Yearly Archives: 2023

The first question obviously is, can LLMs work as product recommenders themselves?

The short answer is: not really, no. Recommender systems develop an understanding of customer engagement with products (purchases, cart/wish list addition, browsing) and use it to recommend products to them. Strategies to do so include finding customer lookalikes, using co-purchase patterns, trending products, content-based filtering etc. However, all these methods are necessarily trained on data about what customers do.

LLMs, on the other hand, are trained on very large corpora of public domain text data; their purpose is to understand and generate natural language at a level comparable to humans. What this means is that, once you’ve decided what to recommend to your customers, LLMs can help you wrap the product recommendations within an appropriate selling story.

Pro tip: You must be aware of the benefits of customer lifecycle management & how you can use predictive modelling software for better targeting your customers & enhance customer lifecycle.

Customers don’t approach purchases the same way every time. For instance, if you’re looking for a restaurant to visit, sometimes you want to try something completely new, sometimes a different restaurant that serves your favourite cuisine, and sometimes you just want to go back to your favourite place once again. It’s the same with selling stories for product recommendations.

You can, of course, tell your customers: “Hey, we’ve got some recommendations especially curated for you!” or you can filter these curated recommendations through various lenses, like: “Hey, we know you love our Men’s wear collection, so here are some recommendations for you” or “Have you tried out our new line of beverages?” or “You’d like the usual once again, sir? Very good”.

Here’s where LLMs come into the picture. Since they are capable of producing very good, human-readable messages, they can write your marketing copy for you in all of these cases. You could even use something like Midjourney to work out what your creative image needs to look like, in case you want to embed these in an email.

But here’s where it gets even more interesting. Suppose you want to run a standard marketing campaign to give a 20% discount on all items in a particular category, say Men’s wear. Rather than send the same message to everyone, you could break your target audience down into two groups: people who have bought men’s wear before, and people who haven’t. You could then ask an LLM to come up with marketing copy, based on your original message, to personalize the message to these two groups. The following is an actual interaction with ChatGPT to accomplish the above task.

These simple variations in product recommendations allow you to take a generic category promotion and make them a lot more personalized and engaging for a better customer experience. The effort involved in getting these variations done simply boils down to telling ChatGPT what to do. You can even provide it with sample messages from your brand and ask it to mimic your preferred tonality.

In conclusion, the collaboration between Large Language Models (LLMs) and product recommenders brings a new dimension to the realm of personalization & machine learning in marketing. While LLMs themselves may not function as product recommenders in the traditional sense, their remarkable capability to generate human-readable content adds a layer of sophistication to the presentation of product recommendations.

By leveraging the strengths of both LLMs and product recommenders, businesses can craft smart shopping campaigns around their offerings, tailoring the messaging to different customer preferences and engagement patterns. Whether it’s suggesting new items, highlighting favoured collections, or introducing special promotions, LLMs have the potential to infuse creativity and personalization into marketing campaigns.

Moreover, the marriage of LLMs with customer data enhances the customer lifecycle journey by creating distinct marketing copies for diverse audience segments, effectively turning a generic promotion into a collection of individualized messages. Harnessing the power of LLMs for product recommendation marketing not only streamlines the process of message creation but also empowers brands to establish deeper connections with their customers by resonating with their unique interests and purchasing behaviours.

Incrementality has been hailed as the gold standard in measurement for marketing. In the BCG-Meta “Measure to Grow,” report measuring incrementality is seen as the ideal for measuring the impact of interventions and spending, even though the ease of use is lower than click-based metrics. However, despite its significance, there are nuances to its measurement that can make it challenging for business users to extract the right insights. In this article, we will explore the intricacies of measuring incrementality and shed light on some key considerations that marketers should keep in mind.

How Incrementality Is Measured

To measure incrementality, a target group (TG) is identified to receive a specific intervention, such as an email campaign, while a control group (CG) is withheld from the intervention. By comparing the conversion rates of the TG and CG, the delta between the two can be calculated as the incremental conversion. This incremental conversion, when multiplied by the target base and revenue per conversion, provides a measure of the incremental revenue generated.

You might also like to read about The Ethics of Using Machine Learning in Marketing: What You Need to Know.

Some Challenging Nuances of Incrementality Measurement

Short-Term vs. Long-Term

It can be conducted in the short term or the long term. Short-term control groups are held out specifically for a campaign, while long-term control groups are held out for an extended period, typically ranging from six months to a year. Both approaches offer valuable insights. Short-term control groups help evaluate the impact of withholding a specific communication, while long-term control groups provide insights into the cumulative effect of not receiving a series of communications. While both approaches have their merits, long-term control groups often offer more actionable insights.

