Building a WordPress AI assistant: lessons from a real deployment
We shipped an AI assistant inside a production WordPress site. These are the integration, latency, and content-grounding lessons that only show up in the real world.
We shipped an AI assistant inside a production WordPress site — answering visitor questions, guiding people to the right content, and drafting replies for the editorial team. It worked, but not for the reasons the demo suggested it would. The lessons that mattered only showed up in the real world, under real traffic and real content mess.
Grounding beats generation
The single biggest quality lever was grounding every answer in the site's actual content. A general model answering from memory hallucinated product details, invented pages, and gave confidently wrong advice. Once we retrieved the relevant posts and pages and constrained the assistant to answer from them, accuracy jumped. The model's job became synthesis and tone, not recall.
The hard part was not the retrieval — it was the content itself. Years of duplicated, outdated, and contradictory pages meant the assistant sometimes grounded itself in the wrong source. Cleaning and structuring the content did more for answer quality than any model change.
Latency is a feature
On a content site, a slow assistant is a broken one. Visitors will not wait eight seconds for an answer. We had to stream responses, cache embeddings, and keep the retrieval index warm, because perceived speed determined whether people used the feature at all. The lesson: budget for latency engineering from the start, not as an optimization pass at the end.
Integration is the real work
The model was maybe twenty percent of the effort. The rest was WordPress itself — hooking into the theme without breaking it, respecting caching layers, handling logged-in versus anonymous visitors, and giving editors a way to correct the assistant when it drifted. A production AI feature is mostly plumbing, and the plumbing is where the schedule goes.
If we did it again, we would scope the integration work first and the model work second. The AI is the easy part to demo and the easy part to overestimate; the deployment is where a real product is won or lost.