I’ve been testing a few different internal linking workflows on a couple of niche sites over the last few months, and I’m starting to think the “AI will solve internal linking” pitch is a bit overstated.

To be clear, I do think AI can help with the grunt work. On one site, I used a simple workflow to suggest contextual links from new posts to older supporting articles. It saved time, especially on clusters where the topical relationships were already obvious. But once I pushed it beyond the easy cases, the quality dropped pretty fast.

A few things I’ve noticed:

– **AI tends to over-link generic terms**
It will happily suggest links for words that technically match but don’t make much editorial sense. That creates noise if you’re not filtering carefully.

– **Exact-match relevance isn’t enough**
For example, it might link “best hosting” to a page about server uptime just because the keyword overlap is there. That’s not always useful for users or search engines.

– **The best results came from tighter rules**
When I constrained the system to only link within specific content clusters and only from pages with clear intent alignment, the output was much better.

– **Manual review is still necessary**
At least in my experience, fully automated internal linking is fine for low-risk sites, but on money pages or competitive niches I’d rather keep a human in the loop.

What I’m curious about is whether anyone here has built a workflow that actually scales without turning into a cleanup job later.

Specifically:

1. Are you using AI for internal linking, and if so, how are you controlling relevance?
2. Do you link from old posts to new posts, new to old, or both?
3. Have you seen any measurable ranking lift, or is the main benefit just efficiency?
4. Are you using plugins, custom scripts, or something like GPT-based content analysis?

I’m not anti-AI at all — I use it in my own processes — but I’m skeptical of any setup that promises “fully automated internal linking” without tradeoffs. In SEO, the details usually matter more than the demo.

Would be interested to hear what’s actually working in real projects.