Ask a vendor for a deflection rate and you will usually get a number in the seventies or eighties. Ask how it was calculated and the answer is almost always the same: conversations the assistant handled, divided by total conversations. Nobody is lying. The arithmetic is just measuring something other than what the word means.
Where the number goes wrong
A deflected ticket is one that would have been filed and was not. A bot conversation is one that happened. Those are different sets, and the gap between them is large in a direction that flatters the tool.
- Most people who open a chat widget were never going to write in. They would have skimmed the page, or left.
- Some conversations are one message long and get no useful answer. They still count.
- A single frustrated visitor can produce four conversations before giving up, and every one of them counts as a success.
- Traffic that arrives because you added a chat widget is new volume, not displaced volume.
The honest version
The measurement that survives scrutiny is dull and slow, which is why it is rare. Take ticket volume per week, normalised against something that moves with your business — active customers, sessions, orders. Record it for several weeks before you launch anything. Then launch, and leave it alone.
Give it at least a month. Weekly volume is noisy, and the first fortnight after any launch is distorted by the novelty of a new widget on the page. Then compare normalised ticket volume before against after. That difference is displacement. It is the only number in this whole exercise that means what it says.
| What you count | What it tells you | Can it be gamed? |
|---|---|---|
| Bot conversations handled | How many people opened the widget | Yes — trivially, by promoting the widget |
| Answer rate | How often it chose to answer | Yes — by answering more carelessly |
| Normalised ticket volume, before vs after | Actual displacement | No |
| Satisfaction over the same window | Whether displacement cost you anything | No |
Falling tickets is not automatically good
Volume drops for two different reasons that look identical in a dashboard. Either the answer got easier to get, or contacting you got harder. Both show up as a smaller number.
relative to baseline
Shape, not data: the left pair is deflection, the right pair is attrition, and ticket count alone cannot tell them apart.
This is why the two numbers only mean anything together. Tickets down with satisfaction flat is the result you wanted. Tickets down with satisfaction down is customers giving up, and it will arrive later as churn, where nobody will connect it to the support project that caused it.
“A support tool that makes it harder to reach you will always beat one that answers well, on the only metric most teams report.”
What to report instead
- Normalised ticket volume, before and after, over at least a month each side.
- Satisfaction across the same window, from the same instrument as before.
- The refusal list, grouped — what the assistant could not answer, ranked by how often it came up.
- How many escalations reached a person, and how quickly.
The third one turns this from a report into a plan. Every refusal is a question your content does not answer yet; written up, it stops arriving. That loop is the part of support automation that compounds, and it is invisible to a deflection percentage.
We do not publish a deflection figure for Matter Chat, and you should treat one from anybody else as an upper bound rather than an estimate. What we can tell you is what to measure and how long to wait — the guide on measuring deflection is the long version of this post, with the baseline arithmetic written out.
