Summer “SLaM” & “CoSMo” Kramer: Investigating “Compressing Search with Language Models,” a Google Research Paper, & Why SEOs Should Care (Probably)

By Ethan Lazuk

Welcome to a new week of Hamsterdam Research! 🐹

If you’re new here, this is where we look at recent AI research papers (usually from Google) to learn just what the heck they’re talking about and explore their hypothetical implications for the future of search and SEO strategies.

This week, we’ll look at a research paper from Google called, “Compressing Search with Language Models.”

It was published on June 24th, 2024, and its contributors are Thomas Mulc and Jennifer L. Steele.

Here’s how Gemini explained the paper’s contents and significance from an SEO’s perspective (ignore my grammatical error in the prompt 😅):

Found a good paper for this week's Hamsterdam Research. Recently published by Google: "Compressing Search with Language Models" – https://t.co/EOlpozIt78 Stay tuned! pic.twitter.com/PmYRqEXXax

— Ethan Lazuk 🪬 (@EthanLazuk) July 3, 2024

🚨 When Gemini mentions “better understand the intent behind search queries,” know that we’re not just talking about the relevance of content for rankings.

We’re talking about real-world outcomes or even predictive search. Imagine Google knowing better which terms lead to specific actions, such as consumer buying habits, and even satisfying those intents before a query gets typed. 😳

Those are the kinds of capabilities that SLaM and CoSMo in this paper could lead to.

But don’t worry about the details just yet.

We’ll cover them in depth, as well as elaborate on the possible implications for SEO strategies.

Even if you’re not an SEO but simply interested in this paper or topic, stick around, because this will help explain it all in easier terms.

Still intrigued?

Let’s rock! 🪨

Oops. Let’s try that again …

Let’s rock! 🎸

To start, we’ll review the paper’s abstract and key vocabulary. 🙌

“Hey, I can’t see that on my phone.” 🤳

Ah, right you are.

Here’s the abstract in HTML (with my bolding and highlights):

“Millions of people turn to Google Search each day for information on things as diverse as new cars or flu symptoms. The terms that they enter contain valuable information on their daily intent and activities, but the information in these search terms has been difficult to fully leverage. User-defined categorical filters have been the most common way to shrink the dimensionality of search data to a tractable size for analysis and modeling. In this paper we present a new approach to reducing the dimensionality of search data while retaining much of the information in the individual terms without user-defined rules. Our contributions are two-fold: 1) we introduce SLaM Compression, a way to quantify search terms using pre-trained language models and create a representation of search data that has low dimensionality, is memory efficient, and effectively acts as a summary of search, and 2) we present CoSMo, a Constrained Search Model for estimating real world events using only search data. We demonstrate the efficacy of our contributions by estimating with high accuracy U.S. automobile sales and U.S. flu rates using only Google Search data.”

We’ll start by breaking down the key terms.

Some of these you’ll already know probably, but just in case, 🐻 with me.

[Note: I’ll be using Gemini Advanced to help explain some of the terminology. All quotes will refer to the research paper, unless otherwise stated.]

Google Search is the predominant search engine worldwide. Its global market share is consistently upward of 90%. It launched in 1998 with a patented algorithm called PageRank and introduced semantic search and the knowledge graph in 2012 and SGE (now AI Overviews) starting in 2023.

As a search engine, Google allows users to search for information on the internet using terms (queries, or strings entered as keywords or phrases).

The intent of those terms refers to a user’s goal when searching them (explicitly or implicitly stated). We as SEOs generally define search intents as commercial, navigational, informational, etc. This Google paper goes beyond that by tying search data to real-world outcomes.

Previously, user-defined categorical filters (like what you’d find in Google Trends, such as “Cold & Flu” or “Autos & Vehicles”) had been used most commonly to group search results based on topics. This is done to reduce the number of unique terms needed for analysis and modeling, or what’s meant by “shrink the dimensionality of search data to a tractable size.”

However, the researchers are proposing a different approach to reduce the complexity of search data, while still keeping the most important information from individual terms.

Instead of relying on user-defined categories, they use pre-trained language models, which are LMs that are already trained on a large corpus of text data to learn the relationships between words and phrases and are able to generate fixed-length vector embeddings that capture the semantic meaning of search terms.

Their SLaM Compression method (or Search Language Model Compression) creates a representation of search data that has low dimensionality, meaning vector embeddings with a small number of values.

