The Age of Conversational Recommendation: Making search more iterative and less rigid for users


Among Airbnb’s “100+ innovations and upgrades” released last month was a number of extensions to its accommodation search mechanism. These are designed to provide renters with an ability to search more flexibly, in a clear effort to counter the impact of ‘over-filtering’.

Airbnb is not alone. Several of the major product search platforms have been on a journey of making their search mechanisms more flexible and user-friendly. In each case, the focus has been different. Some have started to enable user feedback during the search journey, while others (like Airbnb) have tackled the rigidity of their filter system.

As yet, these feel like a somewhat ad-hoc set of adjustments. Each one is for sure improving the User Experience, and it is good to see clear recognition that it is time to move on from the traditional one-shot filter-based search mechanism. However, will the incremental adjustments implemented by players like Airbnb, Google, and Rightmove provide a solid basis for a longer term evolution?

The case of Netflix’s feedback mechanism suggests that it does not. A more holistic conversational and preference-based approach (such as Traverz Conversational Recommendation technology) would provide a superior search foundation that encompasses all these ad-hoc adjustments while enabling a natural super-set of deeper solutions.

Lets review the changes made by three key players: Airbnb, Google, and Rightmove, and wrap up with a look at what went wrong when Netflix sought to incorporate user feedback into its recommendations.

Airbnb – adding ‘flexibility’

Last month, Airbnb released “a whole new way for guests to search on Airbnb”. One of the key changes made was the addition of a new feature called “I’m Flexible”. This allows renters to search for flexible dates, property features, and property types.

Flexible Dates is a pretty cool feature that includes the ability to adjust not only the length of the stay, but also to search for any weekend, week, or month throughout the year. It avoids users having to run several searches for different time periods, in order to see which Airbnb listings are available. For example, you can search for a weekend in either May or June and see the results immediately.

Flexible Matching is designed to help prevent the problem of “over-filtering” that plagues users of classic filtering mechanisms. For example, if you filter locations by search parameters (like Wi-Fi, parking, or hot tub), you may not get so many options to choose from – or indeed none at all. Flexible matching will then show you locations that are just outside of your search parameters – i.e., they are missing one or more of the desired features – and label those features as being missing. As Airbnb stated in its release notes: “This way, you never miss out on a great stay that falls just outside what was specified in a search”.

Images courtesy of Airbnb

The concept of Flexible Matching is absolutely an improvement on the classic filtering approach, where each criteria is rigidly applied. It avoids the frustration of getting few or zero results (and not knowing which criteria is the driver!).

At the same time, it feels like Airbnb has implemented a half-solution. The new matching flexibility is only available for the property features that are listed in Airbnb’s filter UI, which are pre-defined by Airbnb and represent a limited set of common criteria. Users also cannot ‘respond’ right away when they see something that they like in a property listing – they have to tediously go back to the filter menu and see if that feature is indeed available as part of the search criteria available or not. Unless you only have common requirements and are pretty familiar with Airbnb’s range of available criteria, chances are that this will feel like a modest step rather than a large leap forward.

Google image search – recognising that search is iterative

When you enter an image search, Google presents a series of context-specific buttons above the image results list with additional search terms that make your search more specific. So if you search for “dog images”, it adds buttons for “puppy”, “labrador” etc. Once you select one of these, it is displayed in blue background as the first button, with the remaining buttons being adjusted based on the content of the new set of image results.

This is effectively a form of Conversational Recommendation and demonstrates the feedback loop of that search mechanism. The platform presents the user with context-relevant options for feedback, and then reacts upon user feedback with a new set of recommendations.

This works great, and it illustrates the benefits of being able to iteratively build a search. As yet the same functionality is not available for standard searches, even though the rationale and use-case is very much the same.

A logical extension is to not only show feedback buttons at the top of the results list, but to show them alongside each individual result. This approach (as taken by Traverz) encourages and enables the user to respond directly to a specific item rather than having to scroll back to the top of the page and find the relevant tag.

Rightmove – preference based search but only for keywords

Rightmove is the UK’s leading property search portal, and has held a dominant position in that market for very many years. Around 18 months ago, Rightmove introduced a “Keyword Sort” feature that was billed as the way to personalise your property search.

What Keyword Sort adds to the standard search is a mechanism to prioritise properties whose listings contain user-entered key words. Up to 3 of such key words can be entered, and the standard results list (based on the main filter settings) is then ordered to first show properties that meet all keywords, then those that meet one less keyword, etc.

