The ClickedOn AI Wrap — Week of 07 Apr 2026

OpenAI released GPT-5.4 this week, and the headline number is the one that should make every operations and marketing leader pay attention. On OSWorld-Verified — the benchmark that measures whether a model can actually drive a desktop through screenshots, mouse and keyboard — GPT-5.4 hit 75.0%. GPT-5.2 managed 47.3% only a few months ago. Human expert testers score 72.4%.

This is the first general-purpose model OpenAI has shipped with computer use built directly into the base model rather than bolted on as a separate tool. It means GPT-5.4 can open software, configure settings, fill forms, navigate browsers and organise folders without needing a custom scaffolding layer.

Why it matters for marketers and business owners: the things you pay people to do inside ad platforms, analytics dashboards, CRMs and CMSs are exactly the tasks OSWorld measures. Campaign uploads, bulk edits, reporting exports, landing page QA, bid adjustments — the repetitive desktop work that eats junior marketer hours is now something a model can do reliably enough to trust with supervision. We are not at full autonomy, but we are past the point where "AI can't really use my tools" is a safe assumption.

The practical move this week is to audit your team's recurring desktop workflows and rank them by how rules-based they are. The tasks near the top of that list are the ones to pilot with a computer-use agent over the next quarter. The tasks near the bottom — the ones that need judgement, context and stakeholder communication — are the ones your people should be doing instead. This is how you get measurable growth from AI rather than a pile of half-finished experiments.

Google DeepMind released Gemma 4 on 2 April, and it is the most important open model drop of the year so far. The family ships in four sizes — 2B, 4B, 26B MoE and 31B dense — all under a fully permissive Apache 2.0 licence. That licence change matters as much as the benchmarks: unlike earlier Gemma versions with custom terms, you can now use, modify and redistribute these models commercially with no restrictions.

The 31B dense variant posted an Arena ELO of 1,452, beating several closed models with far more parameters on human preference scoring. All sizes are multimodal across text and image, the edge models add audio input, and context windows run up to 256K tokens. Google is positioning Gemma 4 as the foundation for the next generation of Gemini Nano, so the same code targets on-device deployment from Raspberry Pi through to high-end GPUs.

For businesses, this is the moment where "running your own model" stops being a research project and becomes a real option for privacy-sensitive workloads. If you have data you would rather not send to a third-party API — customer records, internal documents, confidential campaign data — a Gemma 4 4B or 26B deployment on your own infrastructure now performs well enough to handle drafting, summarisation, classification and retrieval-augmented generation for most use cases. Combine that with the permissive licence and you can finally build AI into products you ship without an ongoing per-token bill.

Yahoo quietly turned on Scout this week, an AI answer engine built on Anthropic's Claude and grounded with Microsoft's Bing grounding API. It is now live across Yahoo Search, Mail, News, Finance and Sports, which together reach 250 million US users and sit on top of 500 million user profiles.

Why this is the most underrated story of the week: until now the AI search conversation has been Google AI Overviews versus ChatGPT Search versus Perplexity. Scout adds a fourth major surface, and crucially it is one most marketers are not tracking. Yahoo's demographic skews older and higher-income than Google's, and the audience overlap with Finance and Mail means Scout will be used for a lot of commercial and research queries — exactly the ones that drive purchases.

What this means for your GEO and AI visibility work: optimising for Google AI Overviews alone is no longer enough. You need to measure citations across at least four answer engines, and the weighting of each depends on where your customers actually are. For Australian brands selling into US markets — or running US campaigns — Scout is now a meaningful new surface. Bing Webmaster Tools already added AI Performance reporting earlier this year, which gives you the first direct signal of when your content is cited across AI-driven experiences grounded on Bing, including Scout.

The practical play is to pull your last 90 days of high-intent content, check it against Bing's AI Performance report, and identify which pages are being cited and which are not. The gap is your roadmap.