Millions of Customer Reviews, One Model, No Problem
Patricia Osorio is one of the co-founders of Birdie.ai, a consumer insights startup leveraging AI to analyze customer conversations and report insights to consumer brands. I recently had the chance to speak with her about the origin story of Birdie and how it’s providing consumer brands with market intelligence. This interview has been edited for content and clarity.
Hi Patricia, thanks for taking the time to chat. To start can you help me understand what Birdie is doing today?
Birdie reads millions of conversations from consumers on the Internet. So as long as a conversation is published or stored somewhere, we can capture, understand, and process that conversation. From these conversations, we can learn what consumers are thinking and experiencing when buying and using products. We can then give that feedback to our CPG clients so that they can improve how they market their goods or services to them along with the products themselves.
That’s interesting, can you tell me a bit about how Birdie was started?
We started Birdie about two and a half years ago when we started noticing that with the world of eCommerce and more people purchasing products online, there was a lot of content there that could be analyzed. As we started to learn more about the market, we discovered that 55% of product searches start on Amazon but consumers generally don’t buy a product before they’ve read at least a few reviews. We noticed the importance of this behavior and we call it the validation economy — 94% of consumers pursue this type of information and read reviews before making a buying decision. And we realized that brands were capturing review data but still couldn’t systematically learn from the reviews because there was no tool available to help them. There are companies that do social listening to understand what’s trending and what people are talking about, but none of them goes really deep in analyzing that information which is what Birdie is doing. Our product seeks to understand how consumers view product characteristics and aspects of the buying experience. Understanding this experience is a major pain point for marketing executives, marketing teams interested in consumer insights spend up to 80% of their time just trying to make sense of the amount of data that they’re capturing and 85% of marketing executives don’t believe they have the right tools. As a result, 55% of review data becomes dark data [data that’s lost inside an organization]. Birdie looks at all of that data using our AI engine and turns that data into relevant insights for specific products and the buying experience. Our system is set up so that it can deliver consumer insights to marketing teams, they can quickly look at the data, and then take action.
How big is Birdie today?
We have 25 employees, most of them in Brazil.
How much money have you raised and what markets are you focused on?
We’ve raised $1.6 million to date. When we started the company, we were focused on Brazil and partnered with Samsung and P&G in that market, but are now only focused on the United States as 44% of spending on marketing research is from the US — when we have more market share we will go back to working in other countries.
Without revealing your secret sauce, can you talk a bit about what’s happening behind the scenes?
I’ll try to use an analogy. If you go to Google and run a search, the results are all based on keywords and queries, and there are millions of results that can be directly related to what you’re searching for or not at all. The keywords in this instance are not enough to understand the context. After a search, we as users need to go through the results until we find the result we’re looking for. Birdie doesn’t do that because we bring to our clients what stands out instead of them needing to go through our results. To make that happen there are a few components. The first being our web crawlers, we can integrate with tools that are available via APIs, but we also have our own crawlers that can access sources like eCommerce platforms. [A web crawler is a software tool that goes through the internet looking for specific terms or content] So our tool goes to platforms like Amazon or Reddit and reads all the reviews, all the questions and answers, and all the available product descriptions. We then bring all that information into a database and start linking it together using a very comprehensive dictionary where we input the components of a buying experience for a user and customize it further to specific categories. Let’s say if a customer is buying a refrigerator, they’re going to want to understand its cooling capability, how good is its icemaker, etc. So, we became specialists in all the categories that we’re working on. And then everything that flows into our database is connected to that dictionary which is how we connect different data sets. For a more complete breakdown of how Birdie’s system works check out this blog they posted last week: https://www.birdie.ai/post/text-analytics-applied-to-consumer-insights
Very interesting. I think you’ve been hinting at this, what are the major differences between Birdie’s solution and other services on the market today?
So, most of the companies in the market today have already captured the data and are monitoring what’s being said. Ten years ago, not many companies were providing that service, but now pretty much everyone does. But the analysis is mostly manual, one of our clients used to have more than twenty people reading 30,000 reviews per month for just one brand. Those people then had to link those complaints to specific topics. Birdie has helped automate a lot of that work.
Wow, that’s incredible. What are some of the challenges that come with operating in the space?
I think for any startup one challenge is making the right decisions on what to focus on and when. We’ve had opportunities to expand our focus but we’ve turned a number of them down because we want to focus on what we know and get really good at it.
Stepping back for a moment, what does the market look like today?
