I’ve been talking each week with my friend and mentor Barney Cohen about AI.

Barney built and took a company public in the late ’90s, and he’s one of the clearest thinkers I know about business and management. We’ve both been trying to understand how AI will change the way businesses operate, and I thought I’d start sharing some of our conclusions. Take 'em or leave 'em.

One question we keep returning to is: When and where is AI actually useful?

As a software engineer, it’s tempting to give AI everything I used to do. It can think through architecture, write code, test it, debug it and help deploy it much faster than I can.

But it still struggles with the final polish required to get an app across the finish line. Those big little details that make something feel considered and finished.

So is it accurate to say "AI is useful for the first 90% of a project"? Not really.

Something we realized is that AI is also weak at the first 10% of a project. That original idea. Knowing what's worth pursuing and what's not.

The concept that feels accurate today is: AI is appropriate for the middle 80% of work, and poorly suited at the edges. 

I feel "appropriate" and "poorly suited" are the right word too. It's not that AI can't come up with ideas, or push a project over the finish line, it's that the results are diminished.

Humans still matter most at the beginning, when we decide what deserves to be made, and at the end, when we decide whether it’s actually good. I’ve heard a lot about the importance of “taste” and the judgment required to refine and finish something well. But much less has been said about "discernment" and knowing what is worth making in the first place.

I suspect this pattern will appear throughout knowledge work, not just software engineering. Software is simply near the front of the wave.

I’ll keep sharing these little findings as Barney and I arrive at them, hopefully helping you find some footing, and opportunity, in your own work as we go.

In Safari Technology Preview 247, we’re introducing the Safari MCP server — a Model Context Protocol server for web developers that makes your web development and debugging workflow faster and more powerful. 

Can’t tell you how pleased I feel reading this. There are few things in the world I hate more than opening up the web inspector in Safari.

Something I have been watching recently is my SEO traffic dropping, which has had an obvious effect on sales. So I have been trying to figure out whether this is specific to me or part of a broader trend.

Tim Ferriss, the author of The 4-Hour Workweek, published a very interesting piece on how his how-to nonfiction sales have changed as AI tools became the first place people go for quick answers. His sales are down 50% year-over-year, two years in a row.

Something broader is happening here.

For years, the marketing funnel was pretty simple: someone searched Google, found a helpful article, learned to trust the person or business behind it, and maybe eventually bought something.

That was the deal. You created helpful content. Google sent people to it. Some of those people became customers. Potential customers could evaluate your paid products based on how helpful your free products were.

In the Google world, if your article helped answer the question, you had a decent chance of getting the click. In the AI world, your article may help answer the question without the person ever seeing your name, visiting your site, joining your email list, or buying your product.

Tim Soulo shared that Ahrefs has been analyzing more than 1 billion data points across 14 AI search optimization studies, and this finding jumped out: “ChatGPT only cites about 50% of the URLs it retrieves. It fetches dozens of pages per query but uses half as background context without attribution. This means that being retrieved and being cited are very different things.”

A big fear people have is that AI might replace what you sell, but I'm much more concerned with not getting to meet potential customers. Without reaching them, you're just building in the dark.

I do not have a clean solution to this yet. I’m still trying to understand it myself. But the thing I’m watching is not just whether AI can make better products or write better code. I’m watching where people go to get their problems solved. When AI can't solve a problem, where does the person go?

The search → article → trust → purchase funnel is breaking down, so it's time to start imagining different funnel structures and marketing channels.

The more I use AI, the more I realize how important it is to be precise with my words. This is especially important to me in the world of web design, where AI can get me 80% of the way there, but the last 20% still needs fine-tuning. And that fine-tuning depends on being super clear about what I want.

So I enjoyed this list of design language definitions. For instance: do you know the difference between kerning and tracking?

Kerning - The space between two specific characters, adjusted manually. Different from tracking, which affects all characters equally.

Tracking - Letter-spacing applied uniformly across a word or block of text. Uppercase labels almost always need more of it.

I've been using them interchangeably, and never really considered the difference. It's never been more important to think and write clearly.

It seems fairly likely to me that by the end of 2027 the majority of internet traffic will be Agents using the web. AI bots, basically.

I spent some time 'researching' website traffic trends over time, with Claude and ChatGPT and got fairly similar results.

See full graph and prompt here: https://claude.ai/public/artifacts/bd39047b-711f-4cf2-adf7-892651543ac7

The green line is human traffic. The yellow line is "bot" traffic (Google SEO). The red line is malicious bot traffic. The orange line is AI traffic.

To be clear, this graph isn't saying people will be using the internet less, although that might likely be true, it's saying as a share of website requests, AI is going to dominate them.

To me, this tells me the primary visitor of your website at the beginning of 2028 will be an AI Agent. I'm not predicting this uniformly across all industries and niches, maybe not your website, but on aggregate this is my prediction.

So what does that mean for web designers? What does that mean for Squarespace?

Plugin sales have been trending down. Not dramatically, but consistently enough that I'd be lying to myself if I pretended otherwise.

Here's what the last four years look like:

It's important not to misattribute casualty here. This graph is saying "people are buying less plugins", but it doesn't say why.

The "why" is the puzzle.

Are people using AI to build their own plugins?

Is search engine traffic also down, resulting in less top-of-funnel traffic?

Is it because I haven't released any new plugins recently?

Is there even a need for websites anymore?

Probably some mix of all of the above. Or something else entirely I'm not considering.

As we experiment and figure out what the next phase of the plugin business looks like, I'm continuing to iterate and build on Robo-Will and The Toolkit.

And as weird this might sound, I'm enjoying the puzzle.

From SFGate:

Allbirds, the millennial beige shoe brand that was once a mainstay among affluent tech workers in San Francisco, just made the stunning announcement that it will become an AI cloud computing service. 

wut?

According to a Wednesday news release, American Exchange Group, which entered a purchase agreement with Allbirds earlier this year, intends to “build on Allbirds’ legacy and deliver compelling products to Allbirds’ customers” by changing its name to NewBird AI.

“NewBird AI’s long-term vision is to become a fully integrated GPU-as-a-Service (GPUaaS) and AI-native cloud solutions provider.” GPUaaS is a cloud-based service that provides access to computational power without the cost of expensive infrastructure. 

So they changed their name, changed their industry, and have new owners. That's just called a new business.

Squarespace has recently released their contacts API with a some damn fine documentation to boot.

Use the Contacts API to manage the people associated with a Squarespace merchant site — customers, mailing list subscribers, donors, and anyone else who has interacted with the site.

Access is OAuth only, which basically means you have to be an approved extension provider, but they're saying API access "coming soon", so hopefully we won't have to wait too long!

To be honest, I've been starting to lose hope that Squarespace was going to invest in their infrastructure like this. Their closed garden approach hasn't been helpful for them in the age of AI.

But this is shows they've got a lot of work on behind-the-scenes and I hope it continues.

The developer website received a fresh coat of paint as well. Excited for what's happening over there.

This websites has a list of 30 design prompts to make your projects not look like that bland vibe coded style.

Even if you're not about to build anything it's fun to sift through all of these and see the different styles. I'm a sucker for simple names and I love the names for each of them.