the power of eyezon R&D in content analytics: how to get to know your online shoppers better and make your online sales script work wonders.

Who we are, what eyezon analytics does, and how

‍

Who

‍

We are eyezon, and we create on-demand live video sales. By clicking the eyezon widget on your favorite retailer’s website, you can join a live stream to see products with your own eyes, ask shop assistants questions, and receive personalized recommendations.

‍

Why

‍

Understanding what your customer wants is not something you can afford to ignore. In fact, when it comes to online sales, it’s as vital as the air we breathe. Effective streamers always try hard during personal consultations to understand what their customers need so that they can give them as much information as possible and provide them with the best possible service. They help customers find the product perfectly suited to them. Here at eyezon analytics, we strive to improve customers’ lives and make people happier. We offer exhaustive advice so that the customer is certain they really need the item and feel great about purchasing it. Not a bad job, wouldn’t you agree?

‍

Ultimately, if people keep asking the same standard questions about a product, maybe you should save them time and modify the product page on your website. But here’s the thing – you can also change the sales script by homing in on the toughest questions early on in your conversation.

‍

How

‍

eyezon analytics takes anonymized recordings of streams (the streamers’ and customers’ voices) and translates the sound into text using machine learning tools. Our Natural Language Processing tools then kick in, and we study what the customer said – what specific questions they asked, and about what goods in particular. The data is collected, processed, and passed on to the producer or distributor of the product, complete with useful conclusions and accumulated insights.

‍

The result is that the brand is able to work in tandem with streamers, testing the best sales scripts. Another way to use this kind of data is as a tool for subtly configuring advertising. There are many different ways to use our product, but the goal is always the same – to help customers make a choice, solve their problems, and feel a little happier.

‍

eyezon analytics has set up its own system of dictionaries – collections of words and phrases that are identified from the recordings of the streams. We assign a ‘mentioned_event’ to each of these dictionaries. We have dozens of these, for virtually every scenario:

• mentioned question asked – someone asked a question;

• mentioned ordering – someone talked about setting up an order;

• mentioned concerns – someone used words related to doubts or worries when choosing a product;

• mentioned discount – someone mentioned a discount on a good;

• mentioned delivery – someone mentioned deliveries;

• mentioned intent to buy – someone indicated their desire to set up an order or purchase a good;

• mentioned problems – someone mentioned any kind of problem (with an order or when using a product);

• mentioned guarantee – someone mentioned the guarantee on a product;

• mentioned aggression – someone showed signs of aggression;

‍

We don’t just assign these ‘mentioned’ tags to communications and speech; we also count their frequency: how many questions were asked, how often discounts were mentioned, how often people asked about delivery, as well as what people asked, specifically.

‍

The first reason for this is that it makes it possible to find specific communications by ‘mentioned’ events to review exactly what was said. Secondly, we can correlate the number of occurrences of each kind of event with the number of clicks, requests, and streams – and that’s how you get conversions. Thanks to this, we can monitor sellers and make sure that a store’s live consultation services are always operating at peak performance, that conversions are stable, and that customers are satisfied.

‍

The most interesting thing about this kind of work is that it allows us to take a closer look at customer insights.

‍

In one of the tests we conducted, we thought that people would ask most often about discount sand special offers, or about delivery (classic), but it turned out that shelf life was actually customers’ top question.

If we consider the most frequent phrases, then TOP-1 will be “shelf life”, and TOP-10 “production date”.

Of course, all of this is what you would call R&D. Later on, we’ll be able to tighten up the required ML tools, supplementing or replacing the rules-based approach when searching for words. It all depends on a mixture of imagination, technical capabilities, and more than a little expertise.

First case, first results, main goals

‍

First case

‍

How do you make an analytical service accessible and useful? Simple: you approach eyezon’s clients and ask them what they want. And then – make those goals a reality.

‍

So that’s what we did. Our trailblazer was a world-famous cosmetics brand. We took all the voice recordings of their customers, determined what they said, and began to work wonders.