The SLaM method is memory efficient, meaning it takes up less storage space, which is more practical for large datasets, and it effectively acts as a summary of search, meaning the vector embeddings capture the most important information, or a condensed representation of the original search terms and their frequencies (search volume).

Additionally, CoSMo (Constrained Search Model) is their model designed to estimate real-world events (like product sales or disease rates) using only data from Google Searches. Specifically, the researchers applied CoSMo to predict U.S. automobile sales and flu rates in their experiments.

That covers the abstract and its basic concepts.

Feel free to take a moment to fuel up before we proceed. 🥤

Cool. Let’s now delve into the research paper for some deep learning. 🤿

If you wish to follow along, you can grab a PDF or the HTML version on arXiv.

The full paper has seven sections, including an 1) introduction, 2) approach, 3) related work, 4) experiments, 5) model interpretability, 6) ethical use of data, and 7) conclusion.

We’ll summarize each of them below (within reason). 🤗

1. Introduction 👋

The researchers start off by explaining the value of Google Search data for machine learning models. 🤖

They also distinguish between Google Trends and search logs.

Practically speaking, users’ search terms can provide information related to “real world events such as consumer purchases, economic activity, or illness rates,” and that can be used in “forecasting and predictive models.”

“These existing approaches all create machine learning features by summarizing search data from a time period (e.g., day), and then use these features to predict events (e.g., automobile sales),” they explain.

That prior research also “uses two forms of Google search data: Google Trends and search logs.”

Let’s delve into their differences from an ML perspective.

Google Trends data 📈

The researchers explain how Google Trends “groups terms into search categories (such as ‘Cold & Flu,’ and ‘Autos & Vehicles’) and returns an indexed value for the search volume in that category for a particular day and geographic region.”

However, these classification methods produce “a coarse signal of consumer interest, where queries across a spectrum of intent are lumped under a single trend / category as if they were identical.”

As a result, “the dimensionality” of Google Trends data “is relatively small (due to the relatively small number of categories).”

That’s a double-edged sword. It makes Trends data “easily digestible for most downstream machine learning applications,” but given the course nature of its signal, some “approaches rely on additional features such as historical sales or other economic indicators.”

Search logs 🪵

As the researchers explain, “Search logs contain pairs of search terms and their frequency (i.e., search volume) over a given time period in a particular geographic area.”

However, this search data has the opposite constraints as Google Trends.

Given how the “number of unique terms is very large, modeling done using search logs requires that the data for a given time period be summarized into a digestible format and dimensionality for machine learning.”

Therefore, finding “a way to transform millions of distinct textual search terms into useful and tractable features that can be used by downstream machine learning” remains the “primary challenge” for modeling with “raw search logs.”

With that said, past research has shown that “by aggressively filtering the search data and one-hot encoding terms, you can create search features small enough for machine learning.”

To explain those terms a bit more, one-hot encoding is the technique of converting words (in this case, Google Search terms) into numerical representations (vectors) using a unique identifier or index.

One-hot encoding gets its name because all elements are zero except for the one at the index corresponding to the term.

For example, we might have three search terms one-hot encoded as:

• toyota prius [1, 0, 0]
• ford taurus [0, 1, 0]
• kia sorento [0, 0, 1]

SLaM to the rescue 🤏

In order to “summarize the search terms and their frequency” from “large search logs,” the researchers use “language models (LMs) to quantify each search term,” along with creating “a custom model tailored for predicting targets using search data.”

SLaM compresses the search logs “into tractable features for modeling” by using “the embedding vectors generated by LMs to retain the semantics of individual terms.”

The output from SLaM are “features,” which the researchers call “search embeddings.” 🎉

By not relying on user-defined filters, their approach “can be applied at any time granularity (e.g., daily or weekly level), yielding an aggregated search embedding for each time period that is memory efficient while being highly predictive of many events.”

Enter CoSMo 🌌

The search embeddings created by SLaM “are then incorporated into CoSMo, a constrained search model” that “outputs a score between zero and one.”

The score from CoSMo “can be thought of loosely as the probability of the dependent variable occurring for an average search term.” For example, think of “the probability of a sale or the probability of having the flu.”

In their case studies using “reported U.S. flu rates and the U.S. auto sector sales,” the researchers found “that using our search embeddings increases predictive power by 30% in auto sales compared to classification embeddings, and our method is on par or better than existing autoregressive approaches for flu modeling, despite only using search data as a model input.”

