We go where we know someone
In Azerbaijan a paid service is found through a person, not through an advertisement. When something goes wrong the customer tells other people rather than the business, and then quietly stops coming. This report joins three sources, 455 face-to-face surveys covering 910 specific service events, 247 in-depth interviews, plus netnography from public sources and official statistics. All three land in the same place, because the part a company can see is a small fraction of what actually happens.
Percentages are computed over 910 service events, with each respondent describing two. The events are not independent, so confidence intervals are computed at respondent level. The sample is a purposive field sample, so these figures describe the participants rather than the population. Our method is written up here →
We asked every respondent to walk through two services they had actually paid for in the previous twelve months. The split below is not opinion, it is events that happened. The human channel, meaning a friend, a colleague, a neighbour or a referral from another professional, comes to 58.2%, while the internet and advertising together come to 29.5%.
One subtlety is worth measuring separately. Among those who arrived through the human channel, 28.7% cannot remember who told them, which means they effectively found it themselves and later recall it as a recommendation. Counting only the ones where the name is remembered, confirmed recommendations account for 41.5% of all events.
A person's word produced 58.2% of the events we recorded, while the internet and advertising together produced 29.5%. The reason sits in the structure of the market. Total advertising revenue across the country runs at AZN 247.1m a year, while households spend AZN 12.885bn on paid services, so advertising amounts to 1.9% of what is being spent. There simply is not enough advertising in the country to direct anyone.
Across 247 interviews we asked why the recommendation was believed. In only 7.7% of answers had the person recommending actually used the place themselves. In 37.2% the trust was in the person rather than the place, and in 30.4% the respondent simply did not want to spend time searching. In the survey, 47.0% say outright that they picked the recommended option because it was easier, not because it was better.
Something went wrong in 41.7% of events that arrived through a person, against 34.3% of those that arrived through a digital channel, and the gap survives holding the category mix constant. The direction is not uniform though. A recommendation performs 25.2 points worse in insurance, 16.6 worse in education and 16.2 worse in healthcare, but 14.9 points better when buying household appliances.
In 51.8% of events that went wrong the customer said nothing to the business. Of those who stayed silent, 91.3% told other people, on average 7.2 of them. A Poisson model shows that holding the severity of the incident constant, not telling the company raises the number of people told by 20%. What is not said inside travels further outside.
A bad experience is told to 8.61 people on average and a good one to 4.55, a ratio of 1.89. One clarification matters. That ratio is about how many people hear about a single incident, not about the total volume of conversation, because by count of events positive word outnumbers negative by roughly three to one.
Among respondents, 45.5% have stopped recommending a place they once recommended. That decision usually follows no single incident: in 38.1% the relationship wore down gradually, and in 21.1% one event was enough. The company therefore loses over months rather than in a day, and receives no warning signal at any point.
On the listing platforms where tradespeople actually live, a review field does not technically exist. The services section of tap.az carries 38,504 listings, and one plumber listing we checked had gathered 27,228 views with neither a score nor a review field on the page. Service-finding apps sit between 17 and 1,657 installs, while Lalafo in the same country has passed 20 million. The gap is not in digital habit, it is specific to the services category.
Not one person in the interviews asked for their money back. That figure is zero. Instead 52.6% wanted the work put right, 28.7% wanted a sincere apology and 17.8% wanted nothing more than to be asked about it afterwards. For roughly a third of unhappy customers a single phone call would have been enough, which makes it the cheapest and least-used tool available.
In health events, 38% of complaints are about price, 31% about how staff behaved and 31% about waiting, while the quality of the treatment itself never comes up in a face-to-face conversation. A patient cannot assess a clinical decision, so they judge everything they can assess instead. The complaint does appear in online reviews, so the difference is not in the judgement but in whether it gets stated out loud.
We put the breaking points recorded in the field next to a coding of 454 public negative reviews collected from 2GIS Baku. Price and mismatched information land almost exactly on top of each other, which says our coding frame also works on an independent corpus. Behaviour and waiting diverge sharply.
On price and mismatched information the two sources almost coincide, which says the coding frame holds up on an independent corpus. The three notes below say why a review page is not a measuring instrument.
