By Leonardo Giacomet on 31 January 2023
The average MLS has countless possible fields and lookups, which are challenging for agents to populate completely
Across three leading MLSs markets, our AI detects an average of 17 features in each listing
Given shared fields across our solution and the MLS dataset, we are detecting up to 28% more features on every property
AI is detecting many positive features more than twice as frequently as they are being populated on agent listings
Leveraging AI as part of the add/edit process reduces the amount of time agents spend creating listings and improves the overall quality of the MLS dataset
MLS Listing Upload/Creation is Outdated!
Multiple Listing Services (MLSs) pride themselves on being the most reliable source of property data. They spend resources on providing easy-to-use add/edit platforms and dedicate time to continually educate their members on best practices to provide an accurate overview of their market.
However, accurately creating a listing can be a challenge. The average MLS features over 800+ listing fields (and 2800+ lookups!), which means a listing input form can resemble an overwhelming sea of checkboxes and fields. While agents are eager to post their listings as quickly as possible, they must be balanced with ensuring all of the information is correct and complete. A single missing detail can result in a home can be excluded from a search, and a potential sale being delayed or missed entirely.
Example of what the listing feature selection process looks like in the average MLS
Due to the time it takes to populate correctly, image-level details in particular are frequently overlooked or neglected. Based on conversations with our MLS partners, as few as 5-10% of all images include classifications specifying the part of the property depicted in the photo - even fewer have a description or caption of the photo. While it’s understandable many agents aren’t willing to spend the additional few minutes it takes to add these details, this lack of data has profound ADA implications for downstream sites.
Beyond the sheer number of data points each property requires, there is also a challenge in having different agents evaluate properties in a standardized way. How large of a hill is needed for an agent to claim a home has a “mountain view”? Is a 3-foot by 3-foot paved landing outside the backdoor enough for an agent to claim there is a patio? It makes sense that agents are looking to present their property in the best light possible, but this lack of standardization makes it more difficult to accurately compare properties.
Given this reality, MLSs are stuck with two sub-optimal options. At best, their agents spend too much time entering data when they would rather be out working with their clients and closing deals. At worst, valuable property details are overlooked, impacting other agents and home buyers searching the MLS. Either way, MLSs are eager to find better solutions.
Is Artificial Intelligence ready to help ?
Given these challenges, many MLSs have looked to artificial intelligence (AI) to help. Using computer vision, a type of AI that analyzes images to understand them like a human, all of the photos of a property can be scanned in real time. In the same way that tax record information can be pulled to pre-populate a listing, the details the AI “sees” in the images can automatically populate the relevant fields within the MLS.
Instead of an agent having to create a listing from scratch, they can merely edit or review the listing to ensure everything is correct. Not only does this enable properties to be listed more quickly, but it also results in higher-quality listings with more details.
Higher quality listings with more details, without extra work… Really?
I’m sure many of the people reading this are skeptical. If MLSs currently have the most accurate data, how much can AI further enrich it?
To answer this question, we analyzed a year’s worth of listings from three of the top 10 MLSs by member count. We then compared the number of features detected by our RESO standardized auto-pop API and compared it to the number of features populated for each listing by their agent.
As each MLS has slightly different fields, including many that may not be part of the RESO data dictionary, we analyzed and compared the following two scenarios:
1. Total features: The total number of features detected by our API (some of which the MLS may not expose) and the total number of features populated by agents in each MLS (some of which our AI isn’t looking for… yet)
We excluded fields that cannot be determined by a photo, such as HOA dues, etc.
While all of the features our AI detects are RESO-standardized items, those features may not always exist as possible selections in all MLSs
2. Shared features: The total number of features detected among the subset of items our API detects and that were possible selections within the MLS
AI’s impact by numbers
The results were consistent across all three MLSs datasets. In every case, our solutions detected more “Total features,” finding an average of 17.1 total RESO features compared to 11.5 features populated within each MLS. In the most extreme case, our AI found almost 2x as many “Total features.”
Comparison of Restb.ai versus a sample of three MLSs
When looking at “Shared features” that were possible selections for both the MLS and our AI, we achieve up to 28% more detections compared to the MLS.
Comparison of Restb.ai versus a sample of three MLSs, considering only features shared both by the MLS and by our models
Diving deeper into 5 key fields of the data, the areas that our solution outperformed the most were the categories of appliances, interior features and exterior features. Where the MLSs populated more were on Heating and Cooling, which makes sense, as these components cannot always be determined from a property’s photos, or those photos aren’t uploaded to the MLS.
Comparison of only five categories
Interestingly, we saw several instances of MLSs vastly underrepresenting certain positive features. The chart below shows how certain traits like beamed ceilings, vaulted ceilings, cooktops, fire pits, crown molding, etc. are frequently omitted.
In one particular MLS, we saw the Double Oven feature field populated on only two listings, whereas we detected over 3,000 double ovens across the same set of properties. That’s a difference of over 1,000x!
Comparison of how frequently Restb.ai detects specific features versus MLS Data
Future implications for MLSs
Given the results, we believe that leveraging AI is necessary for MLSs to improve their datasets. No amount of member training can provide a similar improvement to an MLS’s data quality. Furthermore, AI is still just scratching the surface. While our current RESO standardized API is detecting 152 data points, nearly 100 more features are to be added by the end of 2023. Additionally, we’re working closely with the RESO Data Dictionary team to ensure alignment and easy adoption of more tags.
Tags available on Restb.ai’s Property API RESO and short term roadmap
For those MLSs that are already taking advantage of this technology through any of the MLS software providers supporting this technology (read about our partners’ in the press releases section), it’s worth considering whether there are any RESO standardized features that make sense to add as possible selections during listing creation. To re-emphasize, we detected an average of over 17 RESO-standardized features per listing (compared to less than 12 populated on analyzed listings)!
Additionally, we believe there are promising implications to real estate working with a more standardized and uniform dataset. RESO is working diligently in this direction, certifying and bringing transparency about data availability. With cleaner and more complete data, agents can compare and find properties more efficiently. It will improve property valuations and make it easier to track trends. We see it as a key building block to create more interactive and visual searches. The possibilities are endless, and the possibility of computer vision is only beginning to be realized.