To elaborate on those “existing autoregressive approaches” they mentioned, this refers to models that use past values of a variable as input features for predictions, thus assuming a variable’s future values would be dependent on its past values.

The search embeddings from SLaM additionally can provide “useful insights into how consumers and patients use search by scoring the individual search terms and highlighting terms with high scores (i.e. high probability of purchase / having the flu).”

This is a key part to understand. 🫡

By scoring individual search terms based on their probability of being associated with a particular outcome (like buying a car or getting the flu), the researchers can then analyze the high-scoring terms and get insights into how consumers or patients would use search to gather information and make decisions.

To give an example, certain high-scoring terms might include specific automotive queries that users search for when they’re more likely to be in the market for a new car.

In short, it’s a more nuanced way to understand how a search query relates to a user’s intent as well as the likelihood of different real-world outcomes.

“This is a new capability in search modeling,” the researchers explain, “because most classification methods (e.g., Google Trends) treat all terms within a category as identical, which makes backing out the importance of any one term impossible, while other classification methods that operate on the individual term-level (e.g., one-hot encoding each term) cannot handle terms outside of their very limited set of included terms.”

Getting back to the implications of SLaM and CoSMo for search and SEO, imagine if Google had even more nuanced predictive power for users’ search intents, not only to determine which content is relevant but also what real-world outcomes (specific actions) a search term may lead to.

Meanwhile, as an SEO or marketer, imagine if you had that same level of insight into your audience’s search behavior. How might that influence the content you produce? 🤔

We’ll now get more into the technical details of the paper and how SLaM and CoSMo work.

I’ll probably tread lightly here to keep this digestible (you know, “shrink the dimensionality” and all). That said, we should still learn the key parts that are generalizable, especially for insights about term embeddings (vectors) and semantic search.

2. Approach 🧭

When it comes to modeling search data, the researchers see it “as a two-step problem.”

The first step is “compressing / aggregating search (feature engineering),” which refers to the transformation of raw search data into a format suitable for machine learning (i.e., reducing the dimensionality and aggregating the data over time).

This is how they create a set of features to capture the most important information from the search logs while still keeping the data manageable.

The second step is “choosing an appropriate model given the features to model the downstream target (model selection),” which involves selecting an ML model that suits the task at hand based on the type of data inputs, desired output, and available computational resources.

In other words, once the search data is transformed into features that are suitable for machine learning (data input), they need to choose the best model to then predict the target variable (real-world outcomes).

As the researchers explain, their approach “leverages LMs to collapse the query space to a tractable size that retains information about the query semantics, without the need for filters or manual data manipulation.”

How is what they’re doing different?

Well, rather than “using a binary classifier to map a search term to a one-hot vector,” they’re instead using “an LM to map the term to a point on the 𝐷-dimensional unit-sphere.”

To break that down a bit more, traditional methods used a binary classifier (yes or no question) to say whether a search term belonged to a category or not, and then it assigned the term a one-hot vector. However, that binary classifier approach is limited because it only considers whether a term fits a predefined category, but it doesn’t capture the semantic nuances of a term’s meaning.

On the other hand, language models are more advanced and can understand the meaning and relationships between words and phrases.

SLaM maps each search term to a point on a “𝐷-dimensional unit-sphere,” which is a type of embedding space of geometric representations, where individual points capture the semantic meaning of a term, different dimensions of the sphere represent different aspects of a term’s meaning, and terms with similar meanings are located closer to each other.

By aggregating “the search terms along these new 𝐷 dimensions,” the result is “a search embedding that has dimensionality 𝒪(𝐷)” — is easier to analyze while still retaining the essential information.

SLaM jam time 🪜

SLaM is “a model that takes search embeddings as the primary input and outputs an estimate for the target variable,” the real-world event it’s trying to predict.

“The model has inductive biases and constraints,” meaning specific assumptions and limitations that “are specific to search data, the underlying distribution of search embeddings, and the limited quantity of targets available for model fitting.”

In other words, those “inductive biases and constraints” help the model generalize better to new data while avoiding overfitting, a problem in machine learning where a model learns its training data so well (memorizing it, basically) that it performs poorly with new, unseen data.

So when are we getting a look at this model?!

Right now.

