Measuring whether multilingual marketing is working

Multilingual work is the hardest marketing to measure because its best channel is invisible. A recommendation moves through a family group chat, six people look you up by name, and every analytics tool you own records it as direct traffic.

So you measure what you can count and give it enough time to mean something.

Why can I not just run a test?

Because your traffic cannot support one, and the number everybody quotes to explain that is itself made up.

Start with the debunk, because it will come up in a meeting. You will be told that valid A/B testing needs about 10,000 monthly visitors. That figure has a trace and it is not research. It appears as reference 29 in Kohavi, Deng, Longbotham and Xu's "Seven Rules of Thumb for Web Site Experimenters," presented at KDD 2014, where the four Microsoft and LinkedIn authors write: "Our advice in previous articles is that you need 'thousands' of users in an experiment; Neil Patel suggests 10,000 monthly visitors, but the guidance should be refined to the metrics of interest." Reference 29 resolves to a QuickSprout blog post dated 14 January 2013. A blog post became an industry threshold by repetition, and the paper that everyone cites for it is actually correcting it.

The correction is the useful part. The same paper: "Formulas for minimum sample size given the metric's variance and sensitivity (the amount of change one wants to detect) provide one lower bound." Sample size is a function of how noisy your metric is and how small a change you want to catch. It is not a fixed traffic number.

Now derive yours, which is the only version that means anything. Say your site gets 620 sessions a month and converts at 3%. That is about 19 leads a month, or 9 per variant if you split the traffic. Over a full year you accumulate 7,440 sessions, so 3,720 per variant.

Compare that to the paper's own worked table for Bing metrics, which shows how metric-specific the answer is. A capped revenue-per-user metric with a skewness of 5.2 needed 9,700 users per variant to detect a 10.5% change at 80% power. Sessions per user, the least demanding row, needed 4,700 per variant for a 5.4% change. Your entire year of traffic, split in half, reaches 3,720. That is 38%derived of the least demanding requirement in the table, after twelve months, to detect an effect far larger than anything you are likely to produce.

And the paper's Rule 2 closes the door: at Bing, experiments that succeed "improve key metrics by 0.1% to 1.0%, once diluted to overall impact." You are not going to see that in nine leads.

This is web-scale desktop data from search engines and large retailers, not a local business dataset, and it should never be presented as one. What transfers is the arithmetic and the epistemics.

So do not test. Count, baseline, and compare over quarters. Less satisfying, and the only method available at your volume.

The baseline matters more than the counting. Write down where the numbers are today, before you change anything, or you will have no way to know what moved.

What are the five numbers?

Small enough that one person can pull them in half an hour.

1. Reviews by language, last 90 days. The best single indicator. It reflects who showed up, who felt understood, and whether your asking process reaches everyone. Count them, do not estimate.

2. Calls by language. Log at the point of answer for two weeks a quarter. The fastest way to find demand you are not serving.

3. Non-English search terms. Search Console and your profile's search terms report. Watch whether they appear at all, then whether they grow.

4. How new customers say they found you. One question, asked consistently, logged. The only instrument that catches the invisible channel.

5. Revenue, total. Not attributed. If everything above improves and revenue does not, the work is not reaching the part of the business that matters.

That is the whole dashboard. Adding more numbers makes it less likely anyone maintains it.

What is actually measurable, and what is not?

Know the shape of the hole before you go looking in it.

Your Google Business Profile performance report publishes interactions, searches and search terms, views, direction requests, calls, website clicks, messages and bookings. It publishes no rank and no position, anywhere. Neither does Search Console. Any tool showing you a local position inside your profile computed it somewhere else, from its own grid crawls.

Google's generative AI performance reports in Search Console, launched 3 June 2026, are similarly bounded. They give impressions, pages, countries, devices and dates. They explicitly do not give clicks, click-through rate, position, or the prompt the user typed, and they do not separate AI Mode from AI Overviews. If you were planning to measure whether your Korean pages are showing up in AI answers, that is the ceiling.

There is a harder problem underneath. Clicks arriving from AI Overviews and AI Mode land in your analytics as ordinary Google organic traffic, with no referral parameter and no separate channel. Google's own documentation confirms sites appearing in AI features are included in overall search traffic under the Web search type. The largest AI surface in local search is fundamentally unattributable, so treat any vendor dashboard claiming to break it out with suspicion.

