This is a perspective on the advertising possibilities discussed in 2022. I have kept that historical framing rather than presenting old predictions as current news.
Targeting is a prediction problem
Advertising systems can use observed signals to estimate which impressions or audiences are more likely to meet an objective. That does not mean the system knows an individual’s intent. A page visit, purchase or demographic label can be incomplete, outdated or a poor proxy for what matters.
The practical question is whether a particular system improves an agreed outcome within appropriate constraints. A higher predicted response rate is not the same as more incremental sales, and more precise targeting is not automatically a better customer experience.
Personalisation needs a reason
Showing different content to different visitors may help when it answers a real need: an appropriate product recommendation, relevant information or a useful next step. It can also feel intrusive or repeatedly show people things they have already bought.
Before personalising, decide which data may be used, why the variation helps the reader and how someone can exercise relevant choices. Then compare the experience with a simpler baseline. Claims that personalisation necessarily creates trust or loyalty go beyond what a generic description can establish.
Automated bidding still follows an objective
A system can adjust bids or delivery using its configured objectives and feedback. That shifts part of the execution work away from manual changes. It does not settle whether the objective is economically useful, whether measurement is reliable or whether the campaign is causing additional business.
A team still needs budgets, exclusions, monitoring and a way to investigate changes. Returns depend on factors including the offer, creative, competition and data. Test changes in ROAS against an appropriate baseline rather than expecting improvement from AI alone.
Creative assistance is different from product evidence
Generating variations or reviewing a draft can make creative exploration quicker. The complete workflow still includes selection, editing and checking the result. Brand requirements, product details and factual claims need independent review.
A useful distinction is between an illustrative concept and evidence. A synthetic scene can explore an idea; a generated before-and-after image cannot establish that a product works. The original GAN research describes one generative approach, but not every advertising or language tool uses that architecture. I discuss that boundary in GANs in advertising.
The risks belong in the evaluation
The same tools can scale misleading copy or reproduce biases. Automation can change people’s roles and create new review work. These are reasons to inspect the actual workflow and its consequences rather than assume that every task should be automated.
What I would carry forward
Choose a task, establish a baseline and test the whole process. Count operating and review costs, record failures and distinguish attributed conversions from incremental contribution. If the question is financial impact, use an appropriate experiment or carefully qualified model rather than treating a platform metric as the final answer.
The possibilities that interested me in 2022 remain useful questions. Their value comes from what a particular implementation demonstrably does. See social media measurement and marketing automation and oversight for related considerations.
Leave a Reply