Short answer: An AI answer engine for a website indexes the company's own pages, documents and FAQs, finds the passages that match a visitor's question, and has a language model write a short answer that links to those sources. It comes as a hosted widget or a custom RAG build, and its quality depends more on the site's content than on the model.
Key takeaways
- Retrieval first, generation second: the model answers only from passages pulled out of the site, and every answer carries source links.
- Hosted tools exist: Algolia Ask AI, Kapa.ai and Google's Vertex AI Search all return generated answers with citations.
- Hebrew needs its own test: spelling variants and attached prefixes make retrieval the weak point.
- Unanswered questions are data: each one points to a missing page or a lead.
How an AI answer engine works
A crawler or CMS feed splits each page into passages and stores them in a search index, usually with embeddings for meaning-based matching. When a visitor asks a question, the engine pulls the best-matching passages, sends them to a language model with the question, and instructs it to answer only from that material and cite the source page. This is retrieval-augmented generation, explained in this guide to RAG and AI document search; terms like embeddings are defined in an AI terms glossary for business. Classic site search returns ten links; an answer engine reads them and replies in two or three sentences, with the links underneath.
Build options: hosted tool or custom RAG
| Option | What it does | Fits when |
|---|---|---|
| Algolia Ask AI | Chat mode in DocSearch, answers linked to indexed pages, own LLM key supported | The site already runs Algolia search |
| Kapa.ai | Widget over docs, forums and tickets, cited answers with feedback votes | A technical product with large documentation |
| Vertex AI Search | Google Cloud site search with cited generative answers, an add-on to advanced site search | The team already works on Google Cloud |
| Custom RAG | Own index, prompts, model and fallback flow, connected to CRM | Hebrew-first content, private documents or custom lead flows |
Hosted tools go live quickly and are built mostly around developer documentation. A custom build, through AI system development on the company's own stack or a document-grounded agent like DocBrain, takes longer and gives full control over retrieval, tone and what happens after the answer.
