AI Reordering for Semantic Search

Search isn’t just about matching keywords – even more so when we’re talking about semantic search.

Semantic search is about finding the right information for the searcher at the right time.

It’s not just about finding the right keywords and concepts and speculating about how searchers will interact with the results.

Artificial intelligence (AI) re-ranking will take information about the people who come to search and customize search results for the individual.

This can be done at the cohort level, changing results based on trends, seasonality and popularity.

It can also be done individually, changing the results based on the wishes of the current searcher.

While implementing AI re-ranking in search engines is not easy, it brings tremendous value to conversions and searcher satisfaction.

Re-ranking with artificial intelligence

AI-driven re-ranking can improve search results regardless of the underlying ranking algorithm used by the search engine.

This is because good search results are more than business metrics like text relevance and raw popularity.

Good results take into account other signals and do so on a per-query level.

To understand why this is important, let’s focus on business metrics of popularity.

This is a good general ranking signal, but may not satisfy specific queries. A search query for “red dress” might show two different dresses in the first result: “backless dress with red accents” and “bright red summer dress”.

Backless dresses may be more popular as an overall dress and product.

But in this case, specifically, that’s not what the client wants.

They want a red dress, not one with red accents, and they click and buy accordingly.

Shouldn’t search engines use it as a signal to improve rankings for summer clothes?

Search Analytics

As the example above shows: understanding what searchers are doing is necessary for re-ranking.

The two most common events to track are clicks and conversions.

Typically, these are the only two necessary events, and must be events from a search.

The example above also highlights another important consideration: events should be associated with a specific query.

This allows search engines to learn from interactions between different result sets and user interactions. It pushes summer dresses higher in search results for the “red dress” query.

The same product may not be as popular as its neighbors in other queries.

When looking at your different events, you’ll want to weigh them differently as well.

Clicking results is a sign of interest, while making a purchase (or any other conversion metric) is a sign of commitment.

Rankings should reflect this.

Weighting doesn’t need to be complicated.

You can simply say that conversions are worth double-clicking.

You should test the correct ratio for your own searches.

You may also want to discount events based on the result rank when searchers see them.

We know that the position of a result affects its click-through rate (CTR).

Disregarding events, you can run into a situation where top results become more entrenched as they get more interactions, which makes them rank higher – and repeat indefinitely.

freshness and seasonality

An easy way to combat this self-reinforcing cycle is to discount events based on the time elapsed since they occurred.

This happens because everything that has happened in the past has less and less impact on re-ranking. That is, until a certain point, it has no effect at all.

For example, you can divide the impact of each event by 2 each day for 30 days. and stop using the event for ranking after 30 days.

A nice benefit of using freshness in a reordering algorithm is that it also introduces seasonality into the results.

Not only are you no longer recommending videos that were very popular a few years ago but are boring to people today; you’re also recommending “Learn How to Swim” videos in the summer and “Learn to Ski” videos in the winter.

YouTube has seasonality and freshness built into its algorithm for exactly this purpose.

Rerank using signals

Now that you have your signals and decay them over time, you can apply them to your search results.

When we see “artificial intelligence”, we often think of something very complex and incredible.

However, AI can also be as simple as taking data over time and using it to make decisions, as we have done here.

An easy way is to take a certain number of results and simply rearrange them according to the score.

For performance reasons, this number of results is usually quite low (10, maybe 20). Then, rank them by score.

As we discussed above, scoring can be as simple as multiplying conversions by 2 and adding clicks.

Adding a decay function adds complexity, as does discounting based on the resulting location – but the same general principles apply.

learning ranking

One downside of this reranking system is that you can only rerank fewer results.

If your result is otherwise popular but doesn’t rank well, it won’t get the attention it deserves.

The system also needs to log events and queries to be reranked.

It is not suitable for brand new product launches or User Generated Content (UGC) that frequently enter and exit the search index.

Learning to Rank (LTR) can solve these problems.

Like the reranking we discussed above, LTR is also based on the idea that the records that searchers interact with are better than the records they don’t interact with.

