The second format.

For thirteen weeks from late February we ran five distinct audience definitions, each isolated in its own campaign, plus a separate retargeting campaign. Isolation is the whole point and it is not how most programmatic media is bought — the common approach pools audiences into one campaign and lets the platform optimise between them, which is cheaper to operate and makes it permanently impossible to say which audience did what. Separating them costs efficiency in the short run and buys attribution you can act on.

Everything is recorded at audience × week × channel granularity. That is finer than platform reporting offers by default, and it is the reason the rest of this document can exist.

1. 01
   Cut the weakest audienceOne of the five had the lowest click-through rate and the highest cost per click of the set across the full thirteen weeks. It was retired at the changeover rather than carried forward on the assumption it would come good. Four audiences continued.
2. 02
   Consolidated the survivors into shared campaignsThe audiences moved from five separate campaigns into shared ones, running as audience segments inside them. At this budget level, splitting a small daily spend across many campaigns starves each one of the volume it needs to deliver efficiently. The test had already given us the per-audience read we ran the structure to get.
3. 03
   Added a second ad formatNative inventory was introduced alongside display, carrying the same four audience definitions at the same negotiated CPM. This was our recommendation and it is the change with the largest measured effect.

What follows is what each change produced. Every figure is like-for-like: same advertiser, same platform, same audience definitions, same media price.

The media is bought at a fixed CPM. Across all fifteen channel-by-flight cells in this dataset the effective cost per thousand impressions sits between $7.63 and $7.97. Price is a constant.

That removes the usual escape hatch. We cannot attribute any of these results to buying impressions more cheaply, because there was no cheaper price to get. Everything below is a delivery result — a consequence of which people saw the ads, how often, and in what format. It also means one thing we are explicitly not claiming: holding CPM flat is the contract, not an achievement.

Native ran across the same ten weeks, against the same four audience definitions, at the same fixed CPM as display. The only variable is the format. That makes it the closest thing in this account to a controlled experiment, and the result is not marginal.

Read the last two rows together. Native produced 76% as many clicks as display, on 19% of the money. Across the flight, 13.5% of spend generated 34.7% of clicks. At a two-proportion z of 21.9 this is not a small-sample accident, and because it replicates across four independent audiences it is not one lucky segment either.

The practical consequence is a standing reallocation case: every dollar moved from display to native in this account has bought roughly four times the click volume at the same media price.

The second change surfaced something the test structure had been quietly hiding.

Retargeting pools shrink. The audience is people who already visited the site, and if the budget holds constant while that pool contracts, the platform does the only thing available to it: serve the same people more often. Over thirteen weeks this campaign climbed from roughly 23 impressions per person per week to 179.

Since the changes, the same campaign has run between 1.5 and 2.1 impressions per person per week, every week, for ten weeks — and its click-through rate is statistically unchanged (0.2036% → 0.2088%, z = 0.28).

That combination is the finding. The campaign was never underperforming on any metric a standard report shows. It was performing identically while spending most of its budget on people who had already seen the ad thirty times. No campaign-level dashboard surfaces this, because no campaign-level dashboard divides impressions by unique people. Ours does, weekly, per audience.

The same pattern held, less dramatically, on prospecting display. Comparing display only, so that the added format cannot flatter the comparison, the restructured flight bought 29% less media per week and put it in front of 71% more people.

One number moved against us, and it belongs here rather than in a footnote. Display click-through rate fell from 0.1168% to 0.0927% — a real decline at z = 4.26 across 1.4 million impressions, holding for each of the four audiences individually. The likely mechanism is the reach expansion itself: repeated exposure to a small, already-warm pool clicks harder than first contact with a cold, broad one, and we roughly halved frequency. We cannot fully separate that from ordinary creative fatigue or inventory drift, so we offer it as the leading explanation rather than a proven one. Our judgement, stated as a judgement, is that for a service with a three-to-six-month consideration cycle we would rather reach 23,000 people twice a week than 13,700 people five and a half times — and note that the format change in Result 1 more than recovered the lost click volume at a quarter of the cost.