How large-language models are changing ecommerce

When you need quick advice about buying something online before doing a search, what do you do?

The answer might be to start a conversation with a ChatGPT-style chatbot, an assistant with access to a large-language-model (LLM). You know, that artificial intelligence–powered helper who greets you from the lower right corner of your screen.

Maybe you need gardening gloves that last longer than a week. You ask the bot which brand’s the most durable.

It gives you a brand name. Fair enough.

The chatbot doesn’t give you a link to a product detail page (PDP) for these highly rated gloves, so you head to the search box to enter the brand name.

Hmm; no results…doesn’t the site sell this brand?

Promising technology with a few caveats

This scenario exposes a couple of drawbacks of using an LLM feature for shopping on an ecommerce site, which we’ll look at in a minute. There’s a promising upside to these AI models, though: when GPT teams up with a search engine on an ecommerce site, the customer experience is likely to be enhanced. Yep, there’s indeed a place for genAI in ecommerce. According to the Harvard Business Review, “for those businesses that can successfully make use of GenAI to reach their business goals, the rewards can only be both promising and huge.”

In line with that sentiment, McKinsey is forecasting the potential economic impact of generative AI to range from $2.6 to $4.4 trillion globally. The growth is expected to be driven by use cases such as efficient content creation, optimized data use, enhanced product discovery, AI-infused search functionality, and streamlined personalized shopping.

Improve pretrained LLMs banner

Language models loom large

LLMs (and one particular application of them, generative AI chatbots) are the current darlings of the tech world, and many ecommerce executives are jumping on the genAI models bandwagon in search of website optimization. With a little help from vectors, semantic understanding, and personalization features, LLMs are driving a conversational AI revolution taking place the world over. And while there are still a few glitches to be worked out for real-world success, companies of all sizes are wading in to see if they can score big with this technology.

Really, genAI in ecommerce?

The core challenge for every ecommerce business remains the same: increase conversion by increasing product selection and lowering costs. Successful AI deployment in ecommerce requires seamless integration with the user experience (UX).

So are large language models applicable and potentially fabulous, inappropriate and a possible threat to the bottom line, or a little of both? The short answer: LLMs work well for certain tasks but not for others, and their strengths make them potentially great team players for generating revenue in the online arena.

As was the case with other tech innovations, many execs have been considering ways to apply generative AI. But if you’re like some businesspeople, you may find it a head scratcher to talk about genAI and ecommerce in the same sentence. After all, ecommerce site shoppers have some nonnegotiable requirements, namely the ability to see multiple purchasing options, all with 100% accurate, contextually relevant product information.

In the context of search, LLMs aren’t there yet. So far, they have a reputation for doing some eyebrow-raising things: there’s that hard-to-ignore possibility that they could hallucinate. They’re also prone to release whatever information they are trained on, so they could share proprietary information if the company has given them access to it. For a website predicated on giving shoppers the correct information on their buying options, even one or two slip-ups in this regard could spell trouble.

So continuing experimentation and fine-tuning are needed to unveil the full potential of genAI for online commerce. Enabling more-dynamic applications that can react to users’ needs in real time is key, says Bharat Guruprakash, chief product officer at Algolia. 

Which LLM apps work well in ecommerce?

As always, we can look to Amazon for the latest experimentation with genAI tools. The company has debuted a chatbot, Rufus, with an LLM trained on the product catalog and customer reviews, to answer user queries at various stages of the shopping journey. So far, reviews for Rufus have been decidedly disappointing

That doesn’t mean LLMs are entirely wrong for ecommerce, though. The problem with many current generative AI applications is that companies are using them to replace existing, working technology instead of simply augmenting online shopping experiences. In ecommerce, alongside mature solutions, such as vector database technology, they can be very useful. They’re making inroads in these areas, where shoppers are finding them more enjoyable to use:

Customer support

“Will this be back in stock soon? If not, what do you recommend?” Shoppers often have decision-making questions before clicking Buy. As an ecommerce seller, it pays to offer excellent-quality, personalized interaction in real time. What happens, though, if your customer support team is understaffed or offline?

