See what your customers ask, before they tell you

A traditional research program — surveys, user interviews, usability tests — has the same chronic problem: it costs time and money to ask people what they want, so you don’t ask very often, you don’t ask very many, and the people who reply are rarely the ones who matter most. Meanwhile, your support inbox quietly fills with the answers, but nobody on the product or marketing side ever reads it.

A chatbot inverts that. Every visitor who has a question types the question. They type it freely, in their own phrasing, with their own assumptions. They are not framing it for a researcher; they are not editing it for politeness. They want an answer. That gives you a continuous, low-cost stream of authentic voice-of-customer data — provided you have somewhere to look at it.

That is what NebulaHex Analytics is for. Not a vanity dashboard with a single “messages this week” number. The layer that takes the raw transcripts your bot is already producing and surfaces the patterns: which questions repeat, which ones the bot struggles with, when traffic spikes, where your knowledge has holes, and how many of those conversations turn into a real lead.

NebulaHex ships with two analytics surfaces side by side. They share the same metric definitions and the same time-range selector. You’ll move between them constantly.

Both surfaces share a time-range selector: 7 days, 14 days, 30 days, or All. Picking a range re-runs every chart, every metric card, every table on the page against that window. Default is 7 days — the right cadence for a weekly health check. Switch to 30 or All when you’re hunting for trends or building a quarterly review.

Across the top of every analytics view sit four headline metric cards. They look simple. They are not — each one tells a specific story about a specific failure mode, and reading them in order is how you keep your bot honest.

Fallback rate gets its own visual treatment because it’s the leading indicator the other three lag. Conversations and Leads only tell you what happened. Fallback Rate tells you whether your bot is starting to lose its grip on your knowledge — before that loss shows up in the conversion numbers.

Below the metric cards sit two charts that turn point-in-time numbers into shapes you can read.

Messages Over Timeis a daily timeline of message volume across the selected window. The shape tells you several things at once. Steady-state usage shows up as a flat line with weekly cycles. A campaign launch shows up as a sharp peak. A bug that took your widget offline shows up as a flat zero day in the middle of the line — usually correlatable to something else that happened in the business. You won’t read this chart every day, but you should read it before every weekly review, because anomalies stand out fast.

Response Confidenceis the chart that rewards careful reading. Every time your bot answers, it computes a confidence score from 0 to 100%. The chart bins those scores into four buckets: Low, Medium, Good, and High. A healthy bot’s distribution is heavily weighted toward Good and High. A bot that’s drifting will start to bulge in the Medium and Low buckets — and the fallback rate climbs in lockstep.

The reason this chart matters more than the average confidence number is that an average hides the tails. A bot whose replies split 50/50 between High and Low can have the same average confidence as a bot whose replies are uniformly Medium — but the first bot is delighting half its visitors and failing the other half, while the second bot is mediocre across the board. The distribution shows you which one you have.

Two tables sit below the charts. The first is the one your marketing team will love.

Most Asked Questionsis a top-10 list of the exact phrases visitors typed into your bot, ranked by frequency, for the selected time window. No preprocessing. No clustering. No “intent” abstraction layer hiding the words. You see what your customers literally typed.

This list is gold. The single most valuable view on the page for anyone outside support. A content marketer reading this list will leave with their next three blog posts. A product manager will find the missing feature their users keep wishing for. A founder will see, in five minutes, what their customers are confused about — without commissioning a study, without running a survey.

Performance is the second table. Three numbers your operations team cares about:

• Average Response Latency — how long, on average, your bot takes to reply. A grounded answer takes longer than a fallback. A bot with rich knowledge will run a bit slower than a thin one — that’s the model thinking, not a bug.
• Model Usage — the cumulative compute across the time range. The number you’ll cite when explaining usage growth, and the number that matters if you’re using BYOK.
• Average Confidence Score — the arithmetic mean of every assistant reply’s confidence. Good for trend tracking. For diagnosis, read the distribution chart above.

Per-bot analytics has an Export CSVbutton at the top right. Pick a time range — 7 days, 30 days, All — then click export. You’ll get a CSV containing the conversations and lead data for that window, ready to drop into a spreadsheet, a BI tool, or whatever pipeline your data team already uses.

The export is the right tool when you want to do something the dashboard doesn’t yet: pivot leads by source, build a custom funnel, hand a slice of conversations to a research firm for tagging, or join chat data against your CRM in a warehouse. The dashboard answers most of the questions most teams have. The export answers the ones nobody could have anticipated.

The same analytics page serves three very different teams in very different ways. Here’s what that looks like.

Your conversations are your data. They’re stored encrypted at rest and transmitted over encrypted connections. Access is gated by role-based permissions inside your workspace — owners and admins see everything, editors and viewers see what their role allows, and nothing crosses the boundary between organizations. Conversations and analytics are scoped per workspace.

Your data is never used to train AI models. Teams using BYOK route inference through their own provider account — same dashboard, same metrics, billing flows through their provider relationship. See our privacy policy and terms for the full posture.

Connect a bot, get visitors typing, and your analytics page has signal in it the same hour. Open the Analytics tab in your sidebar to see the global rollup, click into any bot for its dedicated view, and spend fifteen minutes a week reading the patterns. You’ll know what your customers want before they fill out a survey, before they file a ticket, and well before they leave for a competitor.