Interviews, insight & analysis on digital media & marketing

Training ecommerce algorithms to be fit for purpose

by Andreas Arentoft, Partner & Head of Go-to-market (Technology), Precis Digital

‘Rubbish in, rubbish out’ – or coarser variations on that theme – might be an old adage in digital advertising, but with algorithms beginning to rule the media buying roost it couldn’t be more pertinent.

The potential for algorithms to help digital advertisers go beyond the norm and maximise value from today’s standard ‘e-commerce first’ business model is enormous. But unless we train the algorithm to leap some almost universal hurdles, we might never realise the full benefits it could bring to the media buying landscape.

AI is now so prevalent in deciding all aspects from bids to audiences, and even creative treatment, that it’s easy to overlook the fact so-called ‘smart’ media buying remains in its infancy. It’s therefore simple, not especially sophisticated – and not always as efficient and effective as advertisers might hope.

Firstly, what are the three most common digital advertising problems for e-commerce marketers’ – and how can we get media bidding algorithms ‘e-commerce ready’ to outpace them?

Many unhappy returns

Having the option of sending back an unwanted item has become par for the course for today’s choosy customer. They expect returns policies to be created in their favour.

The only way to handle this and not hit your margins is to gather enough data to meaningfully predict return rate trends. At present, it’s too tempting for brands to figure out the mean return rate per item and simply adjust conversion values accordingly. But constant changes to the product set makes this level of analysis unpredictable at best and, at worst, terrible for forecasting.

Instead, try defining the average return rate per market, product category and month, especially during the Christmas period. This will be a much more applicable adjustment to your conversion value.

Loyalty isn’t everything

It can be tempting to concentrate campaigns on regular customers: after all, it’s said customer retention costs up to seven times less in marketing spend than acquiring new ones. But the ‘total new customers’ metric is becoming key to growth.

When a customer purchases a product, their previous experience with you – not the advertising that draws them in – is the biggest factor in deciding whether they buy from your business or from a rival. This means the incremental effect of every pound of digital ad spend is always less for existing customers than new customers.

An emerging, sophisticated method of customer acquisition is lowering the conversion signal sent to the algorithm. In short, this means making the most of the unique visibility and control you – not the ad platform – have of your customers, tailoring the ad signals accordingly. The algorithm won’t do this for you.

Stock up on best solutions

When an algorithm discovers a best-selling item, it will mercilessly push it – even if you’re running out of stock. Equally, overstocking can eat into the bottom line because it usually means cutting prices to clear inventory. 

Early detection is the key to monitoring and acting on over- or understock. The best measure is ‘Days left of stock’. If this is lower than your restocking lead time or days until next season’s drop, you risk understock – and vice versa. 

It’s straightforward to curate the products your e-commerce site buyers are exposed to, as digital ad solutions are already available. Google’s Performance Max and Meta’s Dynamic Products Ads, for instance, simplify product segmentation.

However, there are drawbacks to only relying on these platforms. The major one is the general lack of transparency and control they afford advertisers. This can make targeting customers and understanding campaign performance more difficult than they need to be. There is limited access for campaign managers to audience data and keywords, which for the most part rule out smart bidding strategies.

The answer to these quandaries, providing the ability to detect over- or understock early, is a more bespoke approach. My go-to measure is ‘Days left of stock’:

No.of items in stock / Avg. sales per day = Days left of stock

If ‘Days left of stock’ is lower than your restocking lead time or days until next season’s drop, you are risking understock and vice versa. 


Test moving high-stock items into discrete campaigns; ramp up aggressive bidding; and remove ad exposure for low-stock items.

The benefits of ‘bulking up’ algorithms

The more we train algorithms, the stronger and more agile they get. That way, we can see beyond standard ROAS measurement and reap the rewards of first-party data strategy.

Data is the ‘protein shake’ of more effective algorithms. In a world powered by AI, quality data and the insights driven from it can be the difference between beating or lagging behind the competition. E-commerce marketers must learn how to leverage first-party data effectively or risk being outmaneuvered.

A good way to look at this is: 

Just do it – Getting started, rather than becoming bogged down in the detail of a 100% accurate strategy, is essential.

And do it fast – It’s better to quickly send a signal, rather than waiting to collect more information and update later. Google’s recency bias will reward you for this approach.

High-impact training – If all product categories have similar return rates, reducing conversion value by a fixed percentage won’t change the algorithm’s choice. Include factors with a high profit impact.

For sure, the e-commerce algorithm left unchecked could leave your digital advertising flabby. Far better to flex its muscles and deliver the customer delight and competitive edge you crave.