No one can perfectly predict the future, but it turns out some are better at it than others.
A study of over 1 million judgments from thousands of people found “super forecasters”: people who have significantly higher success rates at predictions.
What separates super forecasters from lousy forecasters? They change their minds gradually based on new evidence.
“People who are right a lot, they listen a lot, and people who are right a lot, change their mind a lot” is a quote by the man who owns the world’s largest ecommerce economy, the Washington Post, and most wealth on earth.
When people ask me what I’m changing my mind about, I give them two answers: voice and personalization.
The biggest change agent of our times – way bigger than the BlockchAIn – is AI. Large language models (LLMs) have the power to hold voice assistants and personalization to the promises they once made.
Boost your skills with Growth Memo’s weekly expert insights. Subscribe for free!
Don’t Call It A Comeback
The controversial “50% of searches will come from voice by 2020” prediction was blown out of epic proportions and repeated too many times because it made for a good, fear-inducing story.
Reality is that voice assistant devices are not doing so hot. They’re more speaker than assistants.
Investors got scared when Mark Zuckerberg announced to write off $10 billion a year to develop the Multiverse, but no one bats an eye that Amazon is losing the same amount of money on Alexa devices despite having sold 500 million of them.
Apple and Alphabet are in similar boats, boasting almost 100 million Siri and Google Assistant users despite the fact that they’re pre-installed on their vast hardware ecosystem.
In 2020, 54% of U.S. adults 18+ had used voice commands, 24% daily. But the majority of use cases of voice are “play music” and “what’s the weather today?”
My primary voice assistant usage is asking Siri to continue the podcast I’m listening to when I step out of the shower. Commands and searches are two different pairs of shoes.
And yet, there is reason for optimism. Just like LLMs can instantly improve the quality of search boxes, they can make voice assistants actually useful by better understanding user intent and giving more natural answers.
For ad platform giants like Alphabet, Amazon, and Meta, any moment users can’t see ads is a market opportunity.
Shower thoughts and requests are just one small instance. Americans spend an hour per day in their cars, an hour to make food, and half an hour to clean. Just a fraction of that time spent on conversations with AI could unlock significant ad or subscription revenue.
A well-working voice assistant would be an opportunity to not just make more revenue directly but also foster a new app ecosystem and sell more hardware, which Amazon is best positioned to do with their home security and other smart home products.
Just recently, Amazon had to walk away from a $1.7 billion deal to acquire iRobot due to regulatory pressure.
Amazon is already on the train and working on up-leveling Alexa to have better conversations.
Voice assistants need to overcome the challenges of speed and access. Even if an answer only takes a few seconds to generate, the experience doesn’t feel conversational.
Most importantly, the more access to information a voice assistant has, the more it can personalize answers.
The multimodal nature of LLMs could allow voice assistants to see and hear. Multiplied across your phone, computer, doorbell, security cameras, appliances, and cleaning robots, Google’s vision of ambient computing could come to life:
In the mobile era, smartphones changed the world. It’s super useful to have a powerful computer everywhere you are. But it’s even more useful when computing is anywhere you need it, always available to help. Now you heard me talk about this idea with Baratunde, that helpful computing can be all around you — ambient computing. Your devices work together with services and AI, so help is anywhere you want it, and it’s fluid. The technology just fades into the background when you don’t need it. So the devices aren’t the center of the system, you are. That’s our vision for ambient computing.
My Web Or Your Web?
Web personalization is the Blockchain of the early 2010s: a disappointment.
Ten years ago, we believed the future of the web was hyper-tailored websites and search results. But we got filter bubbles and Cambridge Analytica. Patronizable data is hard to collect and scale. Users like free products, but they don’t like the feeling of being tracked.
Google started personalizing search first in beta in 2004, and then globally in 2009. Today, Google personalizes results for a limited amount of queries like “what to watch,” events, previous searches and Discover.
The overall degree is negligible.
In 2013, researchers found that the average degree of personalization in Google Search was 11.7% – of course, varying widely by query and rank. Higher positions, for example, have a higher chance of being personalized than lower positions.
A study from 2019 found that Google personalizes 2/10 results when searching for people and 4/10 for political parties. In other words, not much.
The highest degree of personalization in search goes into Google Suggest: Based on millions of other searches from people just like me, Google suggests query completions that are sometimes plain scary.
However, Google started to offer more personalization for its search canary: fashion.
Users get custom style recommendations based on their choices and can add brands to favorites to customize their search.
In 2022, Google started customizing shopping search more based on previous purchases.
When you’re shopping on Google, just make your selections once — your preferred department and brands — to see more of each in the future. So if you select the “womens” department and the brand Cuyana, next time you’re shopping for something like a messenger bag, we’ll show you women’s messenger bags from Cuyana and similar brands.
Personalized shopping recommendations are much less scary than political news and easier to monetize because Google can shorten the path to conversion while keeping users in search.
Image Credit: Kevin Indig
Image Credit: Kevin IndigPersonalization is most useful in shopping, news, and local, but less for informational searches with a learning intent.
LLMs and machine learning paired with the Shopping graph of over 35 billion data points allow Google to tailor shopping search more than ever before – exactly what Google has been doing over the last months:
When searching for a product, you’ll get a snapshot of noteworthy factors to consider and products that fit the bill. You’ll also get product descriptions that include relevant, up-to-date reviews, ratings, prices and product images. That’s because this new generative AI shopping experience is built on Google’s Shopping Graph, which has more than 35 billion product listings — making it the world’s most comprehensive dataset of constantly-changing products, sellers, brands, reviews and inventory out there. In fact, every hour, more than 1.8 billion listings are refreshed in our Shopping Graph to give people fresh, reliable results.
“Hey Google, Where Is My Data?”
If voice and personalization ever have a chance to make it big, it’s now. Things rarely play out the way you think they do, which would perfectly fit the bill for voice and personalization.
If I’m right – and personalization and voice search do get bigger – the importance of being top of mind before an intent comes up grows even more than today.
On the backend, brands need to convince with superior service, shipping and returns to keep customers. The beauty of personalized shopping search is that once customers are settled on an opinion, it’s hard to get them off.
How would Google personalize results? A combination of first-party data from Google services like Gmail, YouTube, Android & Co., and Topics.
One factor I’m not bullish about is data.
Ever since flooding the SERPs with features like Map Packs, image carousels, shopping carousels, and other carousels, Google hasn’t given marketers much data to understand their impact. I also fear that Google wouldn’t give us data about SGE if it ever rolled out, but that’s another story.
What would change my mind is if I saw how we get the queries and number of searches from current voice devices, but neither Alphabet, Apple, nor Amazon share that data.
How the Best Forecasters Predict Events Such as Election Outcomes
Amazon has sold more than 500 million Alexa-enabled devices, drops 4 new Echo products
Alexa, why are you losing so much money?
Why did Apple’s original HomePod fail? Let’s count the reasons
Roughly 1 in 4 U.S. adults now owns a smart speaker, according to new report
Exploring time spent on cooking, reading, and other activities for National Hobby Month
Average hours per day spent in selected household activities
Amazon is set to supercharge Alexa with generative AI
Measuring Personalization of Web Search
What did you see? A study to measure personalization in Google’s search engine
9 new features and tools for easier shopping on Google
Supercharging Search with generative AI
Featured Image: Lyna ™