Breaking into the ‘Black Box’: An e-commerce case study on unlocking Meta’s product-level ad data

The promoting dilemma: The ‘Black Field’ of dynamic adverts

Ecommerce and Meta usually go hand in hand. You can provide Meta a 20,000-item catalogue and a price range – and with its AI-powered Advantage+ campaigns – it’ll attempt to pair the proper individual with the proper product, whether or not that’s a brand new buyer or somebody who’s already considered these merchandise earlier than.

However what’s really occurring inside that advert? And is there a option to optimise this ‘black field’ Dynamic Product Advert (DPA) format?

Advertisers can see ad-level efficiency, however have no platform-native insights on which particular merchandise are being proven, clicked, or ignored inside a broad DPA.

Is the algorithm making the proper selections?

That’s precisely the query we needed to reply. 

There are three widespread traps manufacturers fall into:

1. Over-segmentation – manufacturers that need extra perception break aside their catalogue into area of interest product units with tons of DPAs
   • Professionals – you can provide every advert a bespoke identify, which tells you precisely what’s being served. Good!
   • Cons – this reduces information density and might kill ROI. There’s additionally a bent to attempt to predict which audiences will reply to which merchandise, which is now not efficient for many manufacturers since Meta’s improved Andromeda updates
2. Convoluted reporting – manufacturers attempt to infer what merchandise Meta is prioritising by pairing GA4 session information (classes by product) to Meta adverts information (the campaigns/adverts that despatched these customers)
   • Professionals – permits some evaluation with out falling into the ‘over-segmentation’ pitfall
   • Cons – time-consuming to arrange, and incomplete. This methodology doesn’t inform us something about product-specific engagement inside Meta – we might solely be guessing at click-through-rate, spend and impressions. 
3. “Set it and overlook“ – manufacturers quit all management and let Meta take the wheel
   • Professionals – avoids over-segmentation points
   • Cons – There’s an enormous threat in trusting the algorithm. You may be pushing merchandise that get excessive impressions however low gross sales, successfully burning your price range and dropping effectivity.

Attempting to deduce efficiency and make selections from Meta Adverts Supervisor UI information is a threat. Many entrepreneurs are nonetheless not assured in AI-powered campaigns – if this is applicable to you, you might want to learn this whitepaper earlier than carrying on.

In my position at Impression, we created our Dynamic Product Explorer (DPEx) to resolve this problem – however worry not, I can stroll you thru the precise steps so you are able to do the identical to your model.

Our pilot shopper for DPEx was a main toilet retailer investing closely in DPAs inside conversion campaigns. 

Let’s undergo the three phases in our journey to overcoming this ecommerce problem – so you possibly can recreate.

Part one: Surfacing engagement information

The primary stage in DPEx was visibility – understanding what was occurring now inside these ‘black field’ DPA codecs.

As I mentioned above, Meta doesn’t instantly report which particular product led to a particular buy inside a DPA within the Adverts Supervisor interface. It’s merely not an obtainable breakdown in the identical method that age, placement and so on. are provided.

However the excellent news is {that a} treasure trove of perception is buried within the Meta APIs:

1. Meta Advertising API (particularly the Insights API) is the principle API we use to get all advert efficiency information. It’s how we’re pulling the important thing metrics like spend, impressions, and clicks for every ad_id and product_id.
2. Meta Commerce Platform API (or Catalogue API), this API gives the checklist of all product_ids and their related particulars (like identify, value, class, and so on)

By becoming a member of these two APIs, we will extract important information, damaged down by product, with their particular clicks and impressions. 

We collaborated with our Media Options colleagues to pipe this information into our information warehouse (we use BigQuery, however there are various choices on the market).

This new, mixed dataset was then visualised in a Looker Studio report template. Once more, different choices can be found. To make sense of the information, we would have liked an simply navigable report, moderately than pages of uncooked information. We constructed the next visualisations:

Product scatter chart – separating every product into 4 distinct classes:

• ‘Star performers‘: Excessive impressions and excessive clicks.
• ‘Promising merchandise‘: Low impressions however a excessive click-through charge.
• ‘Window customers‘: Excessive impressions however very low clicks.
• ‘Low precedence’: low clicks and impressions

High/backside merchandise bar charts – see at a look the highest 10 and backside 10 merchandise by engagement.

Product particulars desk – view detailed metrics for every product.

This might all be filtered by product identify, product kind, availability, and another metrics we needed (color, value, and so on.)

This gave us our first-ever view of the DPEx report – product-level advert engagement!

Even with simply engagement information, this report supplied worth:

• Inventive – we used the information to enhance inventive briefs
  • For our toilet shopper, it was fascinating to see how a lot Meta was pushing non-white merchandise (orange sinks, inexperienced baths) – even though 95% of their product gross sales are conventional white variations. 
  • We hadn’t prioritised these merchandise initially, however have now created heaps extra video and creator content material that includes these extremely clickable variations
• Product segmentation – We are able to now construct highly effective, data-driven product units based mostly on actual engagement metrics
  • For instance, we examined exhibiting solely our most participating ‘Star Performer’ merchandise in feed-powered assortment adverts in our higher funnel campaigns. In these campaigns, the algorithm has fewer indicators to optimise in the direction of, so a broad product set improved because of this.
• Effectivity – this automated a fancy evaluation that was beforehand unwieldy and time-consuming

Crucially, for the primary time, we had sufficient proof to problem Meta’s ‘finest observe’ of utilizing the widest doable product set.

