Link Pilot|AI|suggestions
How AI and rules generate suggestions
Link Pilot does not let an AI invent your link graph. It indexes body HTML, runs deterministic rules in parallel with AI embeddings, scores each candidate, then uses a language model only to pick a safe anchor and write a short Why. You still approve, edit, or dismiss.
What AI does
- Embeds product, collection, article, and page text (per locale—no cross-language matching).
- Retrieves the top 20 similar destinations by cosine similarity (minimum about 0.32).
- Writes an insertable anchor that must already appear in the source copy, plus a Why grounded in signals—not invented product facts.
If AI embeddings are unavailable, rule-based suggestions still run (orphans, broken targets, underlinked hubs, missing commercial links).
What rules do
Rules inspect the body-link graph: who points to whom, which targets are missing, which hubs are thin. They do not replace AI; they catch structural problems embeddings can miss.
Pipeline
- Index and hash live HTML.
- Embed changed content.
- Run rule generators and semantic retrieval in parallel.
- Apply hard filters (no self-links, no existing pair, ignore-list handles, load-balance caps).
- Compute a deterministic confidence score.
- AI: title-substring anchor first; otherwise a validated phrase from the source.
- Persist for review on Suggestions.
Tips
- Prefer high-confidence semantic matches for blog → product paths.
- Use rule types first after a catalog delete (broken) or a thin hub (underlinked).
- If an anchor feels wrong, edit it—apply still requires the phrase in live HTML.
