James & James Is Using MMM To Keep Its AI-Based Ad Buying 'Honest' | AdExchanger

Tuesday, May 27th, 2025 – 10:04 am

When ad tech veteran Tristan Cameron, formerly of LiveRamp and Inuvo, joined furniture brand James & James as CMO in January, she came with a mandate to question everything and bring greater rigor to the marketing budget.

Although it appeared that the brand was hitting its aggressive revenue targets while primarily spending on search and social ads, the facts told a different story, Cameron said while presenting at AdExchanger’s Programmatic IO event in Las Vegas last week.

“At the platform level, everything looked pretty good; things were performing,” she said. “But, unfortunately, that just wasn’t the reality for our bottom line.”

Shortly after starting at James & James, Cameron began working with her former colleagues at Inuvo to create a quarterly marketing mix modeling (MMM) benchmark.

This benchmark “keeps us honest,” she said, even when the big ad platforms over-credit their own work.

The platform playbook

The first big change for James & James was to pull the plug entirely on its Advantage+ Shopping Campaigns (ASC) budget, which is the main AI-based ad product on Meta.

When Cameron joined at the beginning of this year, James & James spent more than half of its monthly ad budget on Meta, most of which went to ASC, she said.

But under the brand’s new attribution lens, ASC wasn’t pulling its weight.

Although ASC was effective at driving a steady stream of general “furniture intenders,” Cameron said, it wasn’t attracting people specifically in market to buy a James & James table or couch. More likely than not, James & James was paying for traffic that would end up converting elsewhere, perhaps with a more budget-friendly brand, such as Ikea.

Still, cutting ASC was a big decision.

According to Eric Tilbury, Inuvo’s senior director of ad operations and solutions engineering, Meta “had awesome in-platform ROAS, and most of the time you wouldn’t touch that.”

“But the MMM hated it,” he said.

Why, though?

For one, as a high-end furniture seller, James & James probably wasn’t generating the overall number of transactions that an AI-based product like ASC needs to effectively optimize conversions for a brand.

AI products must be carefully calibrated. It’s easy for a brand to direct a machine to optimize for real human web traffic or even transactions without realizing that the KPI itself is causing great inefficiencies in how the budget is spent.

For example, James & James set the amount within its Meta ad account that it was willing to pay for real people considering buying a couch to visit its site. Meta dutifully delivered furniture shoppers – just the wrong kind. And the issue trickled down to retargeting, too.

Since James & James wasn’t effectively identifying high-end or handmade wood furniture shoppers, it ended up spending heavily to retarget people who were later converting more cheaply elsewhere. Meanwhile, its lookalike seed audiences also become inefficient.

James & James reset its Meta targeting to focus on people looking for interior designers, say, or who showed interest in handmade walnut furniture. Those types of terms were better proxies for the brand’s higher purchase consideration, rather than people who were less intentionally browsing for tables or couches.

Not that James & James no longer spends heavily on Meta – it does, Cameron said. But these days, its budget on the platform flows through Meta’s manually controlled ad products, not ASC.

The PMax problem

Meta wasn’t the only ad platform claiming massive ROAS for James & James without justification, though. There was also Google’s AI buying product, Performance Max (PMax).

PMax’s problem was also one of marketing incentives, Tilbury said.

Since James & James didn’t generate enough consistent purchases to get a strong signal for PMax, a Google rep suggested the brand use an add-to-cart metric.

“What do you think happened?” Tilbury said. “There were a lot of carts adds.”

But PMax had essentially “started hallucinating,” he said, using the term for when an AI-based service goes haywire.

Some Google Play Store apps, for instance, were either generating bot-based add-to-carts or were driving audiences who weren’t in market to add an item to their carts. Since PMax was optimizing based on the add-to-cart metric, it began homing in on incorrect audiences, and traffic sources generated solid platform-reported ROAS numbers – but not sales.

James & James was also spending heavily on branded search terms, since the Google platform credited that spend with dramatically high ROAS. The quarterly MMM, however, was attributing little in the way of incremental new business to branded search, Tilbury said. So James & James shifted to non-branded search and different types of keywords and targeting.

With better steering, James & James was able to keep its PMax spend going. And the product has begun to get a better sense of what the brand is looking for in a potential customer.

But the AI still needs careful oversight, Tilbury said, which is why brands should use an MMM benchmark, even if a quarterly attribution model means greater delay in the feedback loop.

An AI ad product can easily mistake high ROAS for quality, even when it’s the result of an outright mistake in the campaign calibration or because the platform is claiming credit for people who were going to make a purchase regardless.

“A lot of times, that never gets fixed, because you see good ROAS or good performance out of PMax and so you don’t touch it,” Tilbury said. “And you won’t catch it for a year if you’re not running an MMM.”