Handling “Unsuccessful” Campaigns

Occasionally, smart shopping campaigns may not yield the desired results, with the control group outperforming the target group. In such cases, convention dictates labelling the campaign as “unsuccessful” and setting the incremental revenue to zero. Reporting negative impacts at the campaign level is counterproductive since it does not provide meaningful information. Instead, considering the campaign’s impact as neutral helps maintain its integrity.

Granularity of Measurement

The level of granularity chosen for measurement can significantly impact the calculated lift. Whether measured at a campaign level per day, campaign level per month, or segment level per month, the decision affects the resulting incremental revenue. A more granular approach, such as measuring at a trigger level per day, tends to yield higher incremental revenue values. However, it is essential to strike a balance between granularity and practicality, as excessively granular measurements may become unwieldy and time-consuming.

Non-Comparative Nature

When assessing incrementality, it is crucial to recognize that trigger-level lift and segment-level lift may not align when measured separately for the same period. Different data cuts, such as triggers, segments, revenue centres, or periods, can influence the calculated incremental revenue. This can lead to complex discussions as stakeholders attempt to reconcile incrementality numbers that may not precisely match.

Reading about the Benefits Of Customer Life Cycle Management: How It Can Improve Your Business can also be quite helpful too.

Unreal Numbers

While incremental revenue can be quantified, isolating the specific transactions that contributed to it is challenging. Drill-down analyses, such as identifying incremental bills, customers, or products, are often not feasible. Incrementality focuses on the collective impact of interventions rather than pinpointing individual transactions. Therefore, it is essential to consider the overall incremental revenue rather than fixating on isolating specific components.

Impact of Intervention vs. Selection

Most incrementality metrics primarily measure the impact of interventions—targeted and relevant messaging over some time. However, it is crucial to distinguish between the impact of the intervention itself and the impact of selecting the right audience to target. If predictive modelling is used to identify the best customers for targeting, and the control group consists of similar high-potential customers, the lift observed is primarily attributed to the intervention rather than the selection process.

Conclusion

Understanding and leveraging incrementality effectively requires a thorough examination of its intricacies. Superficial analysis of reported incremental metrics may lead to confusing results upon deeper investigation or contextually inappropriate measurements that fail to provide true insights. At SOLUS, we have invested years in studying it and have incorporated practical and usable approaches into our methodology. Reach out to us to gain a comprehensive understanding of incrementality and lift measurement.

In today’s competitive landscape, personalisation has become the watchword in marketing. In order to treat each customer as a unique individual with their own preferences and interests, we must learn from individual customers’ behaviour rather than aggregate this understanding at a segment or population level. We want our marketing decision models to make decisions not just based on statements like “53% of our customers bought Men’s Formal Wear in the past week”, but on statements like “Customer XYZ has shown a preference for Casual Wear over the past year, but has purchased Formal Wear on his last visit.” And this is exactly where machine learning in marketing takes centre stage.

Understanding The Context

Machine Learning is a highly popular branch of AI that involves constructing data-driven models to comprehend or simulate human decision-making. (FYI, much of the recent popularity of AI has to do with a class of algorithms under machine learning called deep neural networks.) Machine learning models are widely used in modelling and predicting customer behaviour. The importance of machine learning in marketing can be observed in the fact that large, state-of-the-art recommender systems in popular apps and websites like YouTube and Netflix are often powered by ML. To be more accurate, it is possible due to the application of a machine learning-based recommendation system.

However, it is important to remember that this is made possible by data-driven models. So it follows that the more data you feed into these models, the better they get. But the term “more data” doesn’t always mean one thing. It could mean having data about a larger number of customers, thereby making sure that the models learn from a more diverse variety of customer behaviours, and therefore get better.

While we are on the topic of machine learning and what it can do for the field of marketing, how about reading The Benefits of AI in Marketing: Increased Efficiency and Personalisation for insights on that subject?

The Ethical Aspect of Machine Learning In Marketing

An alternative manner of viewing the data-driven system of ML is by considering that it could also be interpreted as having more in-depth data about each customer. And this interpretation is particularly relevant when you think about how you intuitively think about personalisation. The intuitive, one-line pitch on personalisation is: I know my customer inside-out, and am therefore able to recommend the very thing he/she needs, at the right time.

However, this notion of a deep understanding of customers raises several ethical considerations. The quest to gather more detailed data about each customer nudges us closer to a critical question: what data is ethically ours to use and learn from, and what isn’t?

It’s worth noting that data collection and learning from it aren’t inherently problematic. For instance, if a customer purchases several blue shirts over a period, it’s reasonable to infer that they favour that colour. Where it gets murky is when the emphasis on personalisation begins to infringe upon the boundaries of personal data privacy.

