SiteGPT's analytics dashboard is live for everyone: stat tiles and an engagement funnel on every plan, trend charts and knowledge gaps on Growth and above, AI conversation insights on Scale and above. So this is not a list of workarounds for missing analytics.
It is a list for the two situations a dashboard cannot help with. Your plan does not include the view you want yet, or the number is on screen and the actual work, reading the transcripts, drafting the fix, shipping it, is still yours. Connect Claude to SiteGPT over MCP and both become something you ask for.
Each question below has the exact prompt to paste, which of the three connector tools Claude reaches for, what comes back, and the follow-up action. The prompts are the point: copy them.
iShort answer
Connect Claude to SiteGPT from Claude's connector directory and ask it support analytics questions in plain language. Claude works through three tools: search finds the right operation, execute_read pulls conversations, leads, knowledge, and usage read-only, and execute_write changes things only with confirmation. The payoff is twofold. On Starter and Growth, Claude's own analysis of your raw data reaches answers the dashboard reserves for higher tiers, like topic clustering. On any plan, Claude takes the step after the number: drafting the missing help article, saving the fix, triggering the re-sync, writing the Monday summary in your format.
The dashboard is live and tiered; Claude reads the same data through your own permissions and does the step after the number.
The dashboard shows you (live today)
- Every plan: conversations, AI replies, visitors, and leads, plus the widget-to-conversation-to-lead funnel
- Growth and above: daily trends, CSV export, knowledge gaps, and the Monday digest email
- Scale and above: AI-classified topics, conversation outcomes, and top questions grouped by meaning
- By design, no separate ticket-deflection, satisfaction, or resolution-time metrics
Ask Claude when
- Your plan does not include the view: Claude classifies raw transcripts itself on any tier
- You want a cut the dashboard does not offer: by page, by period, by lead quality, charted in the chat
- You want the next step done: draft the missing article, save the fix, trigger the re-sync
- You want it in your format: a Monday summary written for your Slack channel, not a fixed email
Setup is not covered here; the MCP chatbot guide walks through connecting Claude, and the connector directory post covers the two-click version. One security fact carries this whole page: the grant is a subset of the connecting user, never a superset, so Claude sees what your login sees and nothing more.
1. Which questions did the bot fail to answer this week?
The dashboard's answer: knowledge gap identification on Growth and above, and the weekly digest email samples the questions the bot could not answer. The full, current list on any plan is an ask away:
Pull my SiteGPT conversations from the last 7 days. List every question the bot
failed to answer well: fallback answers, visitor downvotes, or the visitor
rephrasing the same question more than once. Group similar questions and rank
groups by frequency.
Claude uses search to find the conversation operations, then execute_read to pull them. What comes back is a ranked list of failure clusters with example transcripts. The follow-up action is question 8: pick the top cluster and draft the fix.
2. What topics keep repeating, and chart them
AI-classified topics are part of AI conversation insights on Scale and above. On Starter or Growth, Claude does the classification itself from raw transcripts:
Pull the last 30 days of conversations and classify each one into topics by
meaning, not keywords. Give me the top 10 topics with counts, and render a bar
chart. Note which topics are new versus 30 days ago.
Claude renders the chart in the conversation; SiteGPT does not generate or store it. This is Claude's analysis rather than the dashboard feature, which cuts both ways: it is available on every tier and shapeable on request, and it is only as current as the data Claude pulled.
3. Which conversations ended with a request for a human, and why?
Conversation outcomes on Scale show the handed-off share. The share is a number; the reasons live in transcripts:
Find conversations from the last 14 days where the visitor asked for a human or
the chat was escalated. Read them and group the reasons: missing information,
policy exceptions, frustration with answers, or something else. Quote one
example per group.
The answer tells you what to fix in training content or escalation setup: missing content is a knowledge fix, policy exceptions are an instructions fix, and frustration clusters are worth a human review of tone.
4. How many conversations became leads, and where is the drop-off?
The engagement funnel is on every plan: widget opens, then conversations started, then leads collected, each stage showing its share. Claude adds the cuts:
Pull conversations and leads for the last 30 days. What share of conversations
produced a lead? Break it down by week and flag the days that underperform.
Then look at 10 conversations that did not convert and tell me the most common
point where the visitor dropped off.
Reads only, via execute_read. The useful output is the last part: the dashboard tells you the ratio, the transcripts tell you whether the lead form asks too early, too late, or for too much.
5. What changed since last month?
Trend charts are Growth and above, and the digest reports week-over-week. For any other window, name it:
Compare this month to last month for my chatbot: conversations, AI replies,
leads, and downvote rate. Show the deltas, then tell me the likeliest drivers
of the biggest change by sampling conversations from both periods.
The delta is arithmetic; the driver analysis is why this is worth asking an assistant instead of reading two numbers off a page.
6. How much of my plan am I actually using?
Message limits are a range, not a single number: the headline quota assumes the cheaper model, and both the question and the reply count as messages. Which makes mid-month checks worth doing:
Check my SiteGPT usage: messages, pages, chatbots, and team seats against my
plan limits. Project message usage to end of month at the current rate, and
tell me if I am on track to hit a limit.
Claude pulls live usage through execute_read. If the projection says you will run out, the options are the model split in billing settings, the extra-messages add-on, or the next plan up.
7. What are people asking that the knowledge base does not cover?
The dashboard's knowledge gap identification (Growth and above) surfaces gaps with one-click Q&A fixes. Claude's version works on any plan and checks the other direction too, what you trained versus what is actually asked:
Pull the last 30 days of conversations and my chatbot's list of trained
documents. Which recurring questions have no matching document or custom
response? Which trained documents answer questions nobody asks? Rank the gaps
by frequency.
