A famous experiment that could not have found anything

View through conversions card showing the number nine, the conversions recorded in a widely cited experiment
Contents 9 sections

View through conversions credit somebody who saw your advert, never clicked it, and converted afterwards. Whether that credit is deserved is the single hardest question in advertising measurement, and the most widely cited attempt to answer it does not survive a look at the arithmetic.

That experiment is worth walking through, because the mistake in it is one this whole industry makes constantly and rarely names.

This is for you if view through activity is inflating your reported numbers and you want to know how much of it is real.

What view through conversions are claiming

They assert influence without interaction.

Somebody is shown your advert. They do not click. Days later they arrive by another route and buy. The platform records a view through conversion, on the basis that the impression contributed.

Sometimes that is entirely true. Advertising does work on people who never click, and pretending otherwise would be its own error.

The difficulty is that the data looks identical either way. Somebody who was always going to buy, and happened to be shown an impression, produces exactly the same record as somebody genuinely persuaded by it.

So the count is not a measurement of influence. It is a count of coincidences, some of which were causal.

Why remarketing inflates view through conversions

The audiences most likely to be shown these adverts are the audiences most likely to buy anyway.

Remarketing targets people who already visited your site, browsed a product, or abandoned a basket. Those people have demonstrated intent before any impression was served.

Show adverts to a group already heading towards a purchase, and a large number of them will convert. Every one of those inside the window becomes a view through conversion.

That is the structural reason this metric flatters remarketing specifically. It is not a fault in the counting. It is a consequence of choosing an audience selected for their likelihood of doing the thing you are about to take credit for.

The experiment everybody cites

WordStream ran the best known attempt to test this, and published the numbers, which is more than most people do.

The setup: a display remarketing campaign ran for thirty days and produced nine view through conversions from 2,418,973 impressions. Then a clone of the campaign ran with a completely blank banner, producing zero conversions from 315,677 impressions.

The conclusion drawn was that view through conversions can be trusted, on the reasoning that a blank advert produced none while a real advert produced nine.

Read that way it sounds convincing. The blank banner is a clever idea and publishing the raw figures is genuinely good practice.

Then do the arithmetic on the second arm, which is where it falls apart.

Why the test could not have found anything

Nine conversions from 2,418,973 impressions is a rate of about 3.72 per million.

Apply that rate to the blank banner’s 315,677 impressions. The expected number of conversions is roughly 1.17.

So the test asked whether a blank banner produces conversions, in a sample predicted to yield about one. Observing zero when you expect 1.17 has a probability of roughly 31% under a Poisson model. That is purely by chance, even if the blank banner worked exactly as well as the real one.

The blank arm was too small to tell a real effect from nothing. A null result from a powerless test is not evidence of absence. It is certainly not evidence that the original nine were causal.

The experiment did not fail because the idea was bad. It failed because the second arm ran at a fraction of the volume needed, and nobody checked the arithmetic first.

The mistake this represents

A well-known publisher drew a confident conclusion from an outcome that was the most likely result regardless of the answer.

That pattern is everywhere in marketing measurement. A test runs, produces nothing, and the nothing gets read as a finding. It usually means the test was too small.

The check takes two minutes and almost nobody does it. Work out the base rate from the arm you have data for, apply it to the size of the other arm, and ask how many events you would expect if the thing you are testing has no effect at all. If the answer is one or two, the test cannot tell you anything.

I would rather cite this experiment as an example of that failure than as evidence about view through conversions, which is the opposite of how it usually gets used.

What to do with the number in your own account

Three practical positions, none of which require resolving the underlying question.

Report view through conversions separately, always. Blending them into a headline figure means part of your number rests on an assumption and nobody downstream can see which part.

Weight them lower when comparing campaigns. A display campaign whose results are largely view through is not directly comparable with a search campaign whose results are clicks, and ranking them on one blended figure will misallocate budget every time.

And know your exposure. On search this is a rounding error. On display, video and awareness activity it can be most of the reported total, which means the same setting has completely different consequences depending on what you run.