Previous reranking methods work by directly boosting or hiding results when associated with a specific query.

At the same time, LTR is more flexible. It works by boosting or hiding results based on other popular results.

LTR uses machine learning to learn which queries are similar (for example, “video games” and “game consoles”).

It can then rerank results for less popular queries based on interactions with more common queries.

LTR doesn’t just generalize queries; it also generalizes records.

The LTR model knows that certain types of results are popular; such as the Nintendo Switch game The Legend of Zelda: Breath of the Wild.

It can then start connecting to other similar results (eg, “The Legend of Zelda: Skyward Sword”) and boost those results.

So, if LTR appears to be much more powerful than typical reranking and provides more query and record coverage, why not just use LTR?

(In other words: it generalizes better.)

In short, LTR is much more complex and requires more specialized in-house machine learning (ML) expertise.

Also, understanding why some results are ranked more difficult in some places.

The first type of reranking allows you to see the number of clicks and conversions for one record compared to another.

At the same time, with LTR, you have an ML model that makes connections that may not always be obvious.

(Are Breath of the Wild and Sonic Color really that similar?)


While reranking works for all searchers, personalization sounds like it: personalization.

The goal of personalization is to take already relevant results and reorder them according to personal preferences.

While there is debate about the extent to which web search engines such as Google use personalization in search results, personalization is often performance that affects results in an on-site search engine.

This is a useful mechanism for increasing search interactions and search conversions.

Search Analytics

Like reranking, personalization depends on understanding how users interact with search results.

By tracking clicks and conversions, you can get a clearer picture of the types of results users want to see.

A notable difference between re-ranking and personalization in this regard is that depending on your search, you may need to adjust the way you apply personalization.

For example, if you sell groceries, you definitely want to recommend previously purchased products.

However, if your website sells books, you won’t want to recommend books that customers have already purchased. In fact, you might even want to move these books down in your search results.

But it’s also true that you shouldn’t push personalization so hard that users only see content they’ve interacted with before.

Search makes discovery and discovery possible. So if they return to the search bar, you should be open to the possibility that they would like to see something new.

Don’t rank results by personalization alone; mix it with other ranking signals.

Like reranking, personalization benefits from event decay.

Reducing the influence of old events can make searches more accurately representative of the user’s current tastes.

In a way, you can think of it as personal seasonality.

Personalization across users

The kind of personalization we’ve seen so far is based on an individual’s own interactions, but you can also combine it with what other people are doing in search.

This approach has shown a huge impact on situations where users have not previously interacted with items in search results.

Because users do not interact with search result items, you cannot, by definition, promote or hide based on past interactions.

Instead, you can look at users who are similar to your current user, and then personalize it based on what they’re interacting with.

For example, let’s say you have a user who has never come to you to buy clothes, but has already bought a lot of handbags.

You can then look for other users who have similar tastes and who have also interacted with the dress.

Intuitively, other customers who like the same type of handbag as our searcher should also like the same clothes.

Rerank and personalize discovery

Search is just one example of how re-ranking and personalization can make a difference. You can also use these same tools for discovery.

The secret is to think of your homepage and category pages as search results.

Then, obviously, you can use the same tools and get the same benefits as search.

For example, a home page is similar to a search page without a query, isn’t it? The category landing page does look like a search page with category filters applied.

If you add personalization and reranking to these pages, they may become less static. They will provide users with content they love to see and can push items that are more popular with customers higher up.

Don’t worry, personalization and re-ranking can get mixed up with editorial decisions on these pages or on internal search.

The best way to deal with this is to fix the desired results somewhere and re-rank around them.

We’ve already seen that personalization and reranking are two methods that improve search through user interaction with relevant signals.

You can let your user base influence the results by using interactions.

These interactions gradually tell search engines which items should rank higher.

Ultimately, searchers benefit from a better search experience, and you benefit from more clicks and conversions.

More resources:

Featured Image: amasterphotographer/Shutterstock

Leave a Comment