Goibibo Data-Driven Flight Search Result Optimisation
In the year 2000 , Sheena Iyengar and Mark Lepper tried out an experiment with Jam flavours. The experiment took place in a store in California . Shoppers were presented with a tasting stall, which carried either 24 flavours of jam, or 6 flavours, depending on the time of day. The test ran for a week and the results were astonishing .
The stall with more jams attracted more shoppers, 60% compared to 40%. However, when it came to sales, the results were completely different. The stall with 6 jams converted at 30% compared to a paltry 3% for the stall with 24 jams. Although more users came to a higher jam selection table lesser number of users actually bought those jams
Bottom-line : Lesser and more relevant the choices , more the conversion
Almost everyone reading this article would have booked flights from one OTA or the other . You put a search sector , a travel date , number of passengers and bang .Depending upon your search sectors , one is presented with a plethora of options . A typical search on Bangalore — Delhi sector on 20 days ahead of travel reveals flights in the range of 90 to 120 depending upon the OTA . Our historical data analysis showed that 60 % of these flights are never / seldom booked .
In short , there is an increase in the complexity for the users to make a decision there by increasing the total end to end transaction time . This is turn decreases the conversion rate from search result page to the booking review page . From a technology point of view , more options of flights increase the payload of the data transfer from server to the various channels like iOs , Android , mWeb and , web . Popular research done by Amazon ,Walmart , Kiss metrics says that a 1 second delay in page load time brings down the conversion by 7% . Data transfer is one of the key parameters affecting page load time .
We at Goibibo wanted to solve this customer problem by removing the clutter from the search result page and make things simpler for our users . A brief about the same is shared below , obfuscating the actual process .
As a first step , our data team , took a dig into our historical data for past 1 year for all the flights those were booked for each sector . From our past user study , we already knew that the three major factors on which user make a flight decision are Price , Stops and Travel time . We took all the booked flights for every sector and created a table extracting these three important factors .
The individual rows were then put into buckets with a frequency term , using this we created a training model . When ever a flight search was made , we checked every flight in the result for the metric called Level of Similarity to the training data , every flight in the search result page is given a Similarity score , the higher the score better the flight . Our flights technology team ensured that the above calculations were done on the fly with minimal impact on the processing time .
Results within 5 days of launch : As a first result our data transfer load for a typical search decreased by 31% . Our conversion from Search Result page to booking review page is showing a very strong positive trend.
