A business added three languages two years ago. Somebody asks at a budget review whether they are working.
Nobody can answer. Traffic is reported as one figure, revenue is reported by product rather than by market, and the cost of maintaining the translations is spread across a content budget, a support budget and an agency invoice.
Two of the three are performing well. The third produces almost nothing, consumes disproportionate support effort and has not been updated in fourteen months, and it has been invisible for the whole two years.
Start by separating the reporting
Traffic, enquiries, conversion and revenue by language version. Without this separation there is no question to answer, and most analytics setups do not do it by default.
Measure conversion, not traffic
A language version attracting visitors who do not convert is a cost rather than a market. Traffic is the easy number and conversion is the one that decides anything.
Measure the enquiry response time per language
Slow responses in one language usually explain a poor conversion rate there, and it is a fixable operational problem rather than a market one.
Measure the support effort per customer
Where serving customers in one language takes considerably more effort, that belongs in the assessment alongside the revenue.
Count the ongoing content cost
Every price change, policy update and new page multiplied by the number of languages. This is the recurring cost that nobody attributes anywhere.
Count the translation maintenance debt
How much of each language version is out of date. A language that is two years stale is producing a worse impression than no version at all.
Compare against the counterfactual
What that market produced before the language was added. Some markets buy anyway in another language, and the uplift is what the investment bought.
Look at the search performance separately
A language version that does not rank is invisible regardless of how good it is, which is the ground in search visibility in other languages.
Look at where the customers actually are
A language version frequently serves several countries with different conversion rates, and the aggregate hides which of them is working.
Set a review point and a decision
Twelve or eighteen months, with an agreed measure and a willingness to withdraw. Without a decision point, an underperforming language persists indefinitely.
Understand what withdrawing costs
Removing a language version is visible to existing customers in that market and should be handled deliberately rather than quietly, which is part of what makes the original decision serious, as set out in deciding which languages to support first.
Consider improving before withdrawing
Most underperforming language versions are under served rather than unwanted, and fixing response times frequently changes the picture entirely.
Watch the abandonment points per language
Where visitors in one language leave at a particular step, the problem is usually something specific to that version such as a payment method, a form that rejects local addresses, or a delivery option that does not exist there. These are fixable and invisible in a blended funnel.
Watch the enquiry mix by language
Where one market generates disproportionate enquiries about the same thing, the content in that language has a specific gap, as in international candidates and multilingual enquiries.
Measure it the way you measure anything else
Cost in, revenue out, with a stated method, which is the discipline in how to measure whether your chatbot is working. Run the arithmetic on the chatbot ROI calculator, see pricing, and the sector view sits under small business.