AI agents are hitting the same accessibility barriers that screen reader users have faced for decades. The solution isn't new tech—it's the semantic HTML we should have been writing all along.
I've watched the web morph from table layouts and Photoshop slices, to Flash intros, to the rise of responsive design, to the mobile-first wave, to single-page apps, to whatever we're calling the AI-driven internet today.
One thing that hasn't changed — or rather, hasn't changed enough — is accessibility.
For most of my career, accessibility sat at the end of the checklist. We'd ship the "real" features first, and if there was time (there rarely was), we'd go back and make things work for screen readers, add keyboard navigation, fix our color contrast sins. It wasn't that we didn't care — it was that nobody was breathing down our neck about it, so it always slid to "later."
But now, something interesting is happening.
⸻
In the last two years, I've been watching a new kind of "user" land on the web: AI agents.
These aren't the SEO crawlers we've been dealing with since the late 2000s. These are language model–driven bots that try to use your website — click buttons, fill forms, read content, maybe even place orders — on behalf of human users.
And here's the kicker: many of them are terrible at it.
They struggle with:
<div>s pretending to be buttons with no aria-label)When faced with this mess, what do they do? They fall back to processing screenshots — like a vision-impaired user relying on OCR instead of clean, semantic HTML.
Sound familiar? It's exactly the same pain screen reader users have been living with for decades.
⸻
For years, we justified accessibility as "the right thing to do" or "a compliance thing." Now, it's also an infrastructure thing.
If an LLM agent can't navigate your site, it can't complete a task for its human user. That's not just bad for that one user — it's bad for your future discoverability, transactions, and integrations.
And this isn't a hypothetical future. I've seen agents fail on common e-commerce flows simply because a cart button had no label. I've seen booking flows choke because a date picker was a custom div with zero keyboard support.
⸻
Think about mobile app stores. They track crash rates and ANRs (Application Not Responding errors). High numbers hurt your rankings and visibility.
Now imagine a future where AI platforms (OpenAI, Anthropic, Perplexity, whoever) measure agent success rates on your site:
Poor scores could mean:
It's not hard to see an "Agent-Friendly" badge becoming as important as SSL locks or mobile-friendly labels.
⸻
The thing is, this "new" agent-friendly web isn't new at all. The very same semantic HTML, ARIA labels, keyboard navigation, and structured data that make a site accessible to people also make it usable for AI.
The only difference is that now there's a fresh economic incentive. And historically, that's when change actually happens.
⸻
If I could give one piece of advice to every dev and product manager right now, it's this: Build as if every single user — human or AI — is using a screen reader.
Because in a way, they are. Some of them just happen to be made of code.