One more thing worth knowing before you panic at a chart. Sterling Sky, using Jepto data compiled across 179 profiles at 34 US law firms over two years, documents clicks-to-call from Google Business Profiles declining even for profiles whose rankings did not move. Website clicks did not fall the same way. If your calls are down and your rankings are flat, that pattern is documented and it is not necessarily your language work failing.

How long before I should expect to see anything?

Ninety days for the leading indicators, a year for revenue.

Reviews by language should move first, because the asking process changes the moment you decide to change it. If ninety days pass with no movement there, your process did not actually change, whatever the plan said.

Search terms move slower, because indexing takes time and non-English query volume is smaller and noisier.

Revenue moves last and is hardest to attribute, which is why total revenue over a year is more trustworthy than any monthly attribution report.

Compare year over year if your business is seasonal. A restaurant comparing October to July is measuring October.

What should I not measure?

Three things that look like progress and are not.

Page views on translated pages. A translated page can collect traffic without producing a customer. A page with low views that converts well is doing its job.

Star rating. It moves slowly, it is dominated by history, and it says almost nothing about this quarter. Count new reviews instead.

Social engagement in the second language. Likes on a post in Korean are not evidence of Korean-speaking customers. They are evidence of Korean speakers with phones.

And do not measure by making the number easier to hit. If you are tempted to ask only the customers you expect to be happy, that is review gating, and it makes you rank worse rather than better.

Say the compliance part precisely, because the loose version circulates and it is wrong. Google bans gating outright: its Maps user generated content policy prohibits merchants from "discourag[ing] or prohibit[ing] negative reviews, or selectively solicit[ing] positive reviews from customers." The Federal Trade Commission's 2024 rule on consumer reviews and testimonials, 16 CFR Part 465, does not name gating in those words. It reaches the adjacent conduct, including incentives conditioned expressly or by implication on a particular sentiment at section 465.4, and misrepresenting that displayed reviews represent all reviews submitted when you have suppressed them by rating at section 465.7(b). So the accurate sentence is that Google bans gating and the FTC bans the conduct next to it. Anyone telling a client the FTC bans gating is overstating a real rule, which is a bad habit to build a program on.

What do I do when the numbers do not move?

Work backward through the chain, because it usually breaks at a specific link.

The chain is: someone searches, finds you, decides, contacts you, is served, and writes about it. Check each step in order.

Did non-English search terms appear at all? If not, the coverage work did not happen or is not indexed.

Did terms appear but no calls follow? People found you and chose someone else. That is a page or profile problem, often a website problem in disguise, since a business still needs a real website even with a strong profile to answer what the profile cannot.

Did calls come but not convert? That is the phone, and it is the most common break.

Did customers come but leave no reviews? That is the asking process, usually because nobody asked in their language.

Working the chain in order beats changing five things at once and learning nothing. It is also the reason to change one thing per quarter. With nine leads a variant, sequencing is the only experimental control you have.

Does this apply to community work too?

Partly, and you should hold community work to a different standard.

Association memberships, event presence, and sponsorships do not respond to quarterly measurement. Track inputs instead: events attended, memberships held, referral sources named. Judge them over years. Community marketing is not advertising, and a monthly return target on it is how it gets cancelled right before it works.

The website and profile side is different. That work should move in a quarter, and if it does not, something is broken rather than slow.

If you are still deciding how far to take the language work, the measurement question and the scope question are the same one, and whether your website needs a second language is answered with these same five numbers.

What to do this week

Count your reviews by language for the last ninety days and write the number down with today's date. That is your baseline.

Pull your Search Console query list and scan for non-English terms. Note whether there are any.

Start logging the language of incoming calls at the point of answer. Two weeks is enough.

Add one question at the counter or the intake form: how did you hear about us. Log it where you will find it later.

Run the sample-size arithmetic above on your own traffic once, so that the next time someone proposes a test you can answer with a number instead of a feeling.

Then put a recurring reminder ninety days out to pull the same five numbers again. The reminder is the system. Everything else is arithmetic.

The underlying decisions about what to build and where to spend follow from these numbers, which is the practical version of building for a bilingual customer base.

Questions about what your numbers are telling you?

Email me at eric@seod.com with your business name and I will pull your review language split and your public profile signals myself, then send you a one-page read of what is working and what is not moving. It is the same five-number check I would run on a client.

I answer these personally. If the numbers say your language work is fine and something else is the constraint, that is the answer you will get.

Or keep reading more on multilingual and community marketing.