Rightmove demonstrates the power of a preference-based search mechanism, which allows users to truly personalise the recommendation order of the results list. However this capability is available only for the user-entered text. Standard filter bar features such as price, number of bedrooms, and property type do not have such a prioritisation method. And as a result, there is no impact on the ‘over-filtering’ problem – if you search for a detached house in York with max 2 bedrooms and a garden, you get just 1 result (a cottage that clearly is not detached). The user is then left to determine which of those criteria is problematic by manually adjusting them one by one.

Another feature worth noting on Rightmove’s platform is the more-like-this button, labelled as “See similar properties”. Clicking this button sets up a search using the property’s postcode, a search radius within 0.5 miles, a price range bracketing the price of this property, and minimum number of bedrooms equal to that of the selected property.

As is somewhat predictable, the new search criteria tend to result in a very small number of results (if any). This is due to the inflexibility of the filter criteria (location, price, bedrooms) that are set, again illustrating that the more-like-this and key-word features are fighting with an underlying filter system that is inflexible.

Another thing to note is that this “see similar properties” button is only displayed if you open the property directly (e.g. from a Google search). We presume that the idea behind it is to provide a way for you to broaden out to other properties, given that you do not yet have an existing results list to go back to. Clearly the same mechanism of adding/setting search criteria could also be provided when the user was already browsing on Rightmove, but this has not been implemented. Which is a shame because such a more-like-this functionality (which is always present in Traverz) is a great way for users who are browsing a diverse results list to indicate their interest and be given a very relevant set of recommendations.

Netflix – insufficient recommendation data to… recommend

A final example of a major platform that is exploring ways of improving the search mechanism is Netflix, the popular streaming service. Although Netflix already changed its five-star rating system to a thumbs-up / thumbs-down approach several years ago, it is nevertheless instructive to briefly review this change and its consequences. In particular, it is a good example of why feedback mechanisms must provide some meaningful clarity as to the user’s intent – or it becomes counterproductive.

Netflix CEO, Reed Hastings, is famous for having cultivated a culture of candid feedback. “Frequent candid feedback exponentially magnifies the speed and effectiveness of your team”, he says. Yet the same cannot be said for the consumer’s ability to provide input and feedback to the Netflix platform. Netflix has always had a very basic filter and search system, and its dominant UX can largely be summed up as “here is a long list of recommended movies/series – go browse”. Those recommendations are based on your viewing history (including anything it can glean from your platform interactions) and the cohort that you are in. Here is a link to a brief explanation from Netflix.

Originally, Netflix provided a 5-star rating system which acted as a key input into their recommendation algorithms. Several years ago this was modified to a simple thumbs-up / thumbs-down feedback system.

The original 5-star system had some clear issues, with users confusing the purpose of the rating (to provide input for individual recommendations) with an external or objective rating for each title. However, given that the user was only ever able to provide one rating for each title without any nuance regarding various aspects of that title (actors, length, genre, etc), then it is not surprising that moving from a 5-tier to a 2-tier system was likely to be fraught with issues.

And indeed, users complained that the new feedback system was close to meaningless, and did not allow them to express what kind of titles they liked. Many stopped using it after discovering that they were being recommended titles that they did not consider interesting. Of course, it is highly likely that they experienced similar recommendation issues under the 5-star system, but the new approach likely led users to feel a strong reduction in their control over the recommendations.

This highlights the importance of providing users with clear feedback mechanisms that are granular enough to be meaningful, and to be transparent about what elements of a recommended product do and do not match their indicated preferences. As many users have reported, the claim “when you look at your Netflix homepage, our systems have ranked titles in a way that is designed to present the best possible ordering of titles that you may enjoy” feels like rather a bold statement when there is so little meaningful input.

The Age of Conversational Recommendation

Airbnb has clearly recognised that searching for properties requires more modern and flexible mechanisms than the traditional filter system. With the Covid pandemic putting the travel industry under severe pressure, it is looking to innovate around the property search process in new ways. Improved search mechanisms that provide flexibility around dates, property features and destinations are a strong sign that preference-based search is entering the mainstream.

Others, such as Google and Rightmove, have already made moves to provide users with mid-search feedback mechanisms. When taken together, these disparate and often apparently ad-hoc adjustments are pointing to the emergence of preference-based Conversational Recommendation. However in each case there has only been one aspect tackled, and/or the implementation has been limited. This is partly due to the difficulties posed by attempting to ‘layer’ further flexibility and interaction on top of a rigid one-shot filter system. For many of these platforms, the search paradigm will need more substantial change to enable further steps to be taken.

At Traverz we have spent several years developing and fine-tuning this search technology, and we look forward to exploring it further with the above-mentioned platforms and their competitors.

Ready to leverage Traverz for competitive advantage?

Contact us to supercharge your product search experience!


Traverz is changing product discovery for good. Are you ready to take advantage?