that finally spell words correctly and handle complex compositions.

Most excitingly, we're building a hybrid approach that combines these new image models with our user-friendly visual studio. Unlike using ChatGPT directly, where you get a "take it or leave it" image, Blaze separates all the elements—images, text boxes, shapes, logos—into modifiable components that you can easily edit.

Our goal is to get you a 90% high-quality first draft, while recognizing that creative excellence always requires human touch. That's why we've invested so deeply in our Visual Designer and Document Editor platforms—so you have control over the final 10% that makes the content truly yours.

Where We're Heading

In the next six months, you'll be able to use a single prompt in Blaze to generate high-quality content—both images and videos—for all your social channels, blog, emails, and websites for months at a time. Blaze will post that content for you and then learn from its performance.

This is what separates us from ChatGPT or Anthropic. While they can generate content, they can't improve it because they're not integrated with the platforms where the content is posted. Because Blaze can post for you and analyze the results, we can continuously improve based on what your audience actually responds to.

The AI companies that win won't just have the best models—they'll have users who trust them enough to share data that continually improves the system based on real-world feedback. And they'll build that trust in the first place by including product functionality that lets you edit and control exactly what gets published.

Looking further ahead, all of this infrastructure—reading your data, generating high-quality content, putting it out into the world, and learning from it—will extend beyond just social posts. It will power your outbound sales, customer service, and expansion efforts.

Building this future takes painstaking work. Tomorrow morning, just like every morning for the past 6 months, I'll wake up and manually review the top 50 most downvoted pieces of content with our team. We'll discuss what went wrong and how to fix it through design, product, or engineering changes. We'll do this every day until our models are good enough to do it on their own.

That's what it takes to get where we're going. It's not just about grand theories or elegant code—it's about grinding through the details, chipping away at the huge wall in front of us to get us closer to consistent, quality content that truly feels like you, one small improvement at a time.