“At a high level,” the researchers explain, “SLaM aims to map individual terms to a fixed-length embedding using a language model, then aggregate the embedding statistics to remove the individual terms from the dimensionality.” This is shown in Figure 1 below:

As you can see above, the search terms are input into the SLaM Compression system (far left), and then each term is passed through a language model (LM) (left).

The LM then generates a fixed-length vector representation (embedding) for each search term (center).

All the embeddings are then aggregated into a single vector (right), which is called the search embedding (far right).

Bring on CoSMo 💫

Recall how we mentioned overfitting a few moments ago. That’s where CoSMo comes into play.

“When predicting real world events at a daily or weekly frequency most models are prone to overfitting due to the curse of dimensionality,” the researchers explain. Although “the number of targets is limited by the time period and regions, it is easy to add dimensionality to the features used for modeling.”

The “curse of dimensionality” refers to how high-dimensional data (with lots of features or variables) can lead to the amount of training data a model needs growing exponentially. In a word, there are always more possible relationships to consider.

Adding too many features to a model can lead to overfitting, where it not only learns the meaningful relationships in the training data but also the noise and random fluctuations, making it difficult to generalize with new data.

“Although there exist approaches like Lasso regression to combat this,” they explain, “we offer a unique modeling approach that is less dependent on regularization tricks in the loss function.”

Lasso regression is a technique used to combat overfitting by adding a penalty term to the model’s loss function (the error between the desired and actual outputs) to discourage it from relying too much on any single feature.

Instead of that approach, CoSMo is built on a structural model that estimates the probability of an average search contributing to a target event (like a product purchase). This means it considers both the total number of searches in a given time period and a function that converts search embeddings into a probability score between 0 and 1. This score represents how likely it is a particular search term is associated with a predicted event.

More specifically, CoSMo is unique because of its two self-regulating characteristics.

One is dynamic search volume (Vt). Given how the total number of searches varies daily, the model learns to adjust its predictions based on those fluctuations, making it more robust to changes in search behavior.

The other is constrained probability (P). Since the output of the probability function is constrained between 0 and 1, this limits the model’s output range and controls its variance, helping to prevent overly extreme predictions. Despite introducing some bias, this leads to better performance on unseen data, the researchers explain.

This section of the paper has a lot of formulas, which is why I’ve summarized it with fewer quotes.

That said, we also get to see what CoSMo looks like in Figure 2:

To explain what you’re seeing above, CoSMo starts with the search embedding (from SLaM) being fed into a probability model (blue).

Geographic indicators (geo indicators) are then incorporated to account for regional variations in search behavior, with GeoMasks applied to filter or adjust the search data based on geographic locations (gray).

Time-based factors then get considered to account for seasonal or temporal trends in search behavior, while calendar multipliers are applied to adjust the model’s predictions based on specific dates or time periods (red).

The search volume (number of searches for a given term) is also taken into account, while volume scaling is applied to adjust the model’s predictions based on the overall search volume (green).

The outputs from the GeoMasks, calendar multipliers, and volume scaling then get combined through a series of multiplications (yellow).

This leads to the final output of the CoSMo model, or an estimate of the real-world event, considering all the input factors and adjustments (dark gray).

3. Related work 📚

At this point, we’ve gotten to know SLaM and CoSMo pretty well. Now we’ll just skim the rest of the paper for interesting stuff before getting to our main takeaways for SEO.

In the related work section, the researchers share information about how Google Search data has been used in research previously.

The researchers note that Google Search is used by “billions of people” every month who type “queries into the search bar to find information on the internet.”

“As early as 2009,” they explain, “researchers were incorporating data around searches into predictive models,” including to predict “U.S. flu rates” and “economic indicators.”

Search representations 🪟

“Individual search queries have been represented as embeddings in many retrieval search tasks,” they explain, yet “for predicting external events using search data, as far as we are aware, all aggregations of search data to date have been through binary 1/0 classification,” the method we described earlier as having shortcomings for not capturing semantic meanings, as “a search query is either a member of the class (1) or not (0).”

They also discuss classification embeddings, “where the embedding vector represents the counts for each mutually exclusive category in each dimension” (one-hot encoding), except that “approach struggles with queries that don’t neatly fit into a single category, and downstream models built on top of this method are sensitive to the classification system.”

“Classification embeddings lose much of the nuance in search queries,” the researchers explain, giving the example of how a query like “best new family SUV” could be classified the same as “New model Lamborghini Urus,” despite that “the intent of the searches might be quite different.”