The only way to actually answer it

A controlled experiment, where a comparable group is deliberately not shown your adverts.

That is what incrementality testing does. It observes what happens without the advertising rather than assuming it. Nothing in a reporting interface can do this, because the interface only ever sees people who were shown adverts.

Those tests are expensive and they need scale. That puts them out of reach for many advertisers, which is a real constraint rather than an excuse.

Where you cannot run one, the honest position is that your view through conversions contain some genuine influence in an unknown proportion. That sentence is unsatisfying and considerably more accurate than the alternative.

What view through conversions should change about your budget

One practical consequence follows from all of this, and it is about comparison rather than deletion.

Rank campaigns on click-driven conversions when deciding where the next thousand goes. That is the part of your reporting resting on an observation rather than an assumption.

Then look at the view through column separately, as context. A campaign producing large view through volume and almost no clicks is doing something, and what it is doing has not been established by the numbers in front of you.

The mistake to avoid is the blended ranking. Sorting all campaigns by total conversions puts display and awareness activity above search on a measure that flatters them structurally, and budget follows the sort order.

I would rather run awareness activity and judge it on branded search volume and direct traffic than judge it on a conversion count built from impressions. Those measures are cruder and they are not assuming the answer.

What to change this week

Three steps.

Find out what share of your reported conversions are view through, per campaign type. Most people have never separated it and the answer on display is often startling.

Then take them out of your headline figure and report them as their own line.

Then apply the power check to any test you are shown, including your own. Base rate times sample size tells you what to expect under no effect, and if that number is around one, the test proves nothing either way.

What conversion numbers actually measure is the counting underneath, attribution windows covers the setting that decides which of these get counted at all, attribution models covers how credit gets divided, platform reported conversions explains why the platform claims what it claims, and cost per lead by channel is where blended figures do the most damage to budget decisions. The case studies show results reported on a stated basis. To find out how much of your reporting is view through, book a teardown.

Frequently asked questions

What are view through conversions?

They credit somebody who was shown your advert, did not click it, and converted later within a window. The claim is that seeing the advert influenced the purchase. That claim is plausible in some cases and impossible to verify from the data itself.

Are view through conversions real?

Some certainly are, and the reported count is not a measurement of how many. The problem is that the same data appears whether the advert caused the purchase or merely happened to be shown to somebody already about to buy.

Should I include view through conversions in reporting?

Report them separately if you report them at all. Blending them into your headline conversion figure means part of your number rests on an assumption rather than an observation, and nobody downstream can tell which part.

What was the WordStream view through experiment?

A display remarketing campaign ran for thirty days producing nine view through conversions from 2,418,973 impressions. A clone with a completely blank banner then produced zero from 315,677 impressions. The conclusion drawn was that they can be trusted.

What is wrong with that conclusion?

The test had almost no chance of detecting anything. At the observed rate, 315,677 impressions would be expected to produce roughly one conversion. Seeing zero when you expect about one is an unremarkable outcome, not evidence about the original nine.

How many impressions would the test have needed?

Many times more. To distinguish a real effect from nothing at that base rate, the blank banner arm needed to run long enough to expect a meaningful number of conversions, not roughly one. A null result from an underpowered test carries no information.

Do view through conversions matter more on some campaigns?

Yes, they dominate on display, video and awareness activity, where most people never click. On search they are a rounding error. So the same setting has wildly different effects depending on which campaigns you run.

How can I actually test whether they are incremental?

A controlled experiment where a comparable group is deliberately not shown the adverts. That is what incrementality testing does. Nothing in your reporting interface can answer the question, because the interface only sees people who were shown adverts.

Want a second pair of eyes on your account?

Book a free teardown

Book a free teardown

Send me your worst-performing campaign.

I will tell you what is wrong on the first call. No obligation, no hard sell, and I will say honestly if I am not the right fit.

Book a free 20-minute account teardown

Replies within one business day, Lahore time.