CXS: the metric you’ve been waiting for to measure customers’ experience
We’ve tried to measure the quality of customers’ experience for a long time. Traditional metrics–CES, CSAT, NPS–rely on surveys, which are highly susceptible to bias introduced by non-random sampling and low return rates. Customer sentiment can be useful, but it only measures tone, not whether the answer given is actually correct. Case deflection is often used for self-service but we know it’s a crude measure. Also, as more support interactions occur DIY-style through third-party LLMs, it’s becoming less and less relevant. We need a better approach.
The Customer Experience Success Index (CXSTM) is a simple, powerful, comprehensive method of gauging the quality of the customer experience. It relies on data, not surveys, allows benchmarking and time comparisons across all support channels including LLMs, and can be understood by everyone from support agents to board members.
How it Works
Computing the CXS index relies on analyzing the outcome of support interactions from the customer’s perspective.
- Some interactions are fully successful: the customer gets a full, correct path to completing the task they are inquiring about. Note that the answer may not be the one the customer is hoping for (as in, “The current version of the product cannot accomplish what you’re trying to do.”) but the answer is clear and accurate. The customer does not need to take any more steps to confirm what to do.
- Some interactions fail: the answer is incorrect, not available, or requires a separate, onerous effort. For instance, the customer is told to contact their account manager, but not told how to do it, or who that person is.
- Some interactions lie somewhere in the middle. The answer is helpful but incomplete, or too generic to accomplish the task at hand. The customer will need to expend more effort to reach a full answer.
To compute CXS, we simply take the share of conversations that were fully successful, and subtract the share that were unsuccessful.
The Formula
Specifically, the formula reads as follows.
100 * ( full successes – failures ) / interactions
For instance, if we rate 200 interactions and find 70 fully successful answers, 40 failures, and 90 partially successful answers, the CXS index would be 100 * (70-40) / 200 = 15.
Because of the way the index is constructed, the same number can represent a different distribution. For instance, with the same 200 interactions, if we find 30 fully successful answers and 0 failing entirely (with another 170 partially-successful answers), the calculation would also yield 15 = 100 * (30-0) / 200.
Note that we include all the attempts (successes, failures, and partial successes) in the denominator. The numerator only considers full successes and failures.
The Range
The scale ranges from -100 to +100, from all-failures to all-successes.
- A score near the top means customers consistently get what they came for.
- A score near zero means any (full) wins are cancelled out by misses.
- A negative score means more failures than successes–and customers encountering more frustration than value.
Based on current benchmarks, many support organizations cluster in the low positive range when it comes to chatbot support. See the isalo survey of 600 interactions with 60 support chatbots.
CXS for Self-Service
CXS is a natural fit for self-service because it does not rely on surveys, which are notoriously unreliable for interactions that occur on your website, and infeasible for the growing tide of DIY interactions via third-party LLMs.
To compute CXS for a support chatbot, we sort the transcripts of conversations, either real customer conversations (all of them, or a sample) or test conversations, into the three categories of full successes, failures, or partial successes. Rating and scoring can be done manually or through an automated tool or service like isalo.
Interactions through third-party LLMs cannot be inspected through customer transcripts. Instead, we need to conduct a test, using well-designed customer-grade questions (not canned questions that perfectly mimic the knowledge base language) and rate the answers.
Because CXS uses the same formula for every kind of support interaction, we can use it to compare trained chatbots with LLM assistants, rated on the same questions. We want to believe that the trained chatbots will win, but they won’t always… And by running time studies you can tell whether improvements in the knowledge base or other aspects of self-service are resulting in improved CXS.
CXS for Traditional Support
CXS can be used for traditional support as well. Here, we have all the relevant data to compute CXS, in the form of case notes. Traditional support yields lots of full successes since customers speak up when they don’t get a proper answer, and escalation processes can ensure that the vast majority of questions do get answered. Still, it’s useful to reflect on what does not get fully answered: can a bug that’s logged but will linger on for months or years qualify as a “full success”? We will tackle creating a solid rubric in an upcoming post.
Separately, there’s the issue of how long it takes to get an answer. We don’t impose a time limit for self-service because self-service interactions complete, successfully or not, within a very short timeframe, but for traditional support we all know that some cases go on for days or weeks, and even if they do get a full answer in the end, it does not feel like success, at least from the customer’s perspective.
We believe there needs to be a time-bounded definition of CXS for traditional support. For relatively simple support, we can use a 24-hour target (in other words, did the customer get a fully-successful answer, a partial answer, or nothing useful within one day). For complex products, a one-week target is probably right. Remember that we are not looking merely for case closure within a day or a week. We are looking for a fully qualified answer being delivered in that timeframe.
If you compare CXS for traditional support to CXS for self-service, traditional support will likely score higher if you eschew time boundaries. With the imposition of a time boundary, the difference may not be so great.
Why CXS is a Superior Metric
CXS beats other success metrics such as satisfaction surveys, case deflection, portal logins, even escalation rate because of the way it’s architected. There are six advantages built into CXS.
- Evidence-based: no survey, no opinions: the interactions speak for themselves.
- Balanced: failures are subtracted from wins, so easy successes can’t hide bad experiences, and half-answers don’t muddy the score.
- Channel-agnostic: the same formula works for support chatbots, LLM assistants, and traditional support so you can benchmark them against each other.
- Granular: you can measure CXS for a particular product or for a specific scenario (e.g., password reset), allowing you to zero in on whatever is most important to the business.
- Responsive: the index moves from week to week as you train team members, update the knowledge base, retrain the chatbot, or launch new features, allowing for adjustments.
- Easy to communicate: one number, one transparent formula that makes sense to anyone from support agents to board members.
We believe that the time has come to replace traditional, misleading, highly-gameable metrics with CXS. Will you give it a try?
The definition of CXS is licensed under Creative Commons Attribution 4.0 (CC BY 4.0). You may share and reuse it with attribution to CXS: the Customer Experience Success Index, version 1.0. November 18, 2026.You may use the name CXS to refer to this metric when it is calculated as defined on that page. Don’t use CXS in the name of a product or service, or for scores calculated a different way.
CXS is a trademark of A6 Group LLC.
