Landing Pages Took Days. Robinson Taught AI Its CMS to Build Them in Seconds

Robinson Club, a resort brand of TUI Group, cut landing-page and competitor-analysis work from one or two days to seconds using an AI experimentation workflow built on Optimizely's Opal AI layer, according to a customer story published by Optimizely. Conversion optimisation manager Michael Richter fed the system Robinson's tone of voice, product details, page frameworks and behavioural-pattern examples, and taught it which CMS components — FAQ sections, selling-point lists and teaser modules — it was allowed to assemble pages from. Optimizely's account gives no independent performance measure and does not say how many generated page versions reached a live A/B test.

6 min read Robinson Club taught AI its brand rules and CMS blocks, and says page work fell from days to seconds. Anyone who has waited two days for a landing page knows how fast a good test idea goes cold. At Robinson Club, a landing page or competitor analysis could take one or two days before a test was ready to move. Its new AI experimentation workflow can produce the first usable version in seconds. That speed claim comes from a customer story published by Optimizely https://www.optimizely.com/field-notes/customer-stories/robinson-club-2025 , the software vendor behind the system. The more interesting part is how Robinson constrained the system, so the output matched the brand, the page structure and the components its content team could publish. The central question is whether those constraints turned a generic writing assistant into a useful part of the testing stack or only made more copy arrive faster. Generic prompts were not good enough Robinson https://www.robinson.com/en/en is a resort brand of TUI Group, the European travel company. Its conversion optimisation manager, Michael Richter https://de.linkedin.com/in/michaelrichter2 , runs experimentation across five TUI brands and more than ten European languages, according to Optimizely’s account. Richter began by asking Opal, Optimizely’s AI layer, for copy and test ideas. The early results came back too generic. He then fed it the missing operating context. That meant Robinson’s tone of voice, product details, page frameworks for different jobs, and examples of when to use particular behavioural patterns. He also taught the system which CMS components were available, meaning the building blocks in Robinson’s content management system CMS , such as FAQ sections, lists of selling points and teaser modules. That changed the job. The model had to assemble a page from an approved kit instead of returning attractive prose in a blank document. In a LinkedIn post https://www.linkedin.com/posts/michaelrichter2 opticon-opal-ai-activity-7379142228738469888-Kq65 , Richter described the workflow as having evolved from simple brand-voice copy into complex landing pages, page analysis and competitor benchmarking. The post supports the direction of the work, although it gives no independent performance measure. The CMS became the constraint Optimizely says the resulting AI landing pages arrive in Robinson’s tone, use an appropriate page framework and are built from components the team already has. Richter told the vendor that the output can be copied directly into the CMS, with an explanation for each decision. This is the most transferable part of the experiment. The AI needed a clear list of what the system was allowed to build, and it received one. The component library acted as a boundary between generation and production. Operators interviewed in our report on when to constrain AI and when to let it fail https://industrycontents.com/when-to-constrain-ai/ describe where they draw their own lines. The company-reported time saving is large. Work that previously took one or two days now takes seconds. Robinson can also request multiple page versions for A/B tests. That is a material change in experiment velocity , although the published account does not say how many of those versions reached a live test. That missing number matters. If review, legal checks, implementation or traffic allocation stay slow, quick drafts will not speed up testing. A comparable team should measure the whole route from brief to launched variant and look beyond the model’s response time. Fyxer’s experimentation workflow https://industrycontents.com/ai-experimentation-workflow-fyxer/ places the useful AI work between idea and decision. Agents checked the agents Richter also built agents to analyse Robinson pages and competing offers. The vendor says they compare storytelling, visual presentation, search engine optimisation SEO and accessibility, then produce summaries and recommendations. Another set audits brand voice, factual accuracy, spelling and grammar. The reported workflow adds probability or confidence scores to some assessments. A score cannot prove an answer is correct, but it makes uncertainty visible to the editor and creates a place to set review rules, such as requiring a human check below an agreed threshold. That makes AI quality control part of the growth system and takes it out of the editorial afterthought pile. The generator produces a page, the checking agents inspect specific failure modes, and a person decides what is safe and useful enough to test. Kuaishou’s A/B Agent https://industrycontents.com/kuaishou-a-b-agent-growth-tests/ shows why that review matters, since its first two experiments hurt key metrics. Richter describes the output as a starting point that removes heavy lifting, and the final call stays with people. The revenue number belongs to older tests Optimizely places the AI work beside more than €4 million in revenue from Robinson’s broader experimentation programme. The chronology matters here. The two revenue examples in the case study came from earlier customer research and predate any AI-generated pages. One test made family room types visible even when they were unavailable for a shopper’s selected dates. Optimizely reports €2.5 million in additional revenue while the test ran. Another moved the coupon field from the payment step to the first checkout step and reportedly generated €1.7 million during testing before being rolled out to all users. Those figures show that Robinson has an established experimentation programme and can connect tests to commercial outcomes. The case study gives no conversion result for an AI-generated page, no number of live AI-assisted tests, no editing or rejection rate, and no manual control group. Nothing in it shows that Opal increased conversion or revenue. The sourcing also leans on the vendor. Optimizely supplies the software and published the case. Richter’s public post corroborates that Robinson uses the agents for page creation and benchmarking, but the detailed speed and workflow claims still come through the vendor. Copy the context pack before the agent A growth team can test the operating idea without a machine-learning group. Pick one repeatable page type and build a small context pack from six ingredients. These are approved brand language, factual product inputs, the page frameworks the team uses, examples of good application, the available CMS blocks and a named review checklist. Then compare ten manually prepared variants with ten AI-assisted variants, and blind the quality review where practical. Track working hours, factual defects, brand corrections, the share of drafts that reach launch and the time from brief to live test. Only after that should conversion decide whether the faster workflow creates better experiments. The same design works with any model or orchestration tool. The part worth copying is the constraint layer, made of a bounded source of truth, an output schema that mirrors the CMS and separate checks for known failure modes. Switching on an agent without those inputs tests the model’s improvisation. Supplying them tests whether the team can compress reliable preparation work. Robinson’s evidence supports a narrower conclusion than the €4 million figure suggests. Its AI experimentation workflow appears to have collapsed the time needed to produce a testable starting point. The next proof point is whether that speed survives human review and raises the number of sound experiments reaching customers while their quality holds. Sources - Optimizely customer story on Robinson Club https://www.optimizely.com/field-notes/customer-stories/robinson-club-2025 - Michael Richter’s practitioner account https://www.linkedin.com/posts/michaelrichter2 opticon-opal-ai-activity-7379142228738469888-Kq65 - Michael Richter on LinkedIn https://de.linkedin.com/in/michaelrichter2 - Robinson https://www.robinson.com/en/en Compare the evidence and operating lessons from more AI experiments in our Growth Signal Index https://industrycontents.com/ic-lab/growth-signal-index/ , or browse benchmark tooling in the Industry Contents Lab https://industrycontents.com/ic-lab/ .