{"slug": "boxios-predictive-preloading-test-speed-kit-reports-9-65-higher-conversion", "title": "BOXIO’s predictive preloading test: Speed Kit reports 9.65% higher conversion", "summary": "Speed Kit reported a 9.65% higher user conversion rate and 10.27% more orders in a split-traffic test on boxio.de, the German Shopify store of Rostock camping-equipment maker BOXIO, according to a Speed Kit case study published 22 October 2026. Half of BOXIO's visitors received Speed Kit's predictive preloading layer, which uses customer-specific machine learning models to predict, cache and sometimes render likely next pages before a click, while an equal-sized control group stayed on the unaccelerated site. BOXIO's Shopify performance manager Susanne Lohbeck said she ran her own evaluation on the raw order data before the company rolled the system out, though the public case omits the test duration, sample size, baseline conversion, confidence interval and revenue per visitor, so the lift cannot be independently checked.", "body_md": "10 min read\n\nBOXIO split traffic on its German store between the normal site and Speed Kit’s predictive preloading layer. Speed Kit reports 9.65% higher user conversion, while BOXIO evaluated the raw order data itself.\n\n## TL;DR\n\n- BOXIO ran an equal traffic split on boxio.de, with half of visitors receiving Speed Kit and half staying on the unaccelerated site.\n- Speed Kit uses customer-specific machine learning models to predict likely next pages, cache them and sometimes render them before the visitor clicks.\n- Speed Kit reports 9.65% higher user conversion and 10.27% more orders in the treatment group. BOXIO received the raw order data and ran its own evaluation before rolling the system out.\n- The public case omits the duration, sample size, baseline conversion, confidence interval and revenue per visitor, so the size and precision of the lift cannot be independently checked.\n\nA shopper on BOXIO’s German store might move from a portable washbasin to a cooker, then open another product page before deciding whether a black plastic box deserves space in the van. During BOXIO’s experiment, one of those pages could already be waiting in the browser. The visitor had not clicked yet, but a prediction model had already made an early guess.\n\nThe commercial question was whether removing a few hundred milliseconds from the next page would increase orders enough to justify another technical layer in the shop. BOXIO tested that on live traffic. It divided visitors between the existing site and an accelerated version, then compared what the two groups bought.\n\n## A slow shop beneath a simple box\n\n[BOXIO](https://boxio.de/pages/presse) is a Rostock company that sells camping equipment in the standard Eurobox format. Its products include a separating toilet, washbasin, passive cool box and gas cooker. The company says more than 90,000 campers use its products, and its German Shopify store is part of a group of country sites selling across Europe and beyond.\n\nThe product looks deliberately simple, but the page underneath it was not. BOXIO wanted faster navigation without rebuilding the Shopify store or changing its existing code. It also wanted evidence from its own customers before making the change permanent, with particular attention to mobile conversion.\n\nThat made ecommerce page speed a growth constraint. A faster Largest Contentful Paint matters only if it changes behaviour that the business owns. BOXIO chose user conversion and order count, then kept half of its visitors on the old experience so it had a counterfactual.\n\n## Half the traffic got the shortcut\n\nAccording to the [case study](https://www.speedkit.com/customers/boxio) from Speed Kit, published on 22 October 2026, half of visitors to boxio.de received the acceleration layer and half formed an equal-sized control group. The treatment sat above the existing Shopify store. BOXIO did not need to rebuild product pages or change the checkout.\n\nThe split is the most useful part of the BOXIO conversion test. Slow and fast visitors arrived during the same period and encountered the same products, prices and campaigns. That does not remove every possible source of bias, but it is substantially stronger than comparing conversion before and after a launch, when seasonality, promotion and traffic mix may have changed.\n\nBOXIO also received the raw order data. [Susanne Lohbeck](https://de.linkedin.com/in/susanne-lohbeck), its Shopify performance manager, reposted the case account and said she ran the evaluation herself. That separates the buying decision from the supplier’s dashboard, although the public material does not include the dataset or analysis.