We are targeting the largest companies in the categories we’re working with (Electronics — including small appliances, Food & Beverages, Beauty & Hygiene, Household items, Toys). And for companies in these categories, we try to look at the most relevant ones. One of the criteria we use to determine if a company is a good target is the company in question has a consumer insights role. Today we are targeting about 300 companies and are very focused on the largest ones, think Fortune 1000, excluding companies that don’t sell consumer products. Thinking about the industries we focus on today, there are more than 2,000 companies that are large relevant companies.
Do you typically find that those companies are multi or single-brand companies?
I would say it’s split almost evenly. Most of our clients today are single brands, but we are negotiating with a few companies that have different brands under their umbrella. P&G for example is one of our clients, they have several brands, while Samsung, another customer, only has one brand. It doesn’t change what we do because as long as we have the category, we bring all the brands into that category. But for our clients, it’s important for them to be able to track that as they might want to compare their brand to certain other brands.
What does a typical deployment look like for a customer? Do they send you a data set? Do they point you to several URLs?
We need two things from customers, 1) the categories that they want to analyze such as TV’s, refrigerators, etc. and 2) the sources the company wants us to use (social media, Amazon, etc.) That’s pretty much all we need from them because our model takes care of everything else.
From start to finish, how long does it take to deploy your solution with a client?
If it’s a category that we are already working on and have gathered the data sources we can give access to customers immediately. If we have done some work on the category but need more sources it can take about two weeks. And if it’s a new category it takes four to six weeks.
How do your customers use the insights you deliver?
We have one example that we always like to use because it was very symbolic. During our analysis for Samsung, we noticed that there was a new niche audience that was buying innovative products. The audience we found was pet owners, at the time we were analyzing washing machines for Samsung and we started noticing a specific type of review that was starting to get more and more common. They were all saying that some washing machines were really good because it can remove pet hair from, there were also comments on how some machines weren’t as effective. And by flagging that information to Samsung they were able to create personalized campaigns that converted customers almost four times more effectively than previous efforts.
Love it. So how does Birdie manage bad data and deal with it today?
That’s one of the main challenges we face for sure, especially as we start processing more data. We have a model that’s constantly learning and today we’re giving feedback internally. But we are starting to also get feedback from clients. They can tag results as good or bad so we can keep learning and improving the model. We also do several steps of data analysis using the taxonomy we’ve developed to make sure that we get to a meaningful level of quality both for sentiment and for the categorization of the data. That said we are not 100 percent there and I think we will get very very good, but I think there will always be this challenge to some extent.
How does Birdie think about bias in data/modeling?
I think one of the main values we give to our clients and one of the main things that we as a company want to do is to remove bias from the process of learning about consumers. To help accomplish that, we as a company, try to have a culture that’s very diverse and helps us in our mission to make everyone feel heard.
Regarding Bias in Birdie’s data, our model starts by extracting a complete list of aspects from user reviews instead of looking for known patterns. This initial approach ensures we use spontaneously mentioned aspects and not a list of keywords created by our team. There are additional steps on the data pipeline based on frequency and combination of aspects that helps Birdie to detect artificially created reviews. By taking this approach, opinion mining from Consumer’s reviews reduces “Response Bias”, commonly found in focus groups and survey results. Still, there is some curation level to cluster and label similar aspects based on trained models that will inevitably carry training bias, but every cluster is transparent, thus allowing users to see contained aspects.
I know we touched on how Birdie is focused on CPG brands right now but how do you see Birdie’s product offering evolving in the future?
In the future, we want to be the platform that’s connected to other data sources in the space such as survey platforms that companies use like Qualtrics and Survey Monkey. In a future state, we also want customers to share sales data and other data sources with Birdie as well. In the end, we want to understand the consumer journey and help our clients discover where they should focus their energy to generate more revenue and keep their customers satisfied. To accomplish that, the more sources of data we can access, the more complete our understanding will be.
Lastly, how has Birdie evolved since its founding?
Funny you should ask, we started with a totally different vision. When we initially started Birdie, we researched issues in the eCommerce industry and our first perception of one of the problems was that it was very hard for consumers to decide what to buy. Our first approach was to help the customer by bringing all the reviews and product information into one place and help customers find the right product for them. That’s when we started learning how to collect data with crawlers and structure our data. Because of challenges with selling to consumers and our team’s contacts in the B2B space we discovered that everything that we were doing was very relevant for brands because they were kind of lost with all the data and no tools to analyze it. So, we decided to pivot from a B2C to a B2B model one year ago.
In case you want to learn more, check out Birdie’s website here: https://www.birdie.ai/