If an LLM using natural language processing (NLP) techniques is provided with the relevant information necessary for question answering, it can meet shoppers’ needs in an economical way. For example, consider someone who’s had previous contact with Support. When they go back on the site and ask the chatbot some follow-up questions, it can draw on the history of the interaction and be aware of any earlier context, then prospectively respond in an coherent way.

Content creation

Just as with other wide-ranging marketing needs, LLMs can quickly do text generation for product descriptions, blog posts, email campaigns, and social media content. Or even all of it at once, adjusted for the channel and tone. One task that’s ideal for an LLM to automate is crafting engaging purchase confirmation email messages that also upsell related items. GenAI can handle this easily if it’s paired with a dedicated recommendations model designed for choosing related items. 

Generative AI can also be used to write product descriptions in advance, which can then be displayed alongside products. An ecommerce site could take this idea even further: with the help of user data, generating product descriptions and individually personalized marketing content on the fly. Imagine the effect that would have on a potential customer. The site would feel less like a static list of products and more like a personal shopping aide. 

Product discovery

LLMs excel at answering questions like “What’s in style this summer?”. Paired with a recommendations algorithm to select relevant products, an LLM could summarize all that fashion information on a customizable category page with built-in styling suggestions.

And for differentiating brand strengths and conveying an authoritative voice, an AI chatbot steeped in a given niche could represent a specific role in the product or service buying journey (e.g., an engineer in the B2B arena). One example of LLMs focused on niches is crewAI, an open-source project with an enterprise upgrade option. An LLM specializing in a particular role can be invaluable for helping shoppers figure out what they’re searching for, even adjusting the query if the AI detects that doing so would be helpful.

Responses to informational queries

A generative AI bot can draw on available information to give shoppers paragraph-like answers to questions. For example, imagine you’re on a clothing site, trying to find a dress shirt you can wear cufflinks with. However, you can’t remember what they’re called. How are you supposed to search for that kind of shirt?

First, you need to jump off of the retailer’s site and do a Google search on the name of the type of shirt (aha, a “French cuff shirt”). However, in addition to answering your question, Google Shopping shows you cufflink-compatible shirts from multiple vendors. You might click on one of those results, meaning the site you were just on has probably lost your business. 

What if, instead, staying on the original retailer’s site, you could query “type of shirt that goes with cufflinks”? Above the search results, you’d get an LLM-generated answer: “The type that goes with cufflinks is a French cuff dress shirt; here are a few options.” And below that, you’d get a list of search results, except that the LLM have modified the query to “french cuff shirts” so that the results are closer to what you wanted to search for. You’d feel like it had read your mind.

Does this sound too good to be true? It’s actually possible right now for an online retailer to set up this type of shopping experience. The trick lies in not relying too much on an LLM but in putting it to work in conjunction with a mature search solution. Using genAI as a partner with other, more mature AI technologies is a good bet to accelerate many companies’ success.

Are ecommerce execs aboard?

Some online retailers are not yet all in with genAI, according to Future Commerce. One stumbling block is trying to set up an LLM so that its output can be trusted, and that’s a daunting task. Another is that many execs haven’t quite figured out how to implement it to substantively impact their bottom lines. However, notes the study, “People are generally optimistic about their professional futures; the benefits of AI experimentation outweigh the potential drawbacks.” Sixty-four percent of those surveyed believe AI will “elevate customer experiences AND deliver a competitive advantage.”

