It looks like AI and machine learning is going to be a major theme for the search industry in 2018. There have already been quite a few announcements around shiny new AI projects in Adtech/SaaS (and we’re still only a month into the New Year). Across the board, I’ve noticed a trend in the number of enterprise tech tools dialing up the messaging around AI.
It’s great for the industry to be innovating, but AI and ML must be the two most misunderstood concepts in tech, and this isn’t being helped by the increasing amount of marketing buzz. There are a lot of trumpets being blown (and in some cases they should perhaps be blowing raspberries instead).
Machine learning has been an integral part of Adthena’s technology for almost six years, and as the Director of Product Marketing I’ve been exposed to a more than a few misconceptions around machine learning, and its application in search.
Here are the 5 most common misconceptions of machine learning in search compiled via feedback from prospects, existing clients and our in-house search experts at Adthena:
1. Expectation vs Reality
These are probably the most frequent, and understandably so, as it can be a bit of a grey area:
Expectation: A machine intelligence, making independent, ‘human-like’ decisions and delivering data and insights.
Reality: The perception that AI represents machine intelligences that can make decisions like humans do (i.e. self-driving cars) is partly true, but doesn’t really take into account how “the machine” comes to a decision. The real strength of AI (as it exists currently) is rooted in the ability to process massive non-numerical datasets, and arrive at a similar result to a human decision-making processes.
For example, self-driving technology takes data from numerous signals, cameras/GPS/map data, and interprets these based on learned algorithms to navigate a vehicle. Another example of this is virtual assistants, think how much better these have gotten in the last 10 years — it’s because having so much user data to interpret and analyse allows these assistants to give better answers and be much more confident in doing so.
Ultimately, there is still a huge gap between AI and the human thought process. If you consider the world’s most advanced Chess AI projects, the engines are really making a call on the best potential move, based on all available data (all the potential lines, permutations, and historic games the machine has investigated), and choosing the best option based on a balance of probabilities.
Expectation: Algorithms that get more intelligent and effective over time, or as it is ‘fed’ more data. Every time I use it, it learns more and gets better.
Reality: This is kind of true, but does tend to humanize the machine learning process. There are many different schools and techniques to machine learning (and sometimes it’s a combination of several or many of these that contribute to an artificial intelligence). Something that unites the many different types of machine learning is the use of large or massive datasets, sometimes used to ‘train’ the machine learning models. Machine learning models do improve or learn over time, but this is more due to the cumulative effect that an enhanced level data has on an output.
2. Doesn’t Machine Learning just mean ‘automation’ for Search?
Whilst machine learning can surface automated insights (i.e. Adthena’s Search Term Opportunities), it is a very common misconception that this is what machine learning is for. Machine learning can surface very powerful data insights, but the quality of these insights is dependent on the quality of the datasets being analysed.
And really, automation is often a byproduct of a machine learning process. The real goal is to refine and interpret complex and multi-dimensional and massive data using technology, to an extent that would be impossible to achieve using more traditional approaches.
A good example of where AI is translating into an automation by-product is with the AdWords DSA feature. Although I hear a lot of good around scaling advertisers campaigns, particularly in the retail space I have experienced the not-so-great results with our own campaigns. Triggering our brand on the term ‘machine learning’ is way too broad, clearly on-topic but not targeted in the right way for us to grow our campaign.
3. Will the Machine understand what is important to my business?
For Adthena’s purposes, the ‘machine’ so to speak, is an advanced data model that has been designed to reflect an advertisers competitive search landscape. It is the product of trillions of calculations and results, creating a model that is beneficial to advertisers. One example of this, is how Adthena uses known data to predict CTRs and CPCs for any client advertiser’s rivals, in effect reverse-engineering the SERP.
It’s a bit like the machine is creating all the quantities and data points that will be important to an advertiser. Ultimately, there will always need to be a skilled human at the end to make to most of this data to get the best insights.
4. How does your suggestion engine work? What will we see?
Adthena’s Search Term Opportunities surfaces search terms which have been scored for being the most valuable opportunities for any given client advertiser. It may for example, show you pure brand search terms that a rival competitor is advertising on but you are not, or generic search terms that your competitors have been seen to be advertising on in the past seven days.
It can also show you ‘Lone-Rangers’, the name we give to search terms where an advertiser is holding both the top paid and organic search listings, with no other competitor ad activity (in these cases, the paid ad is unnecessary). What machine learning enables is for these kinds of insights to be surfaced at scale, something that is necessary for enterprise advertisers managing tens or even hundreds of thousands of keywords.
5. How does AI/ML continue to learn on a limited data-set?
Machine learning and data have an almost symbiotic relationship. Each one improves the power and potential of the other. In search, the amount of data is exception and really quite limitless. Right now there is so much potential to enrich data, collect better data, or segment it further and more powerfully that ‘limited’ is not really a word being used very much.
AI and machine learning really is the future. Not just in search, but in a much, much broader sense. If it’s popularity continues to pick up the pace this year, it will be important for all of us to get a better understanding of the art and science of it AI/ML.