A robot with an awkward smile and possibly two thumbs

This is the fake it until you make it era for Marketing

The integration of machine learning into marketing platforms isn’t new, with many of us first encountering it through Google’s enhanced CPC (ECPC). Google’s system was designed to automatically adjust bids by learning which audiences, times, and locations yielded better results for advertisers. Initially, ECPC’s performance was subpar, but over time it learned (at the advertisers’ expense), and improved significantly. Its primary achievement was boosting Google’s revenue, though it fell short in addressing Google’s challenges in selling non-search advertising inventory. This gap was later addressed by PMax.

PMax represents a novel approach to machine learning. It operates on the principle that a mix of creative content is more effective than search-focused strategies alone. Google observed that Display and Video ads were undervalued in measurable revenue, offering poor standalone return on ad spend (ROAS). However, when combined with search and shopping, these ads contribute to a more acceptable blended ROAS. The biggest impact of PMax has been improving shareholder value at Google. For advertisers it’s been more of a mixed bag.

Initially, PMax faced significant challenges. In cases of insufficient account data, Google would blend data from similar advertisers. This approach could either be beneficial or detrimental, potentially causing a negative feedback loop in an account by continuously underperforming against ROAS targets. Today, with more data available for training, PMax has improved, although it may still be unsuitable for certain accounts, particularly in the small and midsize business sector. The lesson here is that there are advantages to being a late adopter of PMax.

Meta confronted a different issue. They had to reinvent their targeting methods due to the loss of critical behavioral data. Their solution, the Advantage+ machine learning system, relies on attention metrics to predict conversions, analyzing factors beyond mere clicks, such as scroll speed and ad engagement time. After approximately six months, Advantage+ became a dominant targeting strategy.

Machine learning has evolved beyond the “fake it ’til you make it” stage, proving beneficial for both marketers and platforms. However, generative AI in marketing is still finding its footing. Unlike machine learning, which has concrete KPIs to guide its optimization through numerous iterations, generative AI lacks a tangible anchor in reality.

This brings to mind the 90s comedy “Multiplicity,” where the quality of clones diminishes with each generation. A similar phenomenon is observed in AI: training an AI on the output of another AI leads to degradation in quality, a concept humorously termed “Degenerative AI.”

Copies of Copies makes dumb

A recent example is Grok, which was accused of being a derivative of OpenAI’s work. While its developers claim that Grok is original but trained on OpenAI data, leading to similar behaviors, skepticism remains, especially since trust issues plague statements from X.

At issue is AI’s dependence on fresh, real-world data for training. When AI is trained on synthetic data generated by other AIs, its capabilities quickly deteriorate. The growing reliance on AI-generated content on the internet exacerbates this problem, potentially capping AI’s potential for improvement.

Despite these challenges, generative AI has found successful applications. For instance, Midjourney and ChatGPT are useful for generating preliminary images and article outlines, respectively. In e-commerce, Google’s Search Generative Experience has been adept at summarizing products, a technique adaptable to online shopping filters. This approach was effectively utilized in https://tirerobot.com, a website that leverages generative AI to refine tire searches based on specific vehicle generations and use cases, surpassing traditional filter capabilities.

While generative AI has made significant strides in certain domains, its future in content creation and art may be reaching a plateau. As real, diverse data becomes scarcer, both for humans and AI, we might witness a gradual decline in the quality of AI-generated content.