Technical SEO, search intelligence, and the code behind it — notes from 20+ years of making websites work harder.
Same model file, same machine, genuinely different machinery around it. Ollama wraps llama.cpp, so the gap isn't the engine — it's the binary, the defaults and a bit of version drift. Here's what I found once I went CPU-only.
Vectors understand the vibe of what people are looking for, but the synthetic-pollution problem is a real challenge.
llms.txt is not a ranking signal and the agents it was built for don't fetch it. Yet Google added it to Lighthouse under Discoverability. The file's only real shot at mattering is as a map handed to browser agents — next to WebMCP, not in search.
An HNSW vector index is not a property of the column it sits on. It's a property of the query shapes the planner can rewrite. Mismatch the shape and the index does nothing — even when the SQL looks like it should benefit.
Cypher is expressive, graphs fit SEO data, embedded options exist. None of it survived the question 'what happens when the user opens DevTools?'
The Shadow DOM is a secret room inside your web page — with its own CSS, its own events, and the quiet ability to vanish from your SEO audits and every AI crawler that ever visits.
Determining which entities a document fundamentally discusses is a deceptively simple problem. Here's what I tried for Italian — and why none of it worked.
A quietly overlooked EMNLP 2019 paper anticipated the idea now driving modern generative search. Six years later, it finally deserves a name.
By August 2025, nine months after MCP's release, Ollama still had no native support. I built a working prototype in a weekend.