The paper also references filtered one-hot embeddings. Here, researchers choose a smaller set of important search terms to focus on (because using all terms would be too complex for the model). Each selected term is assigned a unique code (vector) using one-hot encoding. The one-hot encoded vectors for all the search terms in a specific time frame are then added together as a combined vector. That vector is then normalized (adjusted to account for the total number of searches in that timeframe).

“While these approaches generated good enough representations to build flu models,” the researchers explain, “search terms that have many synonyms or common misspellings” get ignored.

Instead, SLaM relies “on a language model to handle misspellings, synonyms, and other types of related terms.”

Modeling with search representations 📊

“In terms of nowcasting and predictive modeling, to the best of our knowledge,” the researchers explain, “most of the downstream modeling on top of compressed search takes the form of linear modeling.”

To explain those terms a bit more, “nowcasting” refers to predictions of the present or very near future, while “predictive modeling” is the process of using historical data to predict future events.

“Linear modeling” refers to a type of statistical model that assumes a linear relationship between the input (in this case, compressed search data) and output variables (predicted real-world outcome). Though simple and interpretable, such linear models may not capture complex relationships in data.

“The only non-linear modeling that we are aware of,” they explain, involved “a Gaussian Process” used “in conjunction with autoregressive features.”

A “Gaussian Process” is a type of probabilistic model that can capture more complex, non-linear relationships, while “autoregressive features” refers to features that incorporate past values (historical trends) when making predictions, which is helpful for seasonal patterns or other temporal dependencies.

However, the researchers “believe there is room to improve the predictive capabilities beyond what can be achieved with existing approaches, while preserving the ability to interpret the model,” meaning understand how it’s making its predictions.

With the “search embeddings” from SLaM, “it is likely that there is interaction between the embeddings, which can be captured through non-linear models like neural nets,” meaning neural networks can spot more complex relationships in the compressed search data.

As for “the issue of overfitting,” the researchers address this “through regularization and inductive biases,” the specific assumptions and limitations we mentioned earlier, “and then validate that our model generalizes by reporting our metrics over a test set not included in the train and validation sets.”

4. Experiments 🚘

This is a longer section, so I plugged it into Gemini Advanced and asked for a summary.

Here’s my summary of Gemini’s summary:

The researchers evaluated SLaM Compression and CoSMo modeling using two real-world scenarios, including predicting U.S. flu rates and auto sales. They benchmarked their methods against existing approaches and conducted several experiments:

• Automotive sale predictions: They found their approach improved the accuracy of nowcasting U.S. auto sales, with a 30% improvement in model accuracy compared to traditional classification methods.
• Flu rate predictions: Compared to existing approaches, including autoregressive models, CoSMo performed on par or better, even without using lagged flu rate data.
• Model ablations: The researchers conducted ablation studies to assess the impact of different model components (search volume, regional multipliers, etc.); they found that including search volume as a feature significantly improved the national model’s performance, while regional models benefitted from including regional features.
• LM choice: The researchers experimented with different language models to generate their search embeddings; they found that Multilingual Sentence Encoder (MLSE) outperformed larger models (like sT5), likely due to its ability to handle multiple languages present in the search data.
• Zero-shot inference: The researchers also tested the model’s ability to make predictions for different geographic levels than what it was trained on, discovering it could infer state-level flu rates with some success, even when trained on national-level data (and vice versa).

There are some interesting findings in there, particularly the points about the ablation studies (features for national vs. regional models) and LM choice (factoring in multiple languages).

5. Model interpretability 👀

Interpretability refers to the ability to understand how a model makes its predictions. This section has some cool figures in it.

For example, “Table 7 shows some examples of terms that score high, meaning the model believes queries near those areas of the embedding space are highly predictive of flu cases”:

We can see how it accounts for misspellings of “influenza,” and it makes sense terms containing “flu” would score lower than a query like “do 5 year olds get immunizations.”

There’s also Figure 5, which shows a visualization of auto search terms, which are clustered based on their embedding vectors (semantic similarities) and with their estimated impacts from model predictions shown:

Looking at the data, we can imagine why a query like “Audi financial services” would have a higher impact (on real-world outcomes) than a search term like “what does GMC stand for.”

Traditionally, we as SEOs might depend on SERP analysis, third-party tools that apply search intent labels, or our common sense to make those judgments.

