How AI Can Help in Re-Ranking for Semantic Search

Semantic search is all about making search queries more effective by understanding the user’s intent behind them. It goes beyond just finding the right keywords and concepts and delves into how searchers will interact with the results.

Artificial intelligence (AI) can help to tailor search results more effectively based on the individual user’s needs. This can be done on a cohort level (changing results based on trends, seasonality, and popularity) or individually (changing results based on the current searcher’s desires).

This personalized approach to search results is more effective because it takes into account the user’s specific context and needs. This results in a more relevant and useful search experience that can save time and frustration.

Let’s say you’re looking for a new car. A traditional search engine might show you a list of cars that match your query, but an AI-powered search engine could go a step further. It could show you cars that match your budget, needs, and preferences. It could even show you cars that are similar to ones you’ve considered in the past.

This kind of AI-powered search is called re-ranking. It’s a way of taking the search results that are already returned and changing their order to match the user’s needs better.

Re-ranking is not easy to implement in a search engine, but it brings outsized value for conversions and searcher satisfaction.

Re-Ranking With Artificial Intelligence (AI)

Despite the underlying ranking algorithm a search engine uses, AI-driven re-ranking can improve search results. The reason is that good search results are more than just textual relevance and business metrics like raw popularity.

Ideal results take into account other signals as well, and they do so on a per-query level. This is especially important when you consider business metrics like popularity. While popularity is a good general ranking signal, it can fall short for specific queries.

For example, a search query for “white shoes” might have two different types of results: “sneakers” and “dress shoes.” The sneakers might be more popular as an overall product, but in this case, specifically, they’re not what customers want. They want dress shoes, and they click and buy accordingly.

The search engine should be taking that as a signal to rank the dress shoes higher. Shouldn’t it?

Search Analysis

When it comes to re-ranking, analysis is key. To understand how to re-rank your results, you first need to understand what searchers are doing. For that, you need data.

The two most common events to track are clicks and conversions. In general, those are the only two events necessary and must be events coming from searches. Clicks are important because they show searcher engagement. A searcher who clicks on a result is telling you that they’re interested in what you have to offer.

It is also clear that these events should also tie to specific queries allowing search engines to learn from the interplay between the different result sets and user interactions.

As an example, if a user clicks on the second result for the query “white shoes,” it is a sign of interest. However, if the user ends up purchasing the product from the fifth result, it is a sign of commitment.

Both clicks and conversions are important, but they’re not equal. A click is a signal of interest, while a conversion is a signal of commitment. That’s why when you are looking at your different events, you’ll want to weigh them differently.

Testing the ratio of clicks to conversions (or any other events) is a necessary part of finding the right recipe for your own search engine. It is a well-known fact that clickthrough rate (CTR) varies by position. The first result gets the most clicks, and each subsequent result gets fewer and fewer.

This has been referred to as position bias, and it’s something that you need to take into account when you’re looking at your events.

For example, if you’re looking at a query where the first result gets 100 clicks and the second result gets 50, it’s not an apples-to-apples comparison. The first result may have a CTR of 50%, but the second result has a CTR of 25%.

To compare them accurately, you need to discount the events based on their ranking. In the example above, you would give the second result twice the value of the first result because it’s getting half the clicks.

You can use this same idea to compare any ranking position to any other. You just need to discount the events by the position’s CTR.

Freshness And Seasonality

Discounting for position bias is important, but it’s not the only way to adjust for ranking. You also need to take into account the time since the event occurred, or what’s known as freshness.

Over time, events that happened in the past have an increasingly small impact on re-ranking. That is, until, at some point, it has no impact at all.

For example, you might divide the impact of each event into each day and for 30 days. After 30 days, stop using the event for ranking.

The benefit of using freshness in the re-ranking algorithm is that it also introduces seasonality into the results.

For example, if you’re a clothing retailer, the results for a query in June are going to be different than the results for the same query in December. In June, you might want to surface swimsuits, while in December, you might want to surface sweaters.

You can use freshness to your advantage by surfacing the most relevant results for the time of year.

To do that, you need to track not only when the event happened but also what time of year it happened. Then, when you’re re-ranking your results, you can take into account the seasonality of the event.

For example, if an event happened in December, you would give it less weight than an event that happened in June.

The idea is that an event that happened recently is more relevant than an event that happened a long time ago. And an event that happened in the same season is more relevant than an event that happened in a different season.

You can use this same idea to track other types of seasonality. If you’re a news site, you might want to take into account the breaking news cycle. If you’re a travel site, you might want to take into account the seasons for different destinations.

The key is to track the events not just by when they happened but also by when they’re relevant.

Use Signals to Re-Rank Results

Once you’ve tracked your events and discounted them for position bias and freshness, you can use them to re-rank your results.

We often think of artificial intelligence (AI) as something incredibly complex and inscrutable. But AI can also be as simple as taking data over time and using it to make decisions.

One easy approach is using a score based on the data you’ve collected. For performance reasons, the number of results you re-rank will generally be small. You can use this same idea to compare any ranking position to any other.

Learning To Rank (LTR)

If you want to get more sophisticated, you can use a technique called learning to rank (LTR).

LTR is a machine-learning algorithm that takes data and uses it to understand how to rank results. It is different from the previous approach in two ways.

1. It takes into account a lot more data. In the previous approach, we were limited to the data we could collect on our own site.
2. LTR is able to understand how different types of data relate to each other. In the previous approach, we were limited to using data that was directly tied to a query.

For example, we might use data about how many times a result was clicked for a query. But we couldn’t use data about how many times the result was clicked for other queries.

LTR boosts or buries results based on other popular results. It uses machine learning to understand which queries are similar (e.g., “white label SEO” and “SEO Resellers”). Based on that understanding, it can then re-rank results on the less popular queries based on interactions on the more common ones.

LTR is a bit complex and requires specialized machine-learning expertise. In addition, differentiating why certain results are ranked in certain places can be difficult.

Personalization

Personalization is another approach to re-ranking results that can be used in addition to, or instead of, the methods described above. It is about taking results that are already relevant and re-ranking them based on personal tastes.

For example, if you’re a retailer and someone searches for “women’s shoes,” you might want to re-rank the results based on the shopper’s past purchase history. If the shopper has bought mostly flats in the past, you might want to put flats at the top of the results.

Or, if you’re a news site, you might want to re-rank the results based on the reader’s past interactions. If the reader has clicked on mostly political stories in the past, you might want to put political stories at the top of the results.

To do personalization well, you need data about the shopper or reader. This data can be collected in a number of ways, including cookies, device IDs, and loyalty programs. The key is to have data that can be linked to the individual. Once you have this data, you can use it to understand the individual’s preferences and re-rank the results accordingly.

There are a few challenges with personalization.

• You need to have enough data about the individual to understand their preferences.
• Personalization can lead to filter bubbles. A filter bubble is when someone only sees content that they agree with or that is similar to what they’ve seen before. This can be a problem because it leads to echo chambers where people only see content that reinforces their existing beliefs.

It’s important to strike a balance with personalization. You don’t want to show people only content that they agree with. But you also don’t want to show them content that is so different from what they’re used to seeing that it’s uncomfortable.

Re-ranking and personalization can have a big impact on the shopper or reader experience. They can be used to improve conversion rates, engagement, and even loyalty.

However, it’s important to use these techniques thoughtfully. You don’t want to end up with a filter bubble or an echo chamber. You also don’t want to make the experience so personal that it’s uncomfortable.

The key is to use these techniques to enhance the experience, not replace it. You want to find the right balance that works for your business, your users, and your data.