Why normal search isn’t going anywhere

Word on the street is that if used responsibly, along with proven tools, LLM technology can improve the ecommerce user experience and boost revenue. However, if the technology is not used in a smart way, there’s a risk of doing lasting damage to the user experience and one’s brand image. Many companies are trying to use LLM technology to fully replace, instead of augment, traditional search interactions. That’s a bad idea because LLMs:

Can’t use structured data consistently

GenAI can’t provide traditional search results, as it’s not designed to come up with comprehensive, accurate results from a limited dataset, such as a product catalog. An LLM works in an interesting way; it’s a huge type of neural network trained to “predict the next word”. Given a prompt, it looks at the context and just guesses the most likely next word, so it can’t consistently incorporate structured data, such as details in a product catalog, in its responses. 

 For example, a product record for a pair of jeans might contain information about what other pieces of clothing go well with them and how each style fits. Given this type of context, an LLM could put that information in paragraph form, but only on the page for the item. If someone on the home page were to ask the chatbot for jeans that fit a certain way, it wouldn’t be able to parse the entire product dataset and recommend a SKU. A search engine, on the other hand, could let the user filter by parameters such as category, size, and fit.

Don’t possess search features

Ecommerce experts will tell you that when it comes to consumers tracking down products they want, a conventional search experience is best over all. Why? People need to explore and filter all their options, which isn’t something that lends itself to the conversational interface of a chatbot. LLMs are cool, but they lack the features people expect with search.

The single-answer format is no match for what a shopper can pull up with a finely tuned enterprise search tool. “The lack of context and multiple choices is inescapable within a pure chat context,” says Algolia search expert Dustin Coates.

With search, you get a priority-order-ranked list of results that includes product details and filtering options, with no missing information or inaccuracies. So while an LLM-backed chatbot can be helpful for finding general information about, for instance, a type of product, it’s inadequate for displaying buying options from an ecommerce product catalog.

Can produce inaccurate answers

LLMs make stuff up. “AI hallucination is a phenomenon wherein a large language model (LLM) — often a generative AI chatbot or computer vision tool — perceives patterns or objects that are nonexistent or imperceptible to human observers, creating outputs that are nonsensical or altogether inaccurate,” explains IBM. They’re predicting the most likely next word in a sentence, but they can’t verify whether that word is correct.

This can obviously be disastrous if your whole buying experience depends on genAI. If a chatbot confidently spouts inaccurate information, how can a business expect its shoppers to stick around? Ecommerce customers have proven to be skittish when faced with relatively less irritating problems like slow page-load times, so with something this monumental, they could easily abandon ship. 

Chatbot responses also can vary considerably with specific phrasing of questions. When asked a question in a certain way, the LLM may not peg the context or give the desired answer, but with slight rephrasing — the equivalent of coaching it to think differently — it may get closer to the mark.

Of course, search should simplify the shopping process, and people shouldn’t have to patiently coax a chatbot to come up with the right answer. Relying on an LLM — even a relatively proficient one — to help shoppers with basic needs is akin to making them stand outside a brick-and-mortar store and tell a child what they want purchased. Even if the child succeeds in bringing back the right item (and remembering to get your credit card back), the experience is laughable.

Get the best of both worlds

Are you cautiously optimistic about the potential of AI tools? Want the best toolkit for your ecommerce site: reliable AI search plus easy AI chatbot functionality?

At Algolia, that’s what we provide. Our brand new Generative Shopping Experiences are unleashing state-of-the-art solutions for product comparisons, reviews, buyers’ guides, and query refinement. Get in touch today and we’ll help you use our API to make your ecommerce business a powerhouse.

About the authorCatherine Dee

Catherine Dee

Search and Discovery writer

Recommended Articles

Powered by Algolia AI Recommendations

What does it take to build and train a large language model? An introduction

What does it take to build and train a large language model? An introduction

Vincent Caruana

Vincent Caruana

Senior Digital Marketing Manager, SEO
How and why events drive search ROI

How and why events drive search ROI

Jaden Baptista

Jaden Baptista

Technical Writer
Ben Franz

Ben Franz

Sales Engineering and Product Leader at Algolia
What is ecommerce merchandising? Key components and best practices

What is ecommerce merchandising? Key components and best practices

Catherine Dee

Catherine Dee

Search and Discovery writer