Pitfalls & key concerns

Nonetheless, this work solely went thus far. There have been some key issues lacking that simply tapping into Meta’s API received’t remedy:

• Engagement vs. Conversions: The main downfall with that is that product-level breakdowns are solely obtainable for clicks and impression information – not income or conversions. The “Window Customers” class, for instance, identifies merchandise that get low clicks, however we couldn’t (on this section) definitively say they don’t result in gross sales.
• Context is vital: This information is a strong new diagnostic device. It tells us what Meta is exhibiting and what customers are clicking, which is a large step ahead. The why (e.g., “is that this high-impression, low-click merchandise only a high-value product?”) nonetheless requires our workforce’s evaluation

We knew the above Meta-only information simply explores one a part of the journey. To evolve, we would have liked to affix with GA4 information to search out out what prospects are literally shopping for after they’re interacting with our feed-powered dynamic product adverts.

As an alternative of counting on platform-modelled conversions, we grabbed information from GA4 particularly for buy occasions. This gave us the exhausting numbers: the income and items bought for each transaction.

The essential key right here was getting this information by the merchandise ID, as that is what we used to attach our “revenue” information again to the “engagement” information from Meta.

Pitfalls & key concerns

Becoming a member of Meta and GA4 information sounds simple sufficient – however there have been some key blockers to beat. Our first two have been hygiene-focused: 

• Clear information. The entire mannequin breaks in case your Meta ID doesn’t cleanly match your GA4 IDs. You need to guarantee your product catalogues and your GA4 tagging are completely aligned earlier than you begin.
• UTM self-discipline. To correctly hyperlink a purchase order in GA4 again to a particular Meta advert, we would have liked to seize the ad_id by way of your UTM parameters. This ad_id is the magic bridge that lets us be part of the 2 datasets.

Nonetheless, the final one was more durable to beat – attribution points. The GA4 information will virtually all the time present decrease conversion numbers than Meta’s UI. 

It is because Meta usually ‘over-credits’ – it advantages from longer attribution home windows, together with view-through conversions, and it offers itself full credit score for every conversion it measures (moderately than spreading out throughout a number of channels). 

GA4, nevertheless, usually ‘under-credits’ channels like Meta. It makes use of data-driven attribution to attempt to give credit score to a number of touchpoints – however it’s unable to fully mirror consumer journeys – that means GA4 doesn’t simply recognise when a consumer has interacted with adverts earlier than buying, and credit score accordingly. 

Though we’d like to get a 1:1 match from every product buy again to the precise product interacted with on Meta, neither platform can simply present this perception. There’s nonetheless worth find the relative insights and developments.

Right here’s an instance:

• Meta’s UI: Reported our ‘Luxurious Freestanding Baths – Inexperienced’ product is our prime performer final month, with excessive volumes of clicks and impressions in our dynamic adverts.
• The Downside: Once we joined our GA4 information, we noticed no gross sales for that particular tub final month, in any respect, from any channel! That is the place, earlier than we would’ve panicked and excluded the product from our catalogue to ‘save spend’, and that is the pitfall.

However, by taking a look at all objects bought in these GA4 classes that originated from the ‘Luxurious Tub – Inexperienced’ product, we uncover that many customers who clicked the tub went on to transform however purchased the white variation as an alternative.

The Actionable Perception: The ‘Luxurious Tub’ advert wasn’t a failure; it was a extremely efficient halo product – drawing in aspirational prospects who then convert to purchase different merchandise.

We are able to then confidently fee creator content material, specializing in this tub to attract in new customers even when we all know customers are probably to purchase a distinct color relating to buy.

Part three: Efficiency-enhanced feeds

As soon as we have now this information at your fingertips, the temptation is to deal with it purely for insights and information.

The subsequent stage was even higher – utilizing this information to create automated supplementary feeds. Utilizing those self same 4 product classes from our scatter charts, we created automated logic to fabricate new product units and carried out the next assessments:

• ‘Window Customers’: (Excessive impressions, low clicks/gross sales). Feed these into an exclusion set to grasp if effectivity improves once we take away them from the feed.
• ‘Promising Merchandise’: (Excessive CTR, excessive CVR, low impressions). Feed these right into a scaling set with extra price range to grasp if demand is hidden.
• ‘Star Performers’: (Excessive impressions, excessive clicks). Feed these right into a retargeting set to recapture engaged customers with our signature ranges.

Pitfalls & key concerns

The assessments above are merely examples of hypotheses. Nonetheless, your mileage will range! We strongly suggest testing these use instances strategically with a view to perceive whether or not it provides to your total efficiency.

Is your model prepared to interrupt out of the ‘Black Field’?

Breaking out of Meta’s “Black Field” isn’t only a technical train; it’s a strategic must-do for e-commerce advertisers. The journey strikes from surfacing fundamental engagement information (Part one) to becoming a member of it with gross sales information for true, profit-driven insights (Part two), and in the end, to automating your technique with performance-enhanced feeds (Part three).

That is how you progress from trusting the algorithm to difficult it with proof. In case you’re a decision-maker questioning the place to begin, listed below are the three inquiries to ask:

1. “Are you able to present me which particular merchandise in our catalogue are being prioritised by Meta?”
2. “Are our Meta product_ids and GA4 item_ids an identical?”
3. “Are we capturing the advert.id in our UTM parameters on each single advert?”

If the solutions to those questions are ‘I don’t know,’ you’re in all probability nonetheless working contained in the black field. Breaking it open is feasible – it simply requires the proper information, the proper technical joins, and the need to lastly see what’s really driving efficiency. Get in touch with our team.