In this week's column, Shirley Marschall looks at why attribution in advertising is a lot like working out which mosquito bit you on holiday...

August is over and with it most vacations. But some of those mosquito bites from your vacation are still there, annoyingly flaring up even weeks later. But which mosquito was it? The one from your early morning walk in week two? Or was it the mosquito at dinner at sunset? First bite or last bite? And which exposure even gets 'credit' for this itch you’re scratching at your desk right now? The original bite, the heat that reactivated it, or the moment you noticed it?

No wonder attribution is one of the most debated areas in ad tech. Last-click, multi-touch attribution (MTA), Media Mix Modelling (MMM), platform or independently measured, the options are almost endless, and the challenge of assigning credit to ads and touchpoints that influence a consumer’s decision remains blurry at best.

Once upon a time (and irritatingly still widely relied on even nowadays) last-click attribution dominated digital advertising, giving full credit to the final interaction before a conversion. A walled garden darling, ignoring the role of any earlier touchpoints, however, turned out to be rather problematic. It undervalued channels driving consideration and brand lift, while over-rewarding lower-funnel activity.

To fix this, advertisers turned to multi-touch attribution (MTA) models that distribute credit across multiple interactions. Much better, except MTA’s user-level tracking didn’t fully survive contact with privacy regulation. And it didn’t help that platforms have spent years reporting their own homework instead of submitting it. The Association of National Advertisers has been pushing since 2017 for independent audits of walled gardens, rather than taking self-reported numbers at face value, since every platform swearing it was their mosquito isn’t quite the same as measurement.

That’s part of why marketing mix modelling, the "older", more holistic approach from before anyone could track individual users at all, came back in favour. MMM doesn’t just ask which bite did it. It looks at total exposure, time spent outdoors, footfall near standing water, and works backward to how much itching that exposure probably caused in aggregate. Less precise about any single mosquito, but closer to what can actually be known.

But that was before AI… 

AI assistants became part of the customer journey just as the industry settled into an uneasy truce over how to measure the old one. Now AI is increasingly deciding which brands even make the shortlist, what gets compared, and what gets ignored entirely. And that creates a new attribution problem: the influence happens inside a conversation that may never produce a click.

Someone spends twenty minutes asking ChatGPT to compare three skincare brands, gets a recommendation, closes the tab, thinks about it for three days, then types a brand name straight into Google. Every dashboard downstream logs that as direct traffic while the exposure that actually did the work never shows up anywhere. That’s not a net-new problem but with ads in chats and ChatGPT already reportedly reaching a USD$1bn (£739.2m) annualised advertising revenue run rate after just six months, that missing piece is becoming harder to ignore or more urgent to solve, definitely much more discussed.

Measurement vendors are trying to build something like an itch diary for this: using marketing mix modelling to connect what brands spend to where they surface inside AI answers, and what that surfacing does commercially. Except that diary isn't staying a diary. A recent CIMM report on MMM's expanding role notes that MMMs are increasingly wired directly into automated bidding and budget systems, no longer just informing a human's judgement but executing it. A model’s output about how much credit to assign to a channel doesn’t stay just that. It becomes an instruction that moves real money, with nobody left to pause and wonder, "Was it actually that mosquito?"

Worse, the allocation can end up confirming itself. If a model over-allocates budget to a channel based on incomplete data, the resulting performance data reflects that and reinforces the original allocation further. An endless, potential wrong-loop that keeps generating evidence it was right.

The IAB noticed another issue and released new guidelines that don’t (even) try to fix attribution. They try to get the industry agreeing on what it’s even looking at: shared vocabulary, quality criteria, disclosure requirements, so two vendors both claiming to measure "AI visibility" are describing the same thing. For now, the guidelines punt on attribution itself, offering only an attribution-framework-is-forthcoming promise, and focus instead on getting the industry to agree on measurement quality: directional signals versus results solid enough to actually act on.

Guidelines, attribution methods, AI and all the rest aside, another study that found advertisers only "slightly confident" in measurement generally is more specific about attribution: it's one of the metric areas advertisers expect to matter most over the next three to five years, and one where confidence is currently weakest. The industry is betting harder on the exact thing it trusts least…