There’s a growing tension in Google Ads.
Should decisions around targeting, bidding and optimisation be controlled by the advertiser, or governed by the algorithm?
Much of modern paid search management is about navigating the balance between those two poles.
It shows up in various places…
• Broad match for reach ‘with benefits’ vs Exact match for (diminishing) precision
• Maximise Conversions for machine learning – with its supposedly nimble decisions based on multiple variables for each auction, vs Manual bidding, for human understanding that bypasses a whole lot of trial and error but has blunter instruments to use…
As I discuss in my upcoming update for a course unit on ads,
RSAs are like a microcosm of this wider tension…
You can still exercise control over them, and determine almost exactly how your ad will appear – but it’s not the default option; Google will make you jump through hoops to do so…
It will effectively exert penalties for keeping that control, and will encourage you strongly to go in the other direction – i.e. aggregate your contents and trust the algorithms to make the best use of them…
It will overpromise the benefits of that automation, and will give you frustratingly limited data to judge the effectiveness of those algorithmic decisions for yourself.
(Isn’t Google Ads fun! (well, it is actually – so we won’t be too negative…))
So RSAs present the control vs automation dilemma in a particularly acute form.
At one end of the scale we have Pseudo ETAs, with one asset pinned to each position, and an Ad-Strength party at the other end, with an unpinned array of 15 headlines and four descriptions.
I’ve discussed before how the latter approach runs quickly into the problem of needing too many impressions to uncover useful performance patterns. (There is some unknown-but-too-limited degree of pattern recognition at work alongside that brute trial and error. Park that for now.)
This is worth pausing to consider. The more combinations or options any optimisation mechanism has to test out, the more impressions any decent test will require… and the number required quickly becomes too high for any notion of precise, multivariate testing to be realistic.
That’s certainly the case with RSAs. And the upshot is that this process of optimisation by trial and error needs a lot of help from us.
One of the key ways in which we can provide that help, is by providing the RSA with only good combinations to try
For example, take this fairly typical array of headlines:
Name of brand, name of service being offered, a few CTAs, features, benefits. Looks reasonable…
But of all the possible two-and-three-line combinations of these headlines, there are actually relatively few that would pass your quality control as a hand-written combination.
The algorithm can and will perform some sense checking when putting them together. We know that it tends to avoid clear repetition… but it doesn’t do as great a job of sense checking (or sense making) as we might hope or imagine…
In the real ad on which I based this set of headlines, the second most common headline variation in the early stages was:
…telling the user nothing at all about the product, its benefits, or how to proceed with it.
If the algorithm thinks that’s a reasonable combination to try, then how many impressions are going to be wasted on this and similar non-starters before the data alone persuades it otherwise?
(This isn’t just a theoretical point. As I discuss in the course unit, a slew of studies such as this one by Optmyzr and this by Adalysis cast doubt on the effectiveness of RSAs to produce ads that match ETAs in performance.)
The lesson is: don’t overwhelm the algorithm with an undifferentiated array of options and expect it to weed out the poor combinations efficiently. Machine Learning can do big things with big data sets, but it will be a while before it has our level of ability to cut to the chase (not to mention that big data sets in PPC often rely on big spend).
Instead we need to feed it overwhelmingly with good combinations to start with.
This may mean:
• Not using all 15 headlines
• Pinning a set of CTAs to H2 – CTAs being still the most effective message type to include in an ad (see this excellent study by Adalysis on that topic) and H2 being the best place for them.
• Trying multi-pinning, with e.g. several variations of the company or product name in H1, and a CTA or CTA+benefit in H2 (H3 is still rarely used whether assets are pinned there or not.)
This idea – of placing guardrails around the algorithm’s path – applies to optimisation more widely too.
‘Trying to do too much’ is one of the most common pitfalls in Google Ads.
A £15 per day shopping or search campaign for a company selling 100s of products drowns in the process of trying to determine what works best – whether it’s the smart bidding algorithm or the manual-bidding PPC manager undertaking that process.
And as with RSAs, the solution is again: limit the scope to ‘good options’ only. Advertise fewer products, restricted to those you most expect to prove winners.
This is just one form of that central optimisation principle: Cultivate more of what works and less of what doesn’t…
But some of these decisions to ‘cut out the waste’ need to be made in advance, and are particularly acute when waste traffic produces not only empty spend, but also a huge amount of ‘noise’ for algorithms that are designed to identify winning patterns.