From views to sell-through: the metric that proves social sells.

Every brand asks the same question of social and asks it of an instrument that cannot answer. Does this sell? The dashboard replies in seconds - views, reach, engagement, a return figure - and every one of those numbers is a correlation. None describes a cause. The metric that proves social sells is not a metric at all. It is a comparison: a group that saw the work, and a comparable group that did not.

The dashboard was never built to answer the question you are asking.

The cleanest test of that came in Marketing Science in 2023. Gordon, Moakler and Zettelmeyer ran 663 randomised experiments on Facebook, then asked what observational methods would conclude from the same data. True median lifts were 29 per cent upper funnel, 18 mid, 5 lower. Double machine learning read those campaigns at 83, 58 and 24. Propensity matching read them at 173, 176 and 64. Even with platform-grade data, they could not reliably estimate a campaign's causal effect.

The failure is not arithmetic. It is selection. People who watch your work out, follow and then buy were never a random slice of the market; they were already closest to buying. Sixteen experiments in the Journal of Marketing Research tested in 2017 whether liking a brand on Facebook moved purchasing. It did not - the meta-analytic effect was slightly negative. Engagement correlates with sales because it describes who your customers already were.

A view is a delivery receipt, not a statement of desire.

Look at what the number counts. TikTok defines a video view as a play that started, replays excluded. The Media Rating Council counts a video ad impression viewable at half its pixels in view for two continuous seconds. On 24 August 2026 YouTube redefined a view as the moment playback starts, across Shorts, long-form and live; the older, stricter count became engaged views and still governs monetisation.

So any year-on-year view-growth claim spanning that date compares two different quantities. Views are not a lie. They are an input - a record of how much distribution the platform chose to give the work - and distribution moves for reasons unrelated to anyone deciding to buy. They are evidence about the creative, not about the till.

A view records that the platform delivered the file, not that anyone wanted what was inside it.

Proof has a shape, and the shape is always a comparison.

To learn whether social sold anything you have to withhold it from someone. That is the whole method. Run a holdout: one comparable audience gets none of the work, another gets all of it. Brands selling direct can do that at user level. Brands selling through distributors and retail cannot, and the honest substitute is a geo experiment - non-overlapping regions randomly assigned to test and control, as Vaver and Koehler set out at Google in 2011.

Read the result against sell-out data, never shipments: sell-in is lumpy and typically unusable in marketing mix modelling. The standard now is modelling calibrated by experiments, not instead of them - Google's Meridian, open to all users since 2025, turns experiment results into Bayesian priors, so a few real tests discipline the model reading the rest of the year. Platform lift products sit outside that logic; TikTok's Conversion Lift is a service you neither design nor audit.

Almost no brand can actually run this, and pretending otherwise is expensive.

Lewis and Rao examined 25 large advertising field experiments and found the median confidence interval on ROI more than 100 percentage points wide. The answer arrives, and it is unusable. To separate a 10 per cent ROI from zero, they calculated, the median retail experiment would need to be 61 times larger. Advertising effects are small next to how noisy sales are, and statistics does not care how much your question matters.

The practical constraints compound. Meta's own GeoLift documentation concedes that geo tests carry less statistical power than people-based experiments. Clean tests take quarters and cost media you deliberately do not spend. NielsenIQ said it plainly in July 2026: almost no brand has the foundation required for accurate incrementality measurement, and using ROAS as an incrementality answer is a category error, and an expensive one, because it measures efficiency rather than causality.

So hold the work to the standard it can actually meet.

If you cannot run the experiment, do not claim its conclusion. Use views and retention as a leading indicator of creative quality, which is what they honestly measure - whether the opening seconds earned the ones after them. Hold the work to that standard, leave the sales claim unmade until an experiment supports it, and spend your one affordable test on the comparison that settles it.