Second Order Effects of AI Search: From Keywords to Intent

The way we search is changing rapidly. For years, we’ve learnt to translate our thoughts into keywords, to make it easy for Google or Amazon to find the products we’re looking for. But of late, multiple websites have reported a doubling of their average search query length and an increase in unique queries, likely a result of tools like ChatGPT & Perplexity encouraging us to search conversationally.

This means less queries that look like “cross-trainer shoes” and more that look like “shoes that are good for hiking and football”.

This is because AI tools can now understand query intent, cross-reference it against a catalog/database and come up with a reasoned judgement for why a result might be a match. Contrast this to the old world of naively matching words in the query to words in the product — it’s the difference between using the ctrl + find option on a word document and asking the author of the document a question about its contents.

Second-Order Effects

This implications of this tech for commerce is meaningful:

1. Discovery & consideration search: Not everyone knows what “cross-trainer shoes” are (FYI, they’re what you go for when you’re looking to use the same shoes for multiple activities). Those who don’t, currently spend time in “discovery” mode, watching videos, reading articles or just scrolling through shoe listings with indecision. Or they got lost in “consideration”, i.e. is this particular shoe shown as a top search result good for my needs and is it the best choice for me? Eventually, this ends with the user failing to make a decision due to confusion, returning to Google/Amazon/TikTok & buying from a different store or buying the wrong product leading to returns down the line. With AI search, we get to a world of faster decisions, lower bounce rate, better conversions and fewer returns.
2. Queries give you rich user insights: With a query like “cross-trainer shoes”, it’s hard to tell what the user’s underlying need is. With “shoes that are good for hiking and football”, you can infer that the user is into hiking and football specifically. For category managers whose current alternative is to look at opaque aggregate numerical metrics (e.g. “80% abandoned cart rate”), a new set of filters to break down behavior by query intent attributes can help better inform business decisions.
3. Personalized content marketing: When you know that the user is into hiking and football, you can target them with emails, ad copy and promotions that play to these specific interests. As a bonus, LLM technology can help you not just understand what to do, but go ahead and do it for you.
4. Hyper-personalized on-site experiences: The more you know about each user, and the better that search technology gets at factoring in this information, the easier it becomes to unlock a truly personalized experience. This could include highlighting product attributes that the user may like (e.g. displaying a shoe’s anti-slip properties given that trekking is the use-case) or allowing continuity across sessions (e.g. if the follow-up query is “nike”, then favor showing Nike cross-trainers over any other Nike shoe). Unlike previous generations of AI personalization models, which underperformed for new users and even then were naive at best, new-age AI model can deliver logical personalization even with very little starter information about the user.
5. Chicken-and-egg for AI adoption: Users will attempt AI queries only if they believe that the search they’re using can support such queries. But businesses will invest in AI search only if it seems like this is what users want. Some businesses will react early to add this functionality to get ahead of the market, whereas others will react to changes in user behavior before taking the plunge. The former could unlock a meaningful first-movers advantage.
6. Merchandising moves from tactics to strategy: Currently, if a merchandiser needs to push specific high-margin or promotional cross-trainers to users, then they might “pin” or “boost” chosen products to popular search terms and their variants (“cross-trainers”, “cross-trainer shoes”, “men’s cross-trainers”, etc.). But when queries start becoming one of a kind, variants and intent become harder to pinpoint. It seems likely that simplistic rule-based configurations will give way to higher-level AI controls (e.g. instructions to “boost Nike Metcons when users search for cross-trainers”). The rise of such techniques can in turn unlock more powerful forms of merchandising like persona based promotion (“favor showing Pumas to teenagers but Nikes to adults”).
7. SEM & SEO undergo structural changes: Placing ads on popular keywords (e.g. “cross-trainer shoes”) starts becoming less effective if consumers start searching using one of a kind queries. The share of popular keywords may reduce, and one-off “long-tail” queries may become the norm. How large platforms like Google and Amazon react to this, and whether anybody else steps up to become a search platform, remains to be seen; suffice to say though that this could reshape the search advertising industry.

Change is at the doorstep

This isn’t just a latent trend. In multiple conversations with insiders over the last month, I’ve heard about quantitative signs of these trends. These changes have played out in the public eye too:

• Walmart has already started introducing AI-powered search features. To quote a recent press release — generative AI in retail is particularly exciting as it can help usher in a new way of shopping; shifting from “scroll searching” to “goal searching,” which makes the digital shopping experience more seamless and intuitive.
• Amazon Alexa has announced an AI overhaul for 2024, the backbone of which is search that can support conversational / voice based commerce.
• Perplexity.ai (funded by Amazon’s founder Jeff Bezos) has become a darling of Silicon Valley in leading the charge against Google in general purpose non-ecommerce search.

If you run an Ecommerce store and wish to join the party, do get in touch. At Branch AI, we build Ecommerce-specific LLMs for query rewriting, result ranking, query insights, catalog enrichment and embeddings. We train these on publicly available data across all of these tasks, which implies that you can use them out-of-the-box with low effort and with high performance across new products, categories, user demographics and query types.