‍

And the ultimate result of our efforts was this: automatically updating dashboards where you can see:

‍

1) Current conversions: how many customers came from live streams, and how many sales resulted. We have conversions for every stage of interaction with the eyezon widget, beginning with clicking on the eyezon button and ending with making ‘mentioned event’ or purchase.

‍

‍

2) Flexible pivot tables: this is where you select the columns and rows yourself to create the table you need. You can use pivot tables to find out:

‍

• what products and categories sell the most, which ones get the most questions (and what the questions are), and what products experience the most customer activity.

‍

• how well your streamers are working: who customers talk most willingly to, who’s the best at selling a product, which of your streamers is a little lazy, and which streamer talks too little during consultations.

‍

• your stats for ‘mentioned events’:

‍

3) You can read the identified speech from both streamers and customers by selecting the various tags among the ‘mention edevents’.

‍

4) See what competitors were mentioned in messages, and how often. You can also do a ‘deep dive’ into these messages and find out what customers like and don’t like specifically about your product compared to your competitor’s items.

‍

5) You can analyse N-grams. This is where we count the most popular expressions found in customers’ speech. This can sometimes lead to completely unexpected results.

‍

And lots of other features! eyezon analytics can also adapt some of them to company-specific requests – we’re always flexible and open to new ideas.

‍

‍First results

‍

The dashboards we described above provide a wide range of possibilities for testing. The option is always available to you to give different scripts to different streamers and compare the conversion rates. Experiments, together with careful development of the quality of your live streams, will increase your audience’s loyalty, improve conversions, and help customers make their purchase decisions more effectively.

‍

We’ve also designed our dashboards to collect insights. That’s how we found out live stream customers are a little different from customers in a store. Mathematics is always hard work, but we carried out some thorough analytical work and came up with a formula for the ‘average’ stream:

‍

STREAM (SHOPPER REQUEST) = A+ B + C, where

‍

A – questions that a customer usually asks in a store

‍

B – questions that arise because the customer is not physically at the store, and therefore isn’t able to judge the colour, texture, material, or size of the good in question (and the product page is unable to provide all the required information)

‍

C – ‘candid questions’ that arise because the customer is an anonymous participant in the stream, and can therefore share a personal issue or ask the streamer to give their point of view about a problem, situation, or choice

‍

This is why when we study customers, we often see something like this:

‍

Therefore, ideal script should reduce the category of B questions and increase the category of A and C questions. This is necessary so that the buyer spends as little time as possible on typical questions and asks more complex questions that would allow her to choose the right product.

‍

Main goals

‍

In the future, eyezon analytics aims to add part of the dashboards to the eyezon personal account as an additional add-on, so that using our widget will be as easy as possible for our clients and be of maximum use to them.

‍

Right now, we’re undertaking bold experiments with metrics for our analytics, introducing analysis methods, and working on the development of our R&D.

‍

eyezon analytics is always open to individual cooperation for this kind of R&D. If you have any crazy (and not so crazy) ideas for content analysis with us, or if you just want to replicate our ‘first case’ experience, let us know (for example, write to us at p.gusev@witheyezon.com). When all is said and done, the magic of interaction between streamers and live stream customers needs to continue, and we know better than anyone where and how it begins.

‍

Bonus

For the most curious readers, we have prepared a brief scheme of how our analytics works from the technical side.

Data pipeline 

‍

‍

Pipeline features

‍

A bit more about consumers

‍

Each consumer firstly downloads stream from S3 by its name, converts .mp4 to .wav and starts speech recognition. If result is successful, text will be appended to Postgres.

‍

‍

We can add as many consumers as we want by adding new VDS to boost number of transcribed streams.

A bit more about queues

‍

To change speech-to-text language we can create a new Queue where streams with specific language will be added. This queue will be listened by new Consumers which have exact ML model language.

‍

‍