To simplify, any data that characterise a customer’s interaction with a brand—such as purchase history or browsing behaviour—is considered fair game. On the other hand, any data that describes the customer personally and hasn’t been obtained with explicit permission should be off-limits. Moreover, any practice that denies customer access to a brand’s products or services based on their characteristics is unacceptable.

You might also benefit from reading our article on the Power of Customer Insights: Unlocking Hidden Opportunities for Your Business.

Ethical Machine Learning in Marketing—A Path Forward

All the ethical notions discussed thus far are neither surprising nor for that matter, unregulated. There exist legal safeguards to dictate what brands can and cannot do with data. However, in day-to-day operations, it’s easy to assume that if some data is available, it’s okay to use it. Hence, when applying machine learning in marketing, it’s vital to adhere to ethical standards and respect personal boundaries.

In conclusion, harnessing the power of Solus AI’s campaigns for personalized recommendations, driven by cutting-edge technology, is a game-changer for any marketing strategy. By leveraging this intelligence system, businesses can unlock a plethora of benefits and enhance their marketing efforts like never before.

2022 has been an interesting time for folks in tech– and as you well know we’ve pivoted so much to being a tech-first organization that suddenly everything happening in tech impacts us too! We shifted to being product-first in 2020, and the roller coaster ride that followed is worthy of a small book – or lengthy blog post at the very least. But that’s for another day. Today I’d like to touch upon one of the three main pillars of Solus AI. For context, the three central systems of SOLUS that work together to create magic are: Recommendations Systems | Smart Campaigns | Intelligence Systems

Intelligence Systems in Solus

Part of our entire pivot to being product-first meant that we had to go back to our “greatest hits” of analytics work done over the last decade and see if we could make them available to clients using Solus. But of course, better than before – configurable, always on, and far richer. One decision we did take however is to not create a dashboarding platform – we’re not trying to build a Tableau/ PowerBI, but we’re bringing in our experience of how one should look at CRM and campaign metrics, and making the best of that experience available.

I’m happy to share that the Solus Intelligence System module has matured beautifully over the last year. This is a quick recap of where we are, and where we’re headed with it.

You might like to read how the power of customer Insights helps unlock hidden opportunities for your business.

Customer Intelligence Systems Report

We’ve ramped up the Intelligence systems reports to over everything a CRM manager (or someone dealing with Retention or targeted campaigns) might need. We’re constantly adding to these reports, so while these are currently a set that would work for pretty much any industry, we’re next coming up with industry-specific reports too.

Campaign Performance Reports

In addition, the ability to track campaign performance has been greatly enhanced as well. We’ve focused on Incrementality measurement – which we all know is hard, but also the truest measure of impact, and what we’ve added is a whole lot of reporting that comments on what’s working, what’s the yield it’s generating, what consistently works etc.

A few things we’ve added to these reports that are small, but powerful:

• The ability to filter out triggers where the Control Group is too small
• The ability to switch between Trigger X Day and Trigger X Month level of granularity for Lift measurement
• The ability to make response attribution Time relative rather than Day relative

 There’s so much more, and we’re truly helping clients see beyond clicks and conversions, the true impact of their customer-centric efforts.

Roadmap Stuff: Campaign Diagnostics (and soon enough – What Ifs) 

We aren’t stopping here, of course. A big ask has been – can we help diagnose what campaigns work/ don’t work better? Can we understand what aspect of the campaign works? So we’re now coming up with a campaign diagnostic that does exactly that. What campaign attribute works? The segment being targeted? The use of a personalization engine? The timing? The presence of an offer? The use of a Recommendation? And so much more.

The obvious next step from here is adding a What If, going Prescriptive – recommending campaign types that give you lift and incremental, as well as campaigns that give you coverage and absolute numbers. Answering the Q of – “What If we run these types of campaigns in the coming month”. Getting to goal-seeks “IF I want to attain X, what should I do?” Watch this space!

Looking Ahead

Building Solus has been an immense endeavour. No part of this is easy – CRM is complex, it’s quite literally a space in which everything that can go wrong, will. Personalization and scale are notoriously hard. If you think about it, however, we are driven by a not-complex goal to help brands derive the most value from customer intelligence systems. This translates into – uplifting how brands nudge customer behaviour, whether the context is brand initiated (you nudge them by sending the customer a message), or customer-initiated (the customer comes to your app/ website and we need to decide what to show and how to engage them). We’ve had to break and build like crazy this last year, and the coming year will no doubt be the same. 

In conclusion, if you’re looking to unlock the true potential of your business and drive long-term success, an intelligent campaign & customer insights solution is the key. With our powerful AI-driven platform, you can gain invaluable insights into your customers’ behaviors, preferences, and purchasing patterns. By harnessing these insights, you can create highly targeted and personalized campaigns that will not only increase repeat purchases but also foster customer loyalty.