\n\n## The prediction engine prepared the next page\n\nSpeed Kit combines several systems. Its caching layer creates an anonymous version of dynamic HTML and stores it away from the origin server. A service worker in the browser can then retrieve that version quickly, while current personalised information arrives from the store and is merged into the page.\n\nThe AI component is [predictive preloading](https://www.speedkit.com/docs/predictive-preloading). Speed Kit says its customer-specific models learn from anonymous navigation data, link position, link size, page structure, scroll behaviour and cursor or touch distance. Server-side models also use journey probabilities and semantic similarity. The system combines those signals to estimate which page the visitor will open next.\n\nWhen the prediction is confident enough, the browser fetches cached HTML for the likely destination. On supported browsers it may also prerender that page in the background, so a click can reveal a page that is already assembled. Speed Kit says the preload request comes from its cache instead of BOXIO’s origin, which limits the extra server work created by a wrong guess.\n\nThis machine learning preloading workflow is useful because it treats prediction accuracy as an operational constraint. Preparing every possible page would waste bandwidth and processing power. Preparing nothing would preserve resources and leave the wait untouched. The system has to guess early enough to create lead time, while remaining selective enough that the acceleration does not become its own performance problem.\n\n## The order data changed the decision\n\nSpeed Kit reports that pages in the treatment reached their main content 459 milliseconds faster at the median, a 43% improvement on the pages it accelerated. Category pages improved 56%, product pages 48% and the homepage 14%. Time to First Byte fell 75% and First Contentful Paint fell 46%.\n\nThe system cached 82% of next pages before visitors opened them. It fully prerendered a page for 36% of users. Those are mechanism metrics. They show that the prediction and preparation pipeline worked often enough to change the experience, but not for every visitor or every navigation.\n\nThe reported business result is a 9.65% relative increase in user conversion for the Speed Kit group against control, plus 10.27% more orders. BOXIO then rolled the system out to all visitors on boxio.de and said its other country stores would follow. For a team watching Shopify conversion rate, that decision is more informative than a faster laboratory score because it connects the infrastructure change to an observable commercial outcome.\n\nThe number still needs careful attribution. Speed Kit built the product, ran the instrumentation and published the case. BOXIO’s independent access to raw orders strengthens the account, but the public page gives no sample size, test duration, baseline conversion, confidence interval, revenue per visitor or device-level result. It also does not say whether assignment persisted across repeat visits or whether the test passed a sample-ratio check.\n\n## The model was only one part of the bundle\n\nThe treatment combined machine learning prediction with dynamic caching, edge delivery, a browser service worker and prerendering. The experiment therefore measures the bundle against the original site. It does not establish how much of the 9.65% came from predicting the next page, how much came from faster delivery of the current page, or how much depended on browsers that support prerendering.\n\nOperators deciding what to copy do not need their own prediction model. BOXIO used a managed layer and kept its internal work focused on integration, traffic allocation and order analysis. The copyable lever is to buy or build the smallest performance layer that can be switched on for a random share of traffic, then require it to earn a full rollout on a business metric.\n\nThis is the same discipline behind the strongest Growth Signal cases. [Javo](https://industrycontents.com/javo-ai-sales-agent-experiment/) randomised an AI sales agent across live visitors instead of comparing two periods. [Spotify](https://industrycontents.com/spotify-conversational-agent-experiment/) built repeatable failure tests for its assistant before treating usage as proof of quality.\n\n## A cleaner second test\n\nBOXIO’s next experiment could separate the components. One group would keep the original store, a second would receive caching and faster delivery without prediction, and a third would add the customer-specific prediction model. Revenue per visitor would be the primary metric, with conversion, orders and average order value explaining where any gain appeared.