New research: We scored 60 support chatbots with CXS. Overall CXS was +17. Read the research →

ISALO METRICS · VERSION 1.0

CXS™: the Customer Experience Success Index

The Customer Experience Success Index (CXS) measures how often customers get what they came for. It counts complete, correct answers, subtracts failures, and works the same way for support chatbots, AI assistants, other self-service channels, and assisted support.

Version 1.0 · Published November 18, 2026 · Changelog

01

Definition

CXS (Customer Experience Success Index) is a score from −100 to +100 equal to the share of fully successful customer interactions minus the share that end in failure. Partial successes count in the total only.

CXS is based on what happened in each interaction, not on what the customer said about it afterwards. It needs no survey, so it can measure channels where surveys don't reach, such as AI assistants outside your website.

The formula

CXS=100×S−FN

S = full successes · F = failures · N = all interactions

In words: CXS = 100 × (full successes − failures) ÷ all interactions

All interactions count for the formula. Full successes count for the score, failures count against it, and partial successes count on neither side, so a high share of partial answers pulls the score toward zero.

Worked example

200 interactions: 70 full successes, 90 partial successes, 40 failures.

S = 70, F = 40, N = 200

CXS = 100 × (70 − 40) ÷ 200 = +15

The scale

A negative score means customers met more failures than successes.

When you publish a CXS score, we recommend you give its breakdown and the number of interactions alongside it, because the same score can come from different mixes of outcomes. Write every score with its sign.

CXS +15 (35% full · 45% partial · 20% fail, n = 200)

02

The three outcomes

Every interaction is rated, from the customer's point of view, into one of three outcomes.

Full success

The customer gets a complete, correct path to what they asked about, and needs no further step to confirm it. The answer doesn't have to be the one the customer hoped for. "The current version can't do that" is a full success if it is clear and correct.

Partial success

The answer helps but falls short. It is incomplete, too generic for the customer's situation, or leaves them with more work to reach a full answer.

Failure

The customer gets nothing they can use. The answer is wrong, missing, or sends them somewhere without a usable path, for example "contact your account manager" with no way to reach that person.

03

Choosing the scenarios

When you test, rather than rate real conversations, CXS depends on the scenarios you choose. A scenario is a task a customer needs to complete, such as resetting a password.

Which scenarios to include

  • Common requests. Customers should get a great answer to the questions they ask most, such as how to reset a password.
  • Risky requests. Include tasks where a wrong answer could harm the customer. If removing a user also deletes their data, removing a user is a scenario worth testing.
  • What's new. Include new products, features, and help content, so you can test them early.

How many scenarios you need depends on what you support. A team that mostly handles customer service may need only a handful. Complex products need more. Aim to cover at least half of your known contact volume, using what you can see, such as questions to your chatbot and support cases. You can't see what customers ask AI assistants, so your own data is the best guide. If you have to start small, do. A handful of scenarios is better than none.

Write the questions the way customers ask them

Start from one complete question, such as "How do I reset my password on [product]?". Customers rarely write like that, so add variations:

  • an incomplete sentence: "password reset [product]"
  • a rushed customer, with abbreviations: "reset pwd"
  • different words for the same problem: "I can't log in"
  • a newcomer's words: "I'm blocked"

Real transcripts show the words your customers actually use.

Add context where it matters

Some scenarios need a back-story for the tester. For a password reset: does the customer sign in directly or through single sign-on? Are they signed in and want to change their password, or locked out? Are they new to the product, or an IT administrator?

We recommend a couple of variations per scenario for conversations, and 5 to 10 for single question-and-answer tests.

04

Building a rubric

CXS is only as good as the rubric behind it. A rubric turns "did the customer get what they came for?" into a decision two raters can make the same way. Write one rubric per question, before any rating starts, in five steps.

1. Record the question and the customer's goal

State the question as a customer would ask it, and the task they are trying to complete. "How do I reset my password?" is a question. "Get back into my account without contacting anyone" is the goal. Rate against the goal, not the wording. Also record whether the request needs a person, such as a contract or an account change that requires verification. This decides how a handoff is rated.

2. List the required elements for a full success

List everything a complete answer must contain. Keep each element short and checkable. Most questions need two to four. An answer that covers every element correctly is a full success.

Question: "are you SOC 2 compliant, what type, and how do we get your current report?"

Required elements:

  1. Whether the organization has a SOC 2 report
  2. Which type (Type I or Type II)
  3. How a customer or prospect can get it (a trust center, a request form, or a named team, with any conditions such as an NDA)

3. Name the source of truth

For each required element, record where the correct answer is published: a help article, a pricing page, a trust center, or product documentation, with the link and the date you checked it. Raters check answers against these sources, not against what they believe to be true. If an element has no published source, say so in the rubric. Answers on that element are then rated on clarity and a usable next step, not on accuracy.

4. Define partial success and failure for this question

Decide in advance where this question's middle ground is. Write down the common ways an answer can fall short, and which outcome each one gets.

Building a rubric
OutcomeWhat it looks like for the SOC 2 question
Full successAll three elements, correct against the trust center
Partial successConfirms SOC 2 but not the type, or gives the type but no way to get the report, or says "I can't confirm" and names the team to ask
FailureWrong type, a route that leads to a different document, a general security statement instead of an answer, or an email request before any answer

Any answer that contradicts the source of truth is a failure, even if it covers every element.

5. Test the rubric before you use it

Have two people rate the same 10 to 20 interactions with the rubric, separately. Where they disagree, the rubric is unclear: tighten the wording, refresh the rubric, and rate again. A rubric is ready when the two raters agree on most interactions. Record the agreement rate, and publish it with the score.

