For the Detail Oriented
It Learns From the Frustrating Conversations
Most chat tools show you a transcript dump and wish you luck. SiteChat closes the loop:
- It watches for frustration signals. Repeated questions, a visitor asking for a human, unanswered questions, and chats that end mid task are flagged automatically.
- It proposes the fix. A missing answer becomes a suggested knowledge update. A recurring request becomes the next Script Builder candidate. A confusing answer gets a clearer draft.
- Every change is retested. Before any improvement ships, it runs against a regression test battery, including the tricky questions your visitors already asked, so a fix in one place cannot quietly break another. Updates that fail the battery do not ship.
- You see it monthly. A plain language QA report shows what visitors asked, where the widget struggled, what improved, and what it captured for you.
Your learning stays yours. The system learns from your business conversations to serve your business. Your chat data is not pooled to train anything for other customers.
What Gets Measured, and What It Turns Into
| Signal measured | What it usually means | Fix proposed |
|---|---|---|
| Same question asked many times | Your site buries the answer, or never had it | A clearer passage for the knowledge base, and a suggested page addition for your site |
| Visitor asks for a human | The answer was missing, unclear, or the task needs a person | Knowledge fix if it was answerable; a Script Builder candidate if it is an action |
| Chat ends mid task | Friction: too many steps, a confusing question, a missing price | A shorter scripted flow, or the missing fact added to the knowledge base |
| Question the widget could not answer | A real gap in your site content | Draft answer for your approval, added to the knowledge base once you sign off |
| One action dominates volume | Your number one request is ready to become a script | A Script Builder draft, waiting for your approval before it runs |
Nothing ships on a guess. Every proposed fix is regression tested against the questions your visitors already asked, including the edge cases from your launch battery, and the changes that pass appear in your monthly report with before and after counts.
What the AI Actually Costs: Published, in the Open
Most chat vendors hide the model cost inside credits, multipliers, and per resolution fees. One rival charges US$0.99 for every resolved chat. Another charges about US$0.65 to US$0.78 per AI conversation. Here is what the raw AI actually costs at published API rates, November 2026:
| Model | Input, per 1M tokens | Output, per 1M tokens | One 10 message conversation |
|---|---|---|---|
| Groq GPT-OSS-20B | US$0.075 | US$0.30 | about US$0.002 |
| Google Gemini 2.5 Flash-Lite | US$0.10 | US$0.40 | about US$0.003 |
| OpenAI GPT-4o-mini | US$0.15 | US$0.60 | about US$0.004 |
| Groq GPT-OSS-120B | US$0.15 | US$0.60 | about US$0.004 |
| Anthropic Claude Haiku 4.5 (premium option) | US$1.00 | US$5.00 | about US$0.028 |
Assumptions, stated plainly: each message carries about 1,500 input tokens (the visitor question plus the retrieved passages from your site) and about 250 output tokens, across a 10 message conversation. Indexing your whole site costs under US$0.01 one time at published embedding rates of US$0.02 per million tokens. Sources are listed at the bottom of this page.
The Same Math at Different Conversation Lengths
| Model class | 5 messages | 10 messages | 20 messages | 300 conversations a month |
|---|---|---|---|---|
| Gemini 2.5 Flash-Lite class | about US$0.001 | about US$0.003 | about US$0.005 | about US$0.75 total |
| GPT-4o-mini class | about US$0.002 | about US$0.004 | about US$0.008 | about US$1.14 total |
| Claude Haiku 4.5 premium class | about US$0.014 | about US$0.028 | about US$0.055 | about US$8.25 total |
Read that last column again. A full month of Growth plan conversations, every one answered by a capable model, costs roughly a dollar in AI fees. Even the premium model at full volume stays under ten dollars. Anyone charging per conversation or per resolution is charging for the metering, not the AI.
And It Gets Cheaper as Scripts Take Over
Indexing is a one time cost of under a cent, and weekly refreshes cost the same order of magnitude, because only changed pages are re indexed. Then Script Builder starts eating the volume: if your top action is 40% of conversations and it moves to an approved script at zero inference cost, your monthly AI bill falls by roughly that share, from about a dollar to about sixty cents in the example above. The curve only bends one way, month after month, as the learning loop promotes more repeat actions into scripts.
The conclusion: a conversation costs between US$0.003 and US$0.03 in AI fees, and even the premium model stays under three cents. That is why our pricing is a flat monthly fee in Canadian dollars with a clear conversation allowance, instead of credits that burn at different rates depending on which model answered. And it is why we never route your visitors to a free tier: free API tiers can use conversation content to improve the provider products, which is no place for your customers names and phone numbers. Your chats run on paid tiers with no training on your data.
Sources for the Numbers on This Page
- Frustration statistics: Clutch, Fix Your Chatbot UX, 2026. clutch.co/resources/fix-your-chatbot-ux
- API prices: provider pricing as compiled by Hivebook, LLM pricing comparison, verified 24 October 2026. hivebook.wiki; Groq rates via LiteLLM pricing reference, 2026.
- Rival per conversation pricing: Tidio pricing page, checked 1 November 2026. tidio.com/pricing; Chatbase pricing page, checked 1 November 2026. chatbase.co/pricing
- Cost per lead benchmarks: WordStream and LocaliQ 2026, via PPCFolks, updated 30 October 2026. ppcfolks.com
- Conversational lead study: Isabella et al., Journal of Business Research, volume 201, 2025. sciencedirect.com
- Chatbot liability: Moffatt v. Air Canada, 2024 BCCRT 149, BC Civil Resolution Tribunal.
- Consent rules: CRTC, Canada anti spam legislation FAQ. crtc.gc.ca
SiteChat Lead Widget is built and operated by Daybreak Works, a Canadian company. Support is async and in writing, the same way the product works: no phone queues, and every answer on record.