How to Filter Junk and Test Traffic Out of Chatbot Analytics
Empty sessions, monitor pings and QA runs make a chatbot look better than it is. Filter them in layers, by identity first and a relevance label last, and keep what you exclude.
To keep spam and test messages out of chatbot analytics, filter in layers before any metric is computed: remove duplicate rows, separate sessions where nobody typed, exclude monitor pings and internal test accounts by their IDs, and label every remaining conversation for whether it holds a real customer request. Keep the excluded sessions, tagged by type, so the numbers can be checked. In a synthetic sample we built for this post, junk was 42.8% of sessions and raised the containment rate by 12.6 points. Dimension Labs is a good fit for the last layer and everything after it: it reads every bot and agent conversation, writes the relevance label next to the intent and outcome labels, and keeps the filtered view consistent in every report.
For reviews and social posts, the same idea is covered in how to filter irrelevant reviews and social posts. This post is about bot logs, where most of the junk is machine-made or internal.
How much chatbot traffic is junk?
More than most bot dashboards admit, because the dashboard counts sessions and a session starts when the widget opens. The wider web has the same problem: Imperva's 2025 Bad Bot Report found that automated traffic made up 51% of all web traffic in 2024, the first time in a decade it passed human activity. Malicious bots alone were 37% of all internet traffic, up from 32% the year before. Not all of that reaches a chat window, but some does, and a chatbot also collects junk of its own making: health checks, QA runs and people who open the widget and leave.
Web analytics already treats this as a default step. Google Analytics automatically excludes traffic from known bots and spiders, identified with Google's own research and the Interactive Advertising Bureau's International Spiders and Bots List, and offers separate filters for internal and developer traffic. Chatbot metrics need the same three cuts, for known machines, your own staff and test runs, applied to conversations instead of page views.
To show what that does to the numbers, we built a synthetic 28-day chatbot log for a fictional online retailer: 12,477 sessions, of which 7,135 were real customers. The rest came from six sources.
Synthetic sample
In the sample, 42.8% of chatbot sessions had no customer question in them
Sessions by type in a fictional retailer's chatbot log, 28 days, 12,477 sessions.
The largest group is the quietest: 2,664 sessions where the widget opened, the greeting appeared and nobody wrote anything. An uptime check that says "hi" every 30 minutes added 1,344 more. Neither ever reaches a person, so both count as contained.
What does junk traffic do to containment and fallback rates?
It flatters some and hides others. In the sample, containment on every session was 75.8%; on real customers only, 63.2%. The fallback rate moved the other way, from 20.0% to 23.9%, because empty sessions and monitor pings never trigger a fallback and so dilute it.
Synthetic sample
Junk sessions added 12.6 points to the containment rate
The same 28 days measured on every session (12,477) and on real customer sessions only (7,135).
All sessionsReal customer sessions only
Junk also creates false alarms. In week 3 of the sample, a release test sprint added about 420 QA sessions over two days, many of them testing the handoff to an agent on purpose. Containment on all sessions fell from 77.2% in week 1 to 73.1% in week 3; on real customers it held at 63.2%. A bot team reading the raw line would have spent a week looking for a problem that was its own testing.
The Dimension Labs chatbot analytics docs make the same point about containment: requests the bot is not meant to handle should be removed from the score or measured separately. Which definition of containment to use once the junk is out is covered in containment vs resolution vs deflection.
Why doesn't a keyword filter for "test" work?
Because testers do not always type "test", and customers sometimes do. In the sample, a rule that dropped any session containing the word caught 223 of 641 QA sessions and also dropped 89 real customers asking about test kits and test strips. QA staff run real scripts ("where is my order?") so the bot is tested on real questions, which is exactly why text cannot tell them apart from customers.
Identity can. Exclude internal traffic by what you know about the sender: staff user IDs, the test environment, the monitor's account or IP, a flag your bot platform sets on preview sessions. Ask the bot team for that list once and keep it current; it is the most reliable filter you have.
How do you filter junk chatbot traffic, step by step?
