StackCheck landing page
This is a record of a landing-page redesign for StackCheck, an AI nutrient-tracking app. The purpose of the project was not to prove the new page was better, but to see what we could learn from putting it in front of a small amount of paid traffic.
The page was redesigned, the funnel was instrumented in PostHog, and a $100 Google Ads budget was set aside for an initial test.
At the time of writing, the campaign has not run yet. So the observations below are a mixture of things we know from the existing site, assumptions we are making, and predictions we have deliberately written down before seeing the results.
the principle
Don't claim something worked unless the data supports it.
A landing page can look better, feel clearer, or seem more convincing without actually improving conversion. We wanted to separate those subjective judgments from what we could measure.
The $100 budget is also too small to make strong claims about conversion performance. We expect it to give us some useful signals, but not enough traffic to establish statistical confidence between two versions of a page. That limitation is part of the finding, rather than something to hide.
The existing StackCheck landing page had relatively little traffic and no meaningful conversion tracking. Over approximately one quarter, search generated:
There is another important detail in those 14 clicks: 11 were searches for the StackCheck brand itself. That means the existing search traffic tells us very little about whether the landing page can persuade someone who is unfamiliar with the product — most of the people reaching it were already looking specifically for StackCheck.
This left us with a basic problem. We had opinions about what the page should communicate, but very little evidence about how unfamiliar visitors would respond to it. The redesign therefore became partly a design project and partly a measurement project.
Rather than redesign first and measure afterwards, we decided to set up the measurement before launching the new page. There were three parts to this:
Define the funnel
We first decided what actions would count as useful signals. The main conversion we can observe on the website is a click through to the App Store or Google Play. We cannot currently see whether someone actually installs the app after leaving the site, so we treat the store click as the end of the measurable website funnel — not as an installation or acquisition.
Get some controlled traffic
Organic search was generating too little traffic to learn much from in a reasonable amount of time. The $100 Google Ads budget is intended to create a small, more predictable sample. At roughly $1–4 per click in the health category, we might see somewhere around 25–60 visits — not enough to confidently answer whether the new page converts better, but enough to observe how people move through it and where they stop engaging.
Write down our expectations
Before spending the budget, we wrote down three predictions. This was mainly to avoid changing the interpretation after seeing the numbers. If something doesn't happen, that should be recorded as a failed prediction rather than quietly becoming a different hypothesis.
The previous "how it works" section used a fairly standard four-step icon layout. We replaced it with a scroll-based paper-folding sequence. A folded paper parcel becomes the hero bowl; as the visitor scrolls, different ingredients are identified, the interaction moves into an individual ingredient, and the sequence eventually shows a nutrient the meal is missing. The product is then introduced as the thing that can fill that gap. The intention was to explain the product through one continuous interaction rather than a collection of separate feature statements.
Why paper?
The paper-craft style was a deliberate choice, not just an attempt to make the page more visually interesting. Photography and conventional health-app imagery would make the page look more similar to other products in the category; the paper treatment gave the brand a more distinctive visual language. But this is an assumption. We don't yet know whether being distinctive makes the page more effective — it could just as easily make the experience slower or less immediately understandable. That is one of the things we wanted the test to help us understand.
The main funnel is instrumented in PostHog. Every stage below is a real event; the values fill in once the campaign runs.
The most interesting event for the redesign is probably scroll_100. The page asks visitors to spend time moving through the story before reaching the main payoff and CTA. If a large proportion leave before getting there, that would suggest the interaction is asking too much of them. If people regularly reach the end, that's some evidence the length is tolerable. Neither result, on its own, tells us whether the page converts better — only something about how people are experiencing it.
We are also recording mobile and desktop behaviour separately. The site is designed primarily around mobile use, so a large difference between the two would be useful to identify.
We wrote these before the campaign started, so the interpretation can't quietly shift after the numbers arrive.
Locked 2026-08-23 in commit fcfe5093 — before the campaign runs.
Prediction: at least 50% of visitors who land on the page reach scroll_100. The reasoning is that the visual sequence should be understandable enough to keep people moving through it. If substantially fewer reach the end, we would reconsider how much interaction the page requires before explaining the product. This is a prediction about engagement with the page, not a prediction that the design will increase conversion.
Prediction: we allocate about $30 to compare two ad framings — food-first vs supplement tracker — to see which attracts more clicks.
Limitation: the two groups may not reach us with exactly the same search intent or keywords, so we can't treat this as a clean test of messaging alone. If one framing gets a higher CTR, the reasonable read is that it appears to attract more of the particular demand we reached — not that we've definitively found the better message. Conversion data from this comparison is likely too limited to interpret confidently.
Prediction: the mobile store_click rate lands reasonably close to desktop. The mobile version uses a different treatment below 860px, including a CTA that appears earlier, which we expect to make the main action easier to reach on small screens. Again, this is a prediction, not a conclusion — a mobile/desktop difference could have many explanations, including traffic source, device behaviour, or sample size.
What the $100 can actually tell us
This is probably the most important limitation. With an expected 25–60 paid visits, we should not expect to prove that one landing page has a higher conversion rate than another — a couple of extra conversions could move the percentage substantially, and there won't be enough observations to separate a real difference from random variation. So the purpose is narrower. We are looking for:
If the data is too limited to answer one of these, that result should be recorded as underpowered, rather than turned into a positive or negative conclusion.
There is also a longer-term question about automation. We have previously worked on a system where experiments can be evaluated against a deterministic ground truth. Conversion optimisation is different, because there is no equivalent free and immediate answer to "did this page perform better?" Every conversion experiment requires traffic, and the smaller the sample, the more uncertainty around the result.
For that reason, we don't think there is much value in automating the optimisation process at this stage. The more useful thing for now is to establish a reliable measurement process and understand what the traffic is telling us. If traffic eventually becomes large enough to support more reliable experiments, that could change.
current status
There isn't a conclusion yet.
The redesign is live in preparation for the campaign, the funnel is instrumented, and the predictions have been recorded before the traffic arrives. What we have at this point is a set of hypotheses, not evidence that the new page is better.
The interesting part will be seeing which of our assumptions survive contact with the data — and being willing to leave some questions unanswered if $100 isn't enough to answer them.