What pretesting can tell you, and what only the feed can

A brand that pretests everything before launch is ahead of most of the market, and the habit is worth keeping, so the argument here is about what the test is for. A pretest predicts fitness, not success. It answers the questions that can be answered before anyone has seen the ad, whether it is recognisably the brand's, whether the message is legible, whether the claims and disclosures are in order. The question it cannot answer is the one that decides most of the outcome on a skip medium, whether a person who owes the ad nothing will stop for it, and that is only ever answered in the feed, after money has moved.

What a pretest is good for

The first thing a pretest does is isolate the ad from the media. A brand lift study reads the campaign whole, with the platform that sold the media grading its own homework, while a pretest holds everything else constant so for once the ad is the only variable. That is why the forced-exposure methods, System1, Kantar Link, Swayable, each with a test to sell, have real validation behind them. System1's star rating, tested on 101 Facebook ads and 50 Pinterest ads against platform lift studies, found high-star ads had a materially higher success rate on brand awareness gain than low-star ads. A strong score does predict what an ad will do once it has attention.

The second thing it does is catch the checkable before spend attaches, because from that moment every flaw an asset carries is paid for at auction prices for as long as it runs. One recent asset through a pre-flight gate scored well on every dimension and still came back with three items, a typo in the product name in the overlay, top text drifting into the platform's interface zone, and the product arriving after the three-second mark. None was a matter of taste, each would have been paid for on every impression, and each cost nothing to fix before launch.

Where it falls down

The word doing the work in every pretest is exposure. Forced exposure means the respondent was made to watch, and on a feed nobody is made to watch anything. Dentsu's work with Kantar and Lumen found a view past two seconds is effectively a choice not to skip and that voluntary attention works harder than forced, so the step the pretest removes is the step the medium is built around. A social ad has two jobs, reminding, done in the first second with the distinctive assets on screen, and building, which needs the watched view. A skip can still remind, it cannot build, and a pretest cannot tell you which job an ad will get to do, because in the test it always gets to do both.

Three more limits follow. The norms most pretests score against were built up over decades of television and other formats where the audience could not leave, so a strong score is strong relative to ads that never had to earn their viewer. The cadence does not fit, a portfolio of twenty to thirty concepts refreshed weekly cannot route every ad through a panel and a fortnight of fieldwork, so the test becomes the bottleneck that breaks the weekly rhythm, or gets waved through and standards float. And respondents rationalise, the watch-or-skip decision is made in under two seconds by the fast, automatic mind Daniel Kahneman calls System 1, and a questionnaire asks the slow, deliberate one to explain a decision it did not make. The spread the feed then finds inside pretested work is wide. Meta's own test data has the winning creative in 235 brand lift tests at 13% lower cost per ad recall than the average of the others in the same test, inside sets of ads that had all passed whatever gate preceded them.

What a pretest for paid social has to be

The turn is not to drop the test, it is to rebuild it for the medium. Fast enough for every ad, which means scoring against a standard the brand wrote down once rather than convening a panel per asset. Checking the checkable, brand presence in the first second, codes used correctly, message legible with sound off, claims and mandatories in bounds, the frame built vertical with safe zones clear, failures returned as actions. And handing the attention question to the feed, where stop rate, hold rate and watched rate per ad answer it weekly in observed behaviour across every impression, light buyers included. Attention is read there, it is not bought there. The campaign objective stays set for the brand's goal, reach for brand building, never switched to video views to flatter the number, because the system then finds the people who watch anything and the reach narrows.

The in-market purist

The counter-position belongs to performance teams, and it says the feed is the only test worth running, so pretesting is time spent not learning. They are right about attention, and the weekly loop is their argument carried through. Where the position breaks is the deadline. The feed finds out at auction prices, in public, and some of what it finds out should never have shipped, a code failure that built a competitor's memory for a fortnight, a claim compliance would have stopped, and across a portfolio one such mistake is replicated through every variant that inherited it. The feed tests what the pretest cannot. The pretest catches what the feed charges for. Media creates the opportunity to see your ads. Creative determines whether they work. A separate note on how this system's own pretest was built, what it scores and why, is coming.

What to do with this

Pull last quarter's pretest reports and your Ads Manager export side by side, one platform, one threshold. For every pretested ad, note the score, then the stop rate and watched rate it earned in market, and look at the spread, because the gap between the best and worst ad on the same media plan is what the pretest could not see. Then split the gate in two. Write down the floor, the codes and their timing, the claims register, the mandatories, the platform grammar, and score every ad against it before spend. Leave the attention question to the feed, with the objective held on reach and the portfolio optimised weekly. Why data-driven creative only starts after launch is the natural next read. A pretest is a floor, not a forecast.

Questions

We already pretest with System1 or Kantar. Is that not enough?

It is a good floor and worth keeping, and it answers one half of the question. A forced-exposure score predicts what the ad will do with attention it has been given, which is real information about message, brand recognition and emotional response. It says nothing about whether the ad earns attention in a feed against everything else a person could look at, because the respondent was made to watch, and on a skip medium that is the step most ads fail.

The other half is answered weekly, per ad, from your own account, in stop rate, hold rate and watched rate. Run both, and treat the second as the result.

Can a pretest predict which ad will be watched in the feed?

Not reliably, and no method that removes the skip can. The watch-or-skip decision is made in the first second or two, automatically, on recognition and feeling, and a panel that is paid to watch has had that decision made for it. What a pretest can predict is the response once watched, and what it can verify is that the distinctive assets, the claims and the platform grammar are in order before spend attaches, so the attention the ad does earn accrues to the right memory.

How do you pretest twenty ads a week?

By scoring against a written standard rather than convening a panel. The brand's judgement is exercised once, upstream, in setting down what every ad must do, the codes and when they appear, the claims and their bounds, the mandatories, the frame and safe zones, legibility with sound off. Each asset is then verified against that standard before it spends, with the failures returned as a short list of actions, which is work a system can do at volume without the tenth asset of the afternoon getting a softer look than the first.

Sources

• System1 test of 101 Facebook ads and 50 Pinterest ads against platform brand lift studies, as cited by Tom Roach, "The 3rd age of effectiveness", circa 2023. High-star ads had a materially higher brand awareness success rate than low-star ads. System1 sells the test.
• System1 (Test Your Ad), Kantar Link and Swayable methodology pages. All three are forced-exposure methods that predict creative potential under attention and do not measure in-feed attention capture. Each sells the test.
• Dentsu with Kantar and Lumen, as carried in the paid social measurement framework (How to read paid social results, 2026). A view past two seconds is effectively a choice not to skip, voluntary attention outperforms forced. Dentsu trades media on attention metrics, Kantar sells testing, Lumen sells attention measurement.
• Daniel Kahneman, Thinking, Fast and Slow, 2011. System 1 as the fast, automatic mind that makes the watch-or-skip decision, System 2 as the deliberate one a questionnaire addresses.
• Meta, "Creative and Media Excellence", circa 2023. Winning creative against the average of other creatives in the same test, 235 brand lift tests, April 2021 to 2022. Meta internal test data, Meta sells the media.
• Pre-flight compliance gate, generic description of a platform output, 2026. An asset graded before spend on brand impact, creative quality and compliance, with weighted scores and action items. The three-item example is from one anonymised asset. Our own system, interest declared.