Pinning is a Math Equation, Not an Ideology

How long do you think it takes to complete an ad test when you “feed the machine” with the maximum number of available assets? What if I told you it could take over 1,000 years to get an answer you can rely on?

Before we get to that, let’s back up a little, so I can tell you the story of an account I was looking at last quarter. It was a B2B, mid-market e-Commerce account with spend in the 250k-500k range. As part of my audit, I wanted to get a feel for what they were testing. Like many other companies, they were leveraging RSAs with 15 headlines and 4 descriptions pretty much everywhere, but they had what I felt was a good amount of monthly impressions on both branded and non-brand, and so I initially felt that there was enough volume to justify the approach. I was excited to dig in and see what was working and what to do next.

After reviewing a number of the headlines and descriptions being run, I was having some difficulty in understanding what was being tested. There were appeals to convenience, shipping, price, variety, authority, and more all within the same RSA on top of keyword anchors and various calls-to-action. Eventually, my analysis frustrated, I decided to at least determine what the single best ad combination was to use as a baseline, starting with some of the higher traffic ad groups.

However, when I began looking at the ad combinations reports in several high traffic ad groups, and noticed that the “top combination” ads only had single-digit impression percentage delivery, I was troubled. I tried to parse it out by looking at the data month-by-month, cross-referencing the individual assets in the assets performance report. But what I found is that the same headline shown by one report to be a top-delievered asset and integral part of the winning ad combination, when viewed a month later, reported drastically differing results. I felt like I had no solid ground to stand on when it came to capturing the control ad for the next ad test recommendation.

Frustrated by this lack of clarity. I thought back to the “old school way” of A/B ad testing, back from the ETA era and even before from direct marketing. I decided to run some numbers, and I realized that with the current “assetmaxxing” approach, it would take 3,760 years to produce a single, statistically valid result for clickthrough rate improvement alone, and much longer for a conversion-based KPI.

This was something that the team previously managing the account either did not realize or did not believe in. Another issue is that even if they did resolve that test to confidence, if calls-to-action, benefit statements, offers, and keyword relevance anchors are allowed to appear (or not) in any headline position in the wild, that is, the search results page, what could be learned and used from the outcome of such a “test” anyway?

We inherited a never-ending test that we couldn’t use. There had to be a better way.

To Pin or Not to Pin, that is the Question!

There’s this whole debate in PPC about whether you should pin ad assets or leave them unpinned. The argument for leaving them unpinned is that the algorithm will figure it out and deliver the best ad for each user at auction time. Some proponents of this approach point to the fact that Google uses a multi-armed bandit algorithm and proprietary signals that advertisers don’t have access to make these decisions, and therefore: a) they can be trusted to get things directionally correct at scale and b) they can’t be replicated by an advertiser acting without all of that information. In this situation, impression delivery follows something like predicted CTR, and allows ads to be presented to the user without the need for reaching statistical significance.

Therefore, what I call “assetmaxxing” is what you’ll see Google recommend in official documentation, or what a Google rep will tell you on a call. You will find droves of “unpinned” advocates on LinkedIn and X posts highlighting the correctness of the strategy. Because of that, going unpinned has become the default practice over the last 10 years. “Don’t worry. Let the machine figure it out!” It’s easy to think of the algorithm as a magical black box that will always produce the optimal result for us, allowing us to worry about other things, especially if we believe we can’t compete with it, anyway.

Some people may even think of leaving ad sets unpinned as an ideological choice. People can start dividing themselves into camps as Pinners and Non-Pinners. But thinking about it as an ideology also carries this element of faith with it. The problem is that pinning has nothing to do with faith. At the end of the day, you have to contend with a math problem. One that may ultimately be unwinnable without pinning.

Pros / Cons of Unpinning

Google’s Pinning Guidance Creates Collapsed Structures That Can Dilute Performance

Before we get into the nitty gritty, let’s talk more about Google’s official guidance on pinning. Their advice in their official documentation, and echoed by many reps, is that you should feed RSAs with as much asset variety and quantity as possible, then let the algorithm (increasingly, they are referring to this as “Gemini” or “Google’s AI”) sort it out. Outside of having too many concurrent variables changing during testing, this approach also leads to accounts with collapsed structure (fewer campaigns, ad groups & keywords) and degraded keyword-to-copy-to-landing-page congruence. Let me explain what I mean.

While on the surface it sounds plausible that providing more headlines and descriptions in each ad gives you more testing, and therefore more chances to win, in practice (and by the numbers I’ll share shortly), the account can start devolving into mismatches between keyword impression and asset impressions, leading to decreased efficiency in the way of decreased clickthrough rates and conversion rates. It fundamentally violates the principle of relevance that search advertising was originally founded upon.

Assetmaxxing + Collapsed Structures = Diminishing Returns

This account I was looking at had a pretty collapsed structure compared to how it was built in the past. Based on the change history logs, the account changed hands over several agencies in a span of just over a decade, and while the oldest managers seemed to favor deep segmentation with high traffic single keyword ad groups (SKAGs) and single theme ad groups (STAGs), the newer campaigns were consolidated into just a handful of campaigns with heavy use of broad match and ad groups containing up to dozens of keywords.

One interesting finding was that the current iteration had what might be considered their most important non-brand keyword isolated in its own campaign and ad group, with one responsive search ad. And here, as with elsewhere in the account, they used the strategy I earlier called assetmaxxing, which is the practice of incorporating as many unpinned headlines and unpinned descriptions as possible. A lot of people are into assetmaxxing, and I was once, too. (This is also when I thought ad strength mattered, but that’s another conversation for another day.)

It turned out that the same basic RSA (the same set of headlines and descriptions) also appeared in all of the other search campaigns and ad groups. As I began to investigate both the search terms reports and the ad combinations reports for various ad groups, the issue started to surface that unoptimized, and sometimes downright nonsensical ad combinations were being shown for a given targeting topic. Each time that happened, a little efficiency was lost, because the tight connection between what was being searched by the user, what was being said in the ad and which landing page the user was directed to was incongruent.

The symptom that I noticed that became the diagnostic and part of the insight that led me to this new understanding of pinning in ad testing is that among the targeted keywords within the same high traffic ad group, there was significant variance in the “ad relevance” scores. There were pockets of keywords with “above average”, some “average” and many “below average”. Many of the highest impression volume keywords in ad groups analyzed were “below average”.

Going a layer deeper to the search terms report, in light of the broad match types being used, it was clear that the variety of queries and their semantic distance from the keywords actually targeted further exacerbated the issue. How could you design or trust an ad test when the queries the ads were being shown for were so heterogenous? How could you ever know why a winning ad wins, and design the next experiment?

This “pseudo-testing” problem gripped me straightaway, and I realized that it’s made worse by the fact that Google Ads never reports to advertisers which specific ad combinations or assets are shown for which specific queries! If you want to get close to even approximating this, you need to take matters into your own hands.

At the end of the day, if you don’t know what is really winning or why, how can you possibly scale what’s working or improve your performance in a rational manner? You can’t, because the alternative is that as you scale, the inefficiency scales with you, and the only one who profits from that is Google.