After improving our filter experience, we also noticed a significant increase in the number of searches being run. We needed a way to understand how to users could use a combination of filtering and sorting to achieve the best results set.
We added tracking to our search bar, which allowed us to monitor which queries were being run by customers. To further understand the resulting data, we performed user interviews with our internal services team (who are offered as a service layer to brands while searching for influencers) regarding their perception of our current search offering.
We also compared our searchable influencer information to the data customers most commonly searched for. Additionally, we researched search best practices generally, both within the frameworks and tools we were already using and more forward-thinking features.
Our research into our current technology revealed that we could leverage a "relevance" score included in our search tooling to display results even if they weren't a "perfect match". This was especially interesting because our query monitoring showed that users were still using our search tool to find individual influencers, which would result in a poor experience if that specific influencer was not in our database. A relevance score would help improve this experience by allowing us to quickly parse the influencer’s public profile, and then find similar influencers who were in our database.
In order to fully leverage that technology, however, we would need to improve our data quality, as we noted that the number of influencers with complete profiles was relatively low, especially for some of the more common search areas (such as topics and location).
With regard to best practices, one trend we saw was that increasing the number of search terms produced "better" results - even when compared to directly filtering for the same information (searching for "Boston" versus filtering for Boston as a location).
Our forward thinking research combined with the interviews from our services team allowed us to identify specific opportunities/types of customers where advanced features would be valuable. These included saved searches for users who only wanted to activate a single type of influencer and boolean logic for marketers with very specific needs (e.g., "fitness, NOT vegan").
We made the technological improvements necessary to have the "relevance" sort option and made it the default sort order. Because of this, even when searching for a single influencer, we were able to display other similar ones (or only similar ones if we did not have that specific influencer in our database):
In conjunction, we undertook separate technical projects to improve the data density in key areas using machine learning. We coordinated with our services and implementation teams to standardize "best practices" around influencer search (beginning with a search and then using the filters to narrow the results set), and made sure that the user experience prioritized this type of search. We chose to beta test boolean logic in our search bar to further understand the value it would drive:
Finally, we piloted an alternative search experience focused solely on the search bar in an attempt to "force" users to search before they began filtering.
The key outcome from these changes was that when compared to a "filter-first" experience, the number of results returned was significantly higher. This, combined with our data efforts and relevance sort, reduced the number of reports of poor-quality results we received by 24%. Qualitative feedback from both customers and team members reinforced that search was now returning better results sets than it did when we relied on our old filtering strategies.
In addition to the one-time data improvement, we were able to integrate the technology that improved our data density into our influencer onboarding flow so that it is used every time someone registers with a brand.
Our alternative search experience proved to be generally useful, but was a nonstarter for users with more specific influencer needs. We are currently discussing trade-offs with our business team and may consider disabling it on a per-brand basis. Similarly, boolean logic was deemed to be useful but potentially confusing for the majority of users who didn't need it and weren’t familiar with it. We therefore rolled it out without any visible UI and allowed the implementation team to teach it to users as they saw fit.
Now that marketers are getting results sets that are larger and more accurate, we want to focus on ways to teach users about what number of influencers they should reach out to in order to hit their stated influencer marketing goals. This teaching could be done via additional in-app coaching, which would also aid in rolling out additional "advanced search" features.
To continue to iterate on our alternative search experience we are investigating ways to integrate our most valuable filters into the experience and to measure its value when compared to a traditional search across our customer profiles.