\n\nThe guardrails should cover origin load, bandwidth, JavaScript errors, stale content, checkout failures and analytics contamination. Device and browser splits would show whether the gain concentrates among mobile visitors or among browsers that support prerendering. A fixed test window and a published confidence interval would make the result easier to compare across BOXIO’s country stores.\n\nThe first test has already answered the decision BOXIO needed to make. The accelerated experience beat the control on orders strongly enough for the team to roll it out. The second can answer the more transferable question, which is whether the model’s ability to prepare the likely next page creates incremental value beyond ordinary caching.\n\n| Experiment detail | What the reporting shows | \n|---|---|\n| Constraint | BOXIO wanted faster Shopify navigation without rebuilding the store and required proof from its own order data. | \n| Intervention | Half of boxio.de visitors received Speed Kit’s caching, service-worker and predictive preloading layer. Half remained on the unaccelerated site. | \n| Reported result | Speed Kit reports 9.65% higher user conversion, 10.27% more orders and 43% faster median LCP on accelerated pages. | \n| Decision | BOXIO evaluated the raw order data and rolled the system out to all visitors on boxio.de, with other country stores planned next. | \n| Limits | Speed Kit-published result with no public duration, sample size, baseline conversion, confidence interval, revenue per visitor or component-level isolation. | \n\n## Sources and disclosure\n\n- Speed Kit’s [case study](https://www.speedkit.com/customers/boxio) of BOXIO, published 22 October 2026\n- Speed Kit’s technical documentation for [system architecture](https://www.speedkit.com/docs/how-speed-kit-works) ,[predictive preloading](https://www.speedkit.com/docs/predictive-preloading) , and[traffic splitting](https://www.speedkit.com/docs/configuration)\n- BOXIO’s [press area](https://boxio.de/pages/presse) and[LinkedIn page](https://www.linkedin.com/company/myboxio/)\n- BOXIO Shopify performance manager [Susanne Lohbeck](https://de.linkedin.com/in/susanne-lohbeck) on LinkedIn\n\nSpeed Kit built the system and published the performance figures. BOXIO’s practitioner says the company received the raw order data and performed its own evaluation, but the analysis, sample size and statistical output are not public.\n\n## FAQ\n\n## What did BOXIO test?\n\nHalf of visitors to boxio.de received Speed Kit’s acceleration layer and half stayed on the normal site. BOXIO compared raw order data before deciding to roll the system out.\n\n## Where was AI used?\n\nCustomer-specific machine learning models predicted the next page a visitor was likely to open. The browser could then cache and sometimes prerender that page before the click.\n\n## What result was reported?\n\nSpeed Kit reports 9.65% higher user conversion and 10.27% more orders for the treatment group, alongside a 43% faster median LCP on accelerated pages.\n\n## Was the result independently verified?\n\nNo. Speed Kit published the case. BOXIO says it evaluated the raw order data itself, but the dataset, sample size, duration, baseline and confidence interval are not public.\n\n## What can another growth team copy?\n\nPut the acceleration layer behind a random traffic split, keep an unaccelerated control, use revenue per visitor or conversion as the decision metric, and require a successful holdout before full rollout.\n\nCompare more measured experiments in the [Signal Index](https://industrycontents.com/ic-lab/growth-signal-index/), browse the [Growth Signal](https://industrycontents.com/category/growth-signal/) archive, or use the benchmark tools in the [IC Lab](https://industrycontents.com/ic-lab/).", "url": "https://wpnews.pro/news/boxios-predictive-preloading-test-speed-kit-reports-9-65-higher-conversion", "canonical_source": "https://industrycontents.com/boxio-predictive-preloading-test/?utm_source=rss&utm_medium=feed&utm_campaign=editorial", "published_at": "2026-11-02 11:00:00+00:00", "updated_at": "2026-11-02 11:21:30.467422+00:00", "lang": "en", "topics": ["machine-learning", "ai-tools"], "entities": ["Speed Kit", "BOXIO", "boxio.de", "Susanne Lohbeck", "Shopify"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/boxios-predictive-preloading-test-speed-kit-reports-9-65-higher-conversion", "markdown": "https://wpnews.pro/news/boxios-predictive-preloading-test-speed-kit-reports-9-65-higher-conversion.md", "text": "https://wpnews.pro/news/boxios-predictive-preloading-test-speed-kit-reports-9-65-higher-conversion.txt", "jsonld": "https://wpnews.pro/news/boxios-predictive-preloading-test-speed-kit-reports-9-65-higher-conversion.jsonld"}}