The second half is the underrated one: knowing which content earns its place tells you where re-training effort is wasted.
8. Draft the missing help article for the biggest gap
This is the step no dashboard takes, on any plan or any product, and it is where the agent-ready story stops being abstract:
Take the biggest knowledge gap from the last analysis. Draft a help article
that answers it, in the same tone as my existing trained content, using real
visitor phrasings from the conversations as section headings. Then save the
top question and its answer as a custom response in SiteGPT.
Drafting happens in the conversation. The save at the end goes through execute_write, which is marked destructive, so Claude asks for confirmation before anything lands in your workspace. Nothing ships without you approving it.
9. Which answers lean on stale content, and what should be re-synced?
Auto-sync keeps trained content fresh on a schedule (monthly on Growth, weekly with daily scan on Scale). Between syncs, or on Starter, ask:
List my chatbot's trained documents with their last sync dates. Flag anything
older than 30 days that visitors are actively asking about, and tell me which
documents to re-sync first. Then trigger the re-sync for the ones you flagged.
The listing and cross-referencing are reads; the re-sync trigger is a write, so it runs after your confirmation. Stale pricing and policy pages are the classic case: high traffic, quiet drift, and the bot answers from whatever it last saw.
10. Give me a Monday morning summary I can paste into Slack
Growth and above already gets a Monday digest email: last week's numbers with week-over-week changes, top topics, and a sample of unanswered questions, sent to managers and admins. It is good, and it is fixed. Claude's version is yours:
Write my Monday chatbot summary for Slack. Last week vs the week before:
conversations, leads, downvote rate. Then the top 3 topics, the top 3
unanswered questions, and one recommended action for the week. Keep it under
150 words, no fluff, bold the numbers.
Any plan, any format, any emphasis. Paste it into your channel, and when someone asks a follow-up, the data is one more question away rather than in next Monday's email.
What the dashboard does not measure, and what to do about it
SiteGPT's docs are unusually direct about scope: there are no separate ticket-deflection, satisfaction-score, or resolution-time metrics beyond the conversation outcomes listed. Most vendors leave that for you to discover.
That stated limit is a genuine ask-Claude use case. Claude can read raw conversations and build honest proxies: resolved-without-handoff share as a deflection proxy, downvote rate as a satisfaction proxy, message-count-to-resolution as an effort proxy. Label them as proxies when you report them; they are estimates built from transcripts, not instrumented metrics.
Frequently asked questions
Do I need the Scale plan to get topic analysis for my chatbot? In the dashboard, yes: AI conversation insights, which includes AI-classified topics, conversation outcomes, and top questions grouped by meaning, is part of Scale and above. On Starter or Growth, the honest workaround is to connect Claude over MCP, pull the raw conversations with execute_read, and let Claude do the classification itself. The result is not the dashboard feature, it is Claude's own analysis of your data, which also means you can ask for cuts the dashboard does not offer.
How do I connect Claude to SiteGPT?
Two ways. Find SiteGPT in Claude's connector directory and click Connect, or add https://sitegpt.ai/mcp as a custom connector. Either way, approval happens over OAuth in the browser: you sign in to SiteGPT and choose whether the grant covers all chatbots or only selected ones. The full setup walkthrough is at the MCP chatbot guide.
Is it safe to let Claude read my support conversations? The permission model is the answer to evaluate. The grant is scoped to the connecting user: the client cannot receive permissions or chatbot access your dashboard user does not have, so Claude sees exactly what your own login sees and nothing more. Approval is per-user over OAuth, can be limited to selected chatbots, and ends when the connector is disconnected. No API keys are pasted into chat.
Can Claude change my chatbot, or only read from it? Both, deliberately separated. Reads go through execute_read, which is marked read-only. Anything that creates, updates, or deletes goes through execute_write, which is marked destructive so the client asks for confirmation before running it. That is why the drafting workflow in this list saves a custom response only after you approve it.
Does SiteGPT store or render the charts Claude makes? No. Claude renders charts in the conversation from the data it pulled; SiteGPT does not generate or store them. The dashboard's own charts, the stat tiles, funnel, trends, and insights, live on the Analytics page and are described in the analytics docs.
What analytics does the SiteGPT dashboard include without Claude? Every plan gets stat tiles (conversations, AI replies, visitors, leads) and the engagement funnel from widget opens to conversations to leads, with a date range picker. Growth and above adds daily trend charts, CSV export, knowledge gap identification, and the Monday weekly digest email. Scale and above adds AI conversation insights: AI-classified topics, conversation outcomes, and top questions grouped by meaning. The docs also state there are no separate ticket-deflection, satisfaction-score, or resolution-time metrics.
Sources
- SiteGPT analytics docs for the stat tiles, funnel stages, per-plan gating, and the no-separate-metrics statement, read 14 August 2026
- SiteGPT conversation insights docs for topics, outcomes, top questions, and Scale-and-above gating, read 14 August 2026
- SiteGPT weekly digest docs for the Monday email contents, recipients, and Growth-and-above gating, read 14 August 2026
- SiteGPT MCP server page for the three tools, the server URL, OAuth via Dynamic Client Registration, and the subset permission wording, read 14 August 2026
- Claude connector directory listing for the SiteGPT listing, verified from a signed-in session 14 August 2026
- Model Context Protocol for the open standard
Last updated: August 2026. All SiteGPT docs cited were read directly on 14 August 2026.