How we cut influencer CPA by 55% for Gymshark

The brand is built on influencers. There are hundreds of them, across several Eastern European countries, and at some point the team simply stopped keeping up. Every ambassador has to be found, assessed, agreed and tracked. With twenty people it worked. At two hundred, CPA started climbing and lead quality started falling. The classic ceiling of a manual process.

They came to us with a framing we hear often: "the channel seems to work, but scaling it is frightening." Here is what we did and why it worked — without embellishment, because some of the decisions looked questionable until the numbers came in.

Why reach and likes lie

The standard logic for picking an influencer: look at reach, look at engagement, look at attractive statistics. The problem is that all of those metrics describe popularity, not a willingness to buy. A creator with a million followers and plenty of likes may sell nothing, while a niche author with an audience of thirty thousand delivers a steady stream of orders.

We started from a simple premise: reach does not buy anything, audience-to-product fit does. Which means an influencer should be assessed not by how many people see them, but by how closely those people resemble the ones already buying from the brand.

A model instead of a manager's intuition

We built a model that predicts not reach but the expected conversion rate of a specific collaboration. Inputs: the influencer's audience patterns and the profile of the brand's actual buyers. Output: an estimate of how many orders this creator is likely to bring, and at what cost.

The key word is patterns, not likes. The model looked at audience structure, subject matter and behaviour rather than vanity metrics. The manager still made the decision, but now with a forecast in front of them rather than a feeling that "this one looks alright".

With twenty people, a spreadsheet copes. With three hundred, the spreadsheet stops being a tool and becomes a source of errors.

Attribution you can trust

A good model is useless if you do not know what actually happened after a post went up. With influencers this is a sore point: someone sees a story and buys three days later from a phone — try connecting the two.

We tied every collaboration to a unique promo code linked to server-side events. Not simply "10% off with code BLOGGER", but a code that connects at server level to a specific session and order. That gave honest attribution down to the session — showing not just how much an influencer sold, but the real cost of acquisition rather than the one last-click paints.

The model answers "who to work with". Attribution answers "did it work". Without the second, the first is guesswork.

Automated budget allocation

The last piece was the one the team initially received with suspicion. We arranged for successful collaborations to scale without manual sign-off. As soon as a pairing hit its targets, its budget increased automatically, under rules agreed in advance.

The fear is understandable: "what if the system scales something it shouldn't". So the rules were strict, with ceilings and stop conditions. In exchange, the main bottleneck disappeared — a person who could not physically review hundreds of collaborations every week. The system handled the routine; managers made the strategic calls.

What transfers to other brands

Three things, in descending order of how universal they are.

1. Honest attribution first. Before optimising anything, know what actually happened. This applies at any scale and is usually the cheapest fix available
2. Assess by audience fit, not by reach. Even without a model, comparing a creator's audience against your buyer profile beats sorting by follower count
3. Automate the routine only after the process is measured. Automating an unmeasured process scales the error rather than the result

We build this kind of automation where the volume justifies it, and fix attribution where it does not yet exist.