Your Best-Tested Claim Might Be the One Nobody Reads

A snack brand is a few weeks away from locking a new pack, and the one thing still open is the claim that sits on the front panel just under the flavor name. The team has ten candidates, ranging from "10g of Protein" to a line about real food for busy mornings that the brand manager is quietly attached to, so they do what most teams would do and put all ten into a claims survey, where a few hundred category buyers read each claim as a line of text and pick the ones they find most and least appealing.

The results come back as a clean ranked list. The protein claim finishes first by a comfortable margin, goes to the designer, and the team moves on.

Most people reading this have likely run a study like that, and the result is valid as far as it goes. What it can't show is what happens once the winning claim is on the pack, where it has to compete with the logo, the flavor name and the photography for a shopper's glance that lab research puts somewhere between half a second and two seconds. Adding a claim also changes how everything else on the pack gets read: a line that ranked first as a sentence might make the pack more persuasive and at the same time make the flavor name harder to find, and there's no way to see that in a ranked list.

Testing the Claim Where It Will Live

When my team started building Claims.AI at Behaviorally, this gap was the thing our clients kept coming back to, and it's why we decided early that the workflow had to begin with the pack rather than the list. A brand team picks the one claim it wants to test. Generative AI then writes alternative versions, some factual and some more emotional or aspirational, in the brand's own voice, and places each one directly onto the real design, matched to the existing typeface and layout, while everything else on the pack stays exactly as it was. If the rest of the pack moves, you can no longer tell what the claim did, and the whole exercise stops meaning much.

So the team ends up with ten rendered packs that differ by a single line, each of which looks like something that could go to print, and a decision to make. Each version is then scored by predictive models for how shoppers are likely to respond to it at shelf, with results back in seconds.

The scoring runs on our 4S Framework, which looks at whether a pack is Seen, Shoppable, Seductive and Selected, roughly in the order a shopper experiences it. Seen comes first for a reason, and the research on attention backs that up. In one eye-tracking study using granola bar concepts, the more attention people paid to a claim, the more likely they were to choose the product carrying it. For each variant the team gets a PackPower Score, a single predictive measure of the pack's sales potential, along with the movement on each of the four S's, the lift or loss against the current claim, and a breakdown of what's pushing the result up or down. That last part is where the trade-offs show up. If the protein claim makes the pack more persuasive but crowds the flavor name, it shows up as a gain on Seductive and a loss on Shoppable, not as one blended appeal score. The team can then decide whether the trade is worth making, with the numbers in front of them.

Why the Data Is the Hard Part

Large language models learned to write from an enormous amount of public text, which is why producing a plausible claim has become fairly easy. Predicting how that claim will perform at shelf, meaning how real shoppers respond to real packs and what those packs went on to sell, is a different kind of problem, because the data it needs has to be collected deliberately, study by study, and it isn't sitting on the open internet.

The models we've built draw on the world's largest behavioral packaging database, which captures how shoppers behave in front of packs, in context, and ties that behavior to in-market sales outcomes. The way it's gathered matters too. It reflects what shoppers did rather than what they said they would do, and when a claim is tested only that claim changes, which is what lets a prediction about one line of copy mean something. We've tested 75,000 packs to date.

Back to the snack brand. Tested on the pack, the protein claim may well still come out on top (it's a good claim), and if it does, the team knows it wins as part of a pack shoppers can find and choose. The line the brand manager liked gets a fair hearing too, and so does the bolder option legal would normally have struck before it ever reached a survey. People still decide what fits the brand and what legal can support; they just decide with the pack in front of them.

Whether or not a team uses a tool like ours, it's worth asking of any claims result where the winning line will sit on the pack, and what it will push out of the way to get noticed. Every claim gets judged on the pack in the end, by a shopper with a second or two to spare. The only real choice a brand has is whether it sees that judgment before or after the print run.