ConvTag reads your website, builds an AI agent grounded in your own content, and embeds it on any page with one line of HTML. Here's exactly what happens.
Most teams go live before lunch — no engineering required.
Sign up, paste a URL, and the crawler follows your sitemap and the obvious hub pages (about, pricing, FAQ, docs). Tick the ones you want indexed, and the jobs run in the background — they honour robots.txt, skip authenticated paths, and never wander into internal hosts.
Crawled pages are extracted with Readability, chunked along semantic boundaries (about 500 tokens with overlap), embedded with Cloudflare bge-base-en-v1.5 or OpenAI text-embedding-3-small, and upserted into Vectorize or Qdrant. Every chunk is tagged with workspace and agent so retrieval is strictly tenant-scoped.
Set the persona, tone, default language, theme colours, starter prompts, and the behaviour rules. Pin curated answers for the questions you want answered in your exact words (refunds, pricing caveats, disqualification rules). A live preview shows what the visitor will actually see.
Hitting Publish writes the agent into an immutable version row. The widget runtime always reads the published version, so edits to a draft never leak to live visitors. Any prior version is one click away if you need to roll back.
Paste a single <script> tag before </body>. The widget bundle is under 50 KB gzipped, loads asynchronously, and renders inside a Shadow DOM so it cannot fight your site's CSS. It works on any framework — WordPress, Shopify, Next.js, plain HTML, Webflow.
A visitor types a question. The hot path: curated short-circuit check, embed the query, vector search, rerank, assemble the prompt with sources tagged for prompt-injection defence, stream the LLM response back over SSE. No DB writes, no synchronous webhooks — persistence and analytics fire asynchronously after the stream finishes.
Behaviour rules fire when a visitor signals real intent — asks about pricing, requests a demo, or reaches the third turn — and offer the inline lead form. Captured leads land in your inbox, fire a Slack alert, and POST to your webhook for HubSpot, Pipedrive, or Mailchimp. A human can take over mid-conversation without the visitor noticing the swap.
No DB writes, no synchronous webhooks, no retries. Persistence and analytics are dispatched async after the stream completes. OpenTelemetry spans wrap every phase, so when the budget breaks, the span heatmap points right at the offender.
Free to start. No card required. Live in 5 minutes.