Rules that apply to every rubric

The rubric for each question decides what counts as complete. These rules apply to every rubric, whatever the question:

  • Contradicted claims are failures. An answer that conflicts with the organization's own published source of truth is a failure, however well it is written. A claim that can't be confirmed either way is not a failure on that basis alone. It is rated on everything else in the answer.
  • Handoffs. Passing the customer to a person counts as a full success when the answer gives a usable path: who to contact and how, with no further search. A handoff with no usable path, such as "please contact support", is a failure.
  • "I can't confirm". An honest statement that the assistant can't confirm the answer, together with a specific next step, is a partial success. Without a next step, it is a failure.
  • Contact gates. Asking for an email address or a meeting before answering a general question, and not answering without it, is a failure. Routing a customer to a named contact for an action that genuinely needs one, such as a contract change, can be a full success under the handoff rule.
  • Repetition. Repeating an unhelpful answer after the customer has pointed out the problem is a failure.
  • Effort. In self-service, a complete answer counts as a full success however many turns it took. Effort, such as the number of turns, is worth reporting separately, but it doesn't change the outcome.
  • Handoffs to a person. A handoff counts as a full success when the request needs a person and the customer gets a usable path to one, such as a live transfer, a form, or a direct contact. When the channel could have answered the request itself, a handoff counts as a partial success.
  • Fixed menus. Some chatbots offer only buttons or menu options and don't accept typed questions. If an option matches the customer's request and leads to a complete, correct answer, rate the interaction as you would any other. If no option matches the request, the interaction is a failure, because the customer has nothing they can use.

Keep rubrics stable

A change to a rubric can change the scores, even if nothing else changed. Give each rubric a version and a date. When you change one, re-rate the earlier interactions or report the scores before and after separately. When you compare scores over time, or across channels, use the same version of the rubric.

05

Where CXS applies

CXS version 1.0 covers two kinds of support: self-service and assisted support. The formula is the same for both, so you can compare channels on the same questions, or track one channel over time as content and configuration change.

Self-service

Self-service includes support chatbots, AI assistants on a website or in a product, and third-party AI assistants that customers ask about a company. There are two ways to collect the interactions:

  • Observation. Rate real customer conversations, all of them or a sample. This works for channels where you have the transcripts, such as your own chatbot.
  • Testing. Ask a fixed set of realistic customer questions and rate the answers. This is the only option for third-party AI assistants, where you can't see your customers' conversations. Use questions phrased the way customers ask them, not wording copied from your knowledge base.

Self-service has no time window, because self-service interactions end, successfully or not, within minutes.

Assisted support

Assisted support covers tickets, email, chat with an agent, and phone. In assisted channels, time is part of the outcome. A case that takes weeks to resolve isn't a full success, even if the final answer is right.

Depending on the complexity of the products you support, the standard window ranges from 24 hours to one week from the customer's first contact. Choose the window before you start rating, report it with the score, and compare only scores that use the same window. Rate each interaction by where it stands at the end of the window:

  • Full success: the customer has a complete, correct answer, the issue is fixed, or the customer has a workaround they can use within the window.
  • Partial success: the customer has a useful but incomplete answer, a fix committed for a date after the window, or a commitment to assess the issue within the window, without a fix yet.
  • Failure: the customer has nothing they can use, including a promised future fix with no committed date.

06

Limits

  • CXS depends on the question set. A score from easy questions will be higher than one from hard questions. Compare scores only when they come from the same or comparable question sets, and use a wide range of real customer questions.
  • CXS is one number. It summarizes outcomes. It doesn't explain them. Read it with the breakdown, the transcripts behind it, and your other measures.
  • Rating needs a rubric. Two raters using the same rubric should reach the same outcome on most interactions. We publish our rubrics and an agreement check with our research.
  • Small samples are noisy. With 10 interactions, one interaction moves the score by 10 points. We recommend at least 50 interactions before comparing two scores, and more for small differences.

07

Frequently asked questions

Why don't partial answers count as successes?

A partial answer leaves the customer with more work, and often leads to a second contact. So it doesn't count toward success. It still counts in the total, so a channel full of partial answers scores close to zero.

Can a gated answer be a full success?

Yes, when the gate is part of a usable path. If cancelling requires the account manager and the answer gives their contact details, that is a full success. Withholding a general answer, such as how to reset a password, until the customer shares an email address is a failure.

What about made-up answers?

An answer that contradicts the organization's own published information is a failure. An answer we can't confirm either way is rated on the rest of its content.

Two organizations have the same score. Did they perform the same?

Not necessarily. One may have many successes and many failures, the other mostly partial answers. That's why CXS is always reported with its breakdown.

Does CXS replace satisfaction surveys or NPS?

No. Surveys measure how customers feel. CXS measures whether they got what they came for. In self-service, surveys are a weak measure: only a small share of customers answer them, so the results reflect the few who chose to respond, not everyone who tried. CXS rates every interaction, or a fair sample, so it doesn't depend on who responds. It also reaches channels surveys can't, such as AI assistants outside your website.

Can I calculate CXS myself?

Yes. Rate your interactions into the three outcomes against a written rubric, then apply the formula. Cite this page as the definition.

08

Changelog

Changelog
VersionDateChange
1.0November 18, 2026First published version. Covers self-service and assisted support.

09

How to cite

This definition is licensed under Creative Commons Attribution 4.0 (CC BY 4.0). You may share and reuse it with attribution to isalo.

isalo. CXS: the Customer Experience Success Index, version 1.0. November 18, 2026. https://isalo.ai/metrics/cxs

You may use the name CXS to refer to this metric when it is calculated as defined on this 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.

See how your customers' questions get answered