Imagine having data like what SLaM and CoSMo could provide. 🤤

6. Ethical use of data 🕊️

Speaking of how this data could be applied, let’s touch on ethical data use.

With regard to modeling the flu, the researchers note how their method “is not a substitute for traditional disease reporting.”

As for data privacy, they explain how the search data in their project “was anonymized, with none of the queries associated with any individuals or accounts.” And since they aggregated “individual search terms in an embedding space, we believe our method only increases the privacy of the search data.”

Let’s now check out their conclusions, before drawing a few of our own.

7. Conclusion 🏁

“It has been established in the literature that search data can add efficacy to predictive models,” the researchers explain.

“With billions of Google searches each day,” they write, those search logs contain “valuable signals on everything from flu prevalence to auto brand sentiment.”

As for how their approach is different, “Until now the typical implementation of search data into predictive models has been through incorporating coarse Google Trends data, similarly aggregated data using binary classifiers, or through complex filters for including individual search terms, which leaves the downstream models with either a diluted signal or signals prone to overfitting.”

Meanwhile, SLaM can “include search data in a privacy-safe manner by using the embeddings from language models to create a summary of the search data,” which “allows us to retain much of the information about the queries themselves as well as their relative volumes while greatly reducing the dimensionality.”

They also note how beyond Google Search, their “search compression can be applied to any scenario where statistics associated with natural language need to be summarized to a fixed-length vector.”

As for CoSMo, a constrained search model, it “has inductive biases that greatly improve the accuracy of our models built on search data,” as shown in the experiments for estimating flu rates and auto sales.

“Finally,” the researchers explain, “we demonstrate that our models, despite being highly non-linear neural networks, offer interpretability that explains what terms are related to the variables of interest.”

In short, we can read through the above information now and understand what’s meant by SLaM, CoSMo, as well as terms like “fixed-length vector” and “summary of the search data.”

And that’s pretty 🆒!

But why should SEOs care about SLaM and CoSMo (probably)? 🧐

Well, in general, I think understanding how these approaches work can give us insight into how search engines like Google might use contextual embeddings in vector spaces to classify words (and more) based on their semantic meaning.

More specifically, we can see how Google is pursuing advancements in understanding the relationship between a search query and resulting outcomes in the real world, enhancing its understanding of the nuances of search intent. That trickles down into how Search works and SERPs are presented, as well as what content is relevant for those queries.

For example, maybe the most relevant page isn’t selected based on its topicality but because it fulfills the next step in a user’s journey to complete a specific action.

And on the practical side, if Google can leverage language models to embed search data and draw conclusions about user behavior, why couldn’t we SEOs do likewise?

Imagine how much better our content planning could be if we looked at a query beyond “informational,” “transactional,” or “commercial” intent but instead knew which search terms were most closely assigned with real-world actions.

Since we’ve had a long conversation with Gemini Advanced about SLaM and CoSMo, let’s ask what it predicts their significance to be, from an SEO perspective.

*Note: this is a theoretical exercise now, not predictions or instructions. 😊

1. Improved understanding of search intent ⭐️

Approaches like SLaM and CoSMo could enable Google to better understand the nuances of search intent by analyzing the semantic meaning of search terms and their relationship to real-world outcomes. This could lead to more accurate and relevant search results.

2. Enhanced personalization 👕

The ability to analyze individual search patterns and link them to real-world actions could allow Google to further personalize search results, such as by further tailoring results to a user’s search history, demographics, location, or other relevant factors.

3. Predictive search features 🤓

These approaches could be used to develop predictive search features that anticipate a user’s needs and proactively provide suggestions, such as relevant products or services, even before a user types a query.

4. Content strategy ✍️

By identifying search terms that are highly associated with specific actions, SEOs could create content that’s more likely to drive conversions, including knowing which long-tail keywords most move the needle in terms of audience goals.

5. Measuring real-world impact ✅

Tracking the relationship between search terms and real-world outcomes could enable SEOs to better understand how their content influences user behavior and drives conversions.

Well, all right. 😎

Outro

I hope you’ve enjoyed this week’s Hamsterdam Research article! 🐹

Feel free to comment below or contact me with your feedback.

Stay tuned for another new article, hopefully next week, or check out related posts below.

Until next time, enjoy the vibes:

Thanks for reading. Happy optimizing! 🙂