One of the most frustrating, puzzling situations digital marketers encounter is paid media campaign results that get progressively worse despite their best efforts to optimise for continuous improvement.
Even as their cost per acquisition (CPA) rises, they see the quality of their leads deteriorate. As strange as it might seem, it is often their efforts to optimise that are causing campaign performance to deteriorate.
The downward spiral starts with bad data going into the programmatic platforms a marketer is using for their campaigns. Platforms such as Meta, Google Ads, and Performance Max use machine learning to make automated decisions about a campaign.
In theory, these platforms should optimise for lead quality and customer acquisition cost based on signals such as conversions, audience behaviour, website actions and CRM feeds.
But in many cases, the initial data they are using to make decisions is inaccurate or of poor quality. This leads the system to learn from the wrong examples and then constantly reinforce the wrong learnings because it started from a faulty assumption.
How the negative loop forms
It is a bit like learning the wrong form for an exercise from a YouTube video. The more you repeat the movement, the more your body reinforces the wrong technique and the harder it becomes to correct later.
This is how the negative loop forms: A marketer starts out with poor quality data or incorrect configurations, which leads the programmatic platform to optimise for the wrong outcomes, such as reaching the wrong audience or prioritising low-quality engagements.
Even as the platform receives spam leads or low-quality traffic, it optimises for them as if they were valuable outcomes. Low-quality inputs drive bad optimisation, which generates worse inputs.
This problem cannot be addressed with superficial adjustments like tweaking bids, shifting budgets, or making minor targeting changes.
Even when you later apply recognised best practices, such as targeting high-value conversions or past customers, the algorithm’s foundational learning remains biased to the wrong measures of success. Over time, this may show up as rising CPAs and a steady decline in lead quality.
It starts with bad data
As I mentioned earlier, the vicious cycle usually starts with bad data. Here are some examples:
- Your audience models are polluted because the algorithm has clustered the wrong users together as converters
- Flawed conversion data has led the algorithm to down-bid away from real customers because they appear less likely to convert than spammers
- Ads start over-serving on placements or audiences that attract junk interactions
Once the algorithm is trapped into a vicious cycle of poor data and corrupted learning, it can only be remedied with structural changes to the underlying setup.
Here are four steps towards breaking out of the negative loop:
1. Audit your data inputs
- Verify that conversion events are correct: For example, check there are no duplications and that the system is not registering conversion events too early or too late.
- Ensure real customer signals are feeding platforms: Make sure you are using trustworthy data such as offline conversions, CRM uploads and value-based signals.
- Remove fake or low-value events such as add-to-cart or time-on-site if they’re misleading optimisation
2. Audit your platform settings
- Check networks: Exclude Google Search Partners if it is generating spam and review display expansion and automatic placements.
- Check demographics: Look for over-delivery in low-quality age ranges or genders.
- Check location settings: Ensure “Presence: People in or regularly in the location” vs “Presence or interest” is prioritised in the location setup
- Check audiences: Make sure that lookalikes built from real customers rather than from weak top-funnel activity. Also check that broad campaigns are actually broad.
3. Audit website behaviour via GA4
Look at engagement rate, bounce rate, scroll depth and session quality. This will help you identify patterns like traffic hitting irrelevant landing pages, high bounce rates from paid sources and poor on-site conversion flow.
4. Structural reset
In many instances, applying the fixes in steps one to three will be enough to restore performance. But if the algorithm is too far gone, a complete reset may be necessary.
This could entail rebuilding Performance Max from scratch with correct conversion signals and audience lists or rebuilding Meta ad sets to remove conflicting optimisation signals. This sort of system reset allows the machine to relearn everything from a blank slate.
From vicious circle to virtuous cycle
Once you correct the feedback loop, you can create a virtuous cycle. The algorithm prioritises better quality user signals, in turn resulting in a rise in lead quality before any noticeable increase in volume. Cost per conversion will start to level out and then decline as audience targeting becomes more accurate and efficient.
What’s more, search terms and creative delivery will align more closely with genuine customer intent, shifting the system from reinforcing errors to compounding improvements. The result is a point where optimisations finally hold because the underlying learning path has been set in the right direction.
Tanika Corneleus is campaign lead for digital media at iqbusiness.