What is visible is the tip. Only 18% of those who talk about an experience write anything online. Of 278 health and beauty businesses, 79.5% carry fewer than fifty reviews, with a median of 10 and a median of 3 for beauty salons. Across 99 garages and repair shops the entire review count is 883, median 2. Only 7.2% of businesses that have any reviews score below 4.0, while in the field something went wrong in 38.8% of events.
What is visible is also distorted. Complaints about staff behaviour inflate roughly threefold online, while waiting and lost time shrink threefold. Rudeness makes a story worth writing down. Time lost in a queue is only ever lived through.
The local customer's voice is barely in this channel. Of the 454 negative reviews we collected, 89.9% are in Russian and only 6.2% in Azerbaijani. On top of that, a business replied to just 18 of the 454, and among clinics, shops and salons there are no replies at all.
What follows are not quotations. The interviews were written up from field notes rather than audio recordings, so the texts are not verbatim transcripts. For that reason we publish no sentence as a direct quote, only the typical formulations an interviewer recorded, given in generalised form.
A friend said so, therefore it must be reliable, and they checked nothing themselves.
They do not know whether the person recommending had ever been there, they only heard the word good.
They said nothing to the business, because they expected no result from it and did not want an argument.
They saw no point in saying anything, since they were not going back anyway.
They told the story laughing, and finished it with do not go there.
It happened once, but it was bad enough that once was sufficient.
Had anyone come back and asked, they say they would not have told so many people, one call would have done it.
Not a dry apology but the work put right, and then they say they would believe it.
The figures come from tables published by the State Statistical Committee and are 2024 data. Total annual advertising spend across the country amounts to 1.9% of what households spend on paid services. Against that background, finding a service through a person is not surprising.
To show how ordinary and how unusual our figures are, we set them against measurements taken elsewhere.
State Statistical Committee tables for advertising, paid services, healthcare, transport and demography. Public data from Google Maps, 2GIS, Google Play and the App Store. For international comparison, TARP 1981, Nielsen Global Trust in Advertising, Keller Fay TalkTrack, Engagement Labs and the European Commission’s Consumer Conditions Scoreboard. The full source list and the methodology document are shared on request.
Terms of use. The figures on this page are open under CC BY 4.0. Quote them, chart them and republish them freely. The only condition is that the source is named, meaning Emotix Insights and the address of this page.
In our fieldwork 58.2% of paid service events were found through a person, meaning a friend, a relative, a colleague, a neighbour or a referral from another professional. Online search and advertising together account for 29.5%. The ratio is roughly two to one.
No. Something went wrong in 41.7% of events that came through a person, against 34.3% through a digital channel. On top of that, in only 7.7% of recommendations had the person recommending actually used the place. A recommendation helps when buying an ordinary product and hurts when choosing a specialist.
Usually not. In 51.8% of events that went wrong the customer said nothing to the business. Of those who stayed silent, 91.3% told other people instead, on average 7.2 of them. A statistical model shows that not telling the company raises the number of people told outside by 20%.
No. Of the 278 health and beauty businesses we collected, 79.5% carry fewer than fifty reviews, with a median of 10. Only 7.2% of businesses with any reviews score below 4.0, while in the field something went wrong in 38.8% of events. Online reviews also inflate complaints about staff behaviour threefold and shrink complaints about waiting threefold.
Cheaper than expected. Not one person in the interviews asked for their money back. Instead 52.6% wanted the work put right, 28.7% wanted a sincere apology and 17.8% wanted only to be asked about it afterwards. For roughly a third of unhappy customers a single phone call would have been enough.
Face to face, with 455 surveys and 247 in-depth interviews across 21 locations. Each respondent walked through two services they had actually paid for in the previous twelve months, which gives 910 specific events. Netnography from public sources and official statistics were layered on top. The sample is a purposive field sample, not a probability sample.
This page is the national picture. Sector cuts are prepared separately: where a patient referral breaks, how many cars one bad repair costs a garage, how a shop owner picks a supplier and when they switch. We hold cuts for healthcare, vehicle servicing, trades and beauty. The step after that is a separate wave run with your own customers on the same method, measuring your network loss rather than the market’s. A short call is enough to work out which one fits.