TV, radio, podcast and out-of-home attribution means measuring channels nobody clicks. The click-trail playbook is out from the start: no sessions, no UTM tags, no trail. What’s left is inference: timing signals, exposure matching, promo codes and aggregate modeling, each honest about a different slice of the effect.
And then there’s connected TV (CTV), which promised digital-style measurement for television and only partly delivered. This guide works through the methods channel by channel: what each honestly measures, what each misses, and how they stack into a working setup.
Why the unclickable channels resist attribution
The structural facts: no click means no trail; exposure happens on one device (the living-room screen) and response on another (the phone); and effects spread over days to weeks, so the response rarely lands next to the ad in time.
Every offline attribution method is therefore an inference: a reasoned estimate of the connection between exposure and outcome, never a followed trail. That’s not a flaw to hide: it’s the honest starting point, and vendors who claim otherwise are the ones to worry about. The useful question for any method on this page: is it an honest description of an inference, or a claim dressed as certainty?
For the trail-based world these channels sit outside of, see what marketing attribution is.
Spike analysis: the timing signal
The classic method for TV, radio and podcasts: match ad airings to short-term response spikes: site visits, brand searches, app opens in the minutes after a spot runs.
The homely version: a coffee shop buys a radio spot, then sees twelve site visits in the hour after airing against a usual eight. Those four extra visits are the spike, and across many airings the pattern becomes solid evidence for comparing creatives, slots and stations.
What spike analysis honestly measures: immediate direct response. What it structurally misses: the delayed brand effect, the listener who wasn’t ready to buy but will remember the name in three weeks. That effect is usually the larger share, and it never spikes. Celebrate the spike; just don’t mistake it for the whole return.
CTV attribution: digital-ish, not digital
Connected TV serves ads through internet devices, which genuinely improves on linear TV’s panel estimates: impressions are logged, and household-level matching (usually by IP address) can connect exposures to later site visits and conversions.
Now the honest limits. A household is not a person: four people watch, one buys, the match can’t say which saw the ad. IP matching is coarse and increasingly squeezed by privacy rules. CTV inventory lives in walled platforms that share aggregate reporting on their own terms. And view-through windows: how long after an exposure a conversion still counts, are vendor-defined choices that move the numbers materially.
Fair verdict: CTV sits between digital precision and TV inference, closer to the latter than the sales decks suggest. Use its reporting for in-flight tuning; don’t let it grade its own homework on budget questions.
Codes, URLs and panels: the direct-response bridges
The other bridges, each honestly rated.
Promo codes and vanity URLs are clean: the code is used, the ad gets credit. But they only capture responders who use the mechanism: everyone influenced who bought without the code is invisible, so codes give you a floor estimate, never the total effect.
Podcast attribution runs on the same logic: mention-this-show codes plus timing analysis, and the same undercount: the host’s recommendation works on plenty of listeners who never type the code.
Out-of-home is measured through footfall estimation (aggregated location signals suggesting how many people passed and visited) and geo-lift comparisons: areas with the campaign versus matched areas without. Coarse, directional, and still genuinely useful for the biggest-spend decisions.
The shared property: every bridge undercounts. Treat their numbers as minimums, and let the aggregate layer find the rest.
The aggregate answer: MMM was built for this
Here’s the satisfying historical fact: marketing mix modeling was invented in the TV era, precisely because the biggest advertising channels of the day resisted trails. Measuring the unclickable isn’t an edge case for MMM: it’s the founding use case.
The mechanics fit perfectly. MMM reads total sales against spend and airing schedules across every channel, clickable or not, and estimates each channel’s contribution: with adstock handling the delayed effects that spike analysis misses, and the baseline separating advertising’s work from sales that were coming anyway. No user-level data anywhere. Geo-experiments (one region gets the campaign, a matched region doesn’t) then validate the biggest claims causally.
The output is the thing no bridge can produce: the whole picture: every channel’s contribution to total sales, delayed effects included, on one comparable scale. The fundamentals are in what a marketing mix model is, and MarSci’s version runs always-on: MarSci’s marketing mix modeling solutions.
A working setup for offline-heavy mixes
For a mix heavy in TV, CTV, radio and out-of-home, the blueprint:
Spike analysis for fast creative feedback: which spots, slots and stations pull immediate response: answers in days, not quarters.
CTV platform reporting for in-flight tuning, held with the skepticism self-reported numbers deserve, and validated against independent signals where possible.
MMM over total outcomes for budget truth: the complete, delayed, cross-channel effect, which none of the faster methods can see.
A geo-experiment before any big bet: the causal proof that the model’s biggest claim survives contact with reality.
The unclickable channels are measurable: just not by pretending they’re clickable. If your mix is heavy in them, talk to MarSci.
Frequently Asked Questions
Through inference, not trails: spike analysis matching airings to immediate response, promo codes as floor estimates, and marketing mix modeling reading total sales against airing schedules for the full delayed effect. No single method sees everything; together they do. Each method covers a slice; the model completes the picture.
Connected TV logs ad impressions and matches households (usually by IP) to later visits and conversions: genuinely better than linear TV’s panel estimates, but household-level, coarse, privacy-squeezed, and often self-reported by the platform selling the inventory. Use its reporting for tuning, not for budget verdicts.
Partly: promo codes, vanity URLs and timing analysis capture direct responders, and always undercount listeners influenced who never use the code. Aggregate modeling recovers the fuller effect that the codes can’t see. Treat code-based numbers as the floor of the true impact, never the ceiling.
Combine methods: fast bridges (codes, spike analysis) for creative feedback, geo-experiments for causal proof on the biggest decisions, and MMM for the complete delayed effect inside total sales. No single method carries offline measurement alone. Combining methods is the method.