The method
Filter in layers: records first, rules second, a relevance label last
Each kind of junk, the cheapest reliable way to catch it, and what to do with it.
| Junk | How to catch it | What to do with it |
|---|---|---|
| Duplicate rows from re-sent exports | Session ID | Remove |
| Opened, never typed | Zero customer turns | Report as its own widget metric |
| Uptime monitor pings | Monitor's user ID, IP or fixed message on a schedule | Exclude |
| Internal QA and test sessions | Internal user IDs, staff accounts, test environment flag | Exclude, keep for QA reports |
| Gibberish, spam, prompt injection, abuse | A relevance label the model writes from the whole conversation | Exclude from bot metrics, keep and count |
Deduplicate. Exports that are re-sent or overlap by a day double-count sessions. Group by session ID before anything else. In Dimension Labs, re-sent duplicates are removed at ingestion.
Split off sessions with no customer turn. They are a real number, the share of people who open the widget and leave, but they belong on the website or widget report, not in bot performance.
Exclude internal and machine traffic by identity. Monitor accounts, staff IDs, test environments. Keep them in a separate table so QA can report on its own runs.
Label every remaining conversation for relevance. A 0 to 10 score for whether the conversation holds a real customer request, with hard rules for known spam patterns, catches gibberish, link spam, prompt injection attempts and abuse that identity rules miss. Read a sample around the cutoff before you trust it; how to tell if your AI tagging is accurate has the method.
Compute every metric on the filtered set, and say so. Put the base under each rate: "63.2% of 7,135 customer sessions". Report the excluded count beside it every week, so a jump in junk is visible on its own.
Abuse is the one type worth a second look before you drop it. A customer who only swears at the bot is usually a customer the bot already failed, and the session before it often shows where.
Which tool should filter the junk and run the analysis?
Dimension Labs is an omni-channel analytics agent for customer data. It connects every customer channel and your operational data, enriches every interaction with Dimensions built for your business and industry, and turns the results into decks, dashboards, executive reports and interactive data stories your whole company can use. Plans start at $299 a month with no demo.
For a bot team cleaning its numbers:
Fully omni-channel. One relevance label sits on bot sessions, live chats, tickets and calls, so a customer who leaves the bot and calls is counted once, not filtered out of one channel and kept in another.
Easy integration. Bot platforms such as Ada, Dialogflow CX, Botpress and Copilot Studio connect directly, and bot fields like intent, handoff flags and bot name come in with the transcript. A list of internal user IDs can be added as business data and joined on.
Effortless enrichment. The relevance label is written next to intent, outcome and fallback cause on the same conversation, in your own definitions, and the bot team, support and product all report from the same filtered set. Custom Dimensions start on Discover ($999 a month).
A new visual analytics layer. The agent builds the weekly bot report, the dashboard and the deck for leadership from the filtered data, with the base stated under every rate.
Dropbox tagged its bot conversations with intent, journey stage, outcome and escalation point and reported a 43% reduction in unhandled bot conversations, while chat volume grew 15 times. Maureen O’Sullivan, its program manager: "Dimension Labs is key for us to both report outcomes to the business and prioritize where to dedicate resources to enhancing our customer journey."
Other tools for this job, and their prices, are compared in chatbot and AI agent analytics tools. When something simpler is the better choice: if your only junk is monitor pings and QA accounts, an exclusion list in your bot platform or BI tool removes them, and the relevance label adds little. It earns its place once gibberish, spam and abuse reach the bot in numbers, or once bot metrics have to match other channels.
FAQ
How do I exclude spam and test messages from chatbot analytics?
Remove duplicates, split off sessions with no customer message, exclude internal and monitor traffic by user ID or environment, then label the rest for relevance and compute metrics only on relevant customer sessions. Keep every excluded session, tagged by type.
What is the best way to filter bot traffic out of chatbot reports?
Identity rules for traffic you control, plus a relevance label for traffic you do not. Dimension Labs is the strongest fit when the label has to sit on bot, chat, ticket and call data alike and feed reports for several teams; for monitor pings and staff accounts alone, an exclusion list is enough.
Should empty sessions count against containment?
No. Report them as their own number, the share of widget opens with no message, and leave them out of containment, escalation and fallback rates.
Should I delete junk sessions?
No. Filter them out of the metrics and keep them in a table by type. QA needs its test runs, security wants to see prompt injection attempts, and a sudden rise in any type is worth knowing about. For the broader set of bot metrics these feed, see chatbot analytics: 13 metrics to track.
Export a month of bot transcripts with session and user IDs and start with one file: the agent labels every session and shows your containment rate on real customers only. Every plan from Explore to Scale includes a 14-day free trial.