Test It in 24 Stores Before You Trust It in 4,000

Grupo Boticário tested AI skin analysis in 24 stores, reported an 80% increase in average skincare order value, then expanded it to roughly 4,000. The lesson for AI in marketing is simple: prove that a tool improves real work before giving it more reach, money or autonomy.

A lot of AI marketing still gets demonstrated backwards. The impressive bit comes first: generate more assets, automate more decisions, connect more agents, then work out whether the output actually helped. This week's stronger examples followed a different order, starting with something already worth improving and asking AI to make that work more useful.

Start with something proven

Google is now using generative AI to create missing aspect ratios for existing Performance Max video ads.¹ The advertiser provides a video it already chose to run, and the platform handles some of the repetitive work required to make that asset fit other placements. The idea, product, footage and basic creative direction already exist. The machine extends work that has already earned a reason to exist.

Designkit launched with a similar premise. Its system turns existing ecommerce product photography into marketing videos rather than asking a business to begin every asset with a blank prompt.² The useful input is something the brand already owns and recognises. AI handles adaptation around that source instead of pretending the source itself is disposable.

Higgsfield shows why this distinction is becoming more important. The AI video company reached a reported $5.4 billion valuation after rapid growth in demand for generated video, with business customers contributing heavily to that rise.³ Production capacity that once required specialist teams is becoming available on demand. That is valuable, but it also means another video is becoming easier for everyone to make.

Taken together, those stories suggest a better use for content automation. Start with the product photo, the customer conversation, the creator idea or the campaign concept that already contains something specific, then let AI handle the repetitive adaptation around it. As production gets cheaper, source quality matters more because the source is carrying more of the distinction. A weak idea multiplied into ten formats is still a weak idea.

Give autonomy in stages

The same discipline appears further down the marketing workflow. Xnurta introduced an MCP connection that brings retail-media data into assistants such as ChatGPT and Claude, with write capabilities on its roadmap.⁴ That means an assistant could eventually move from explaining a campaign to changing one. The consequences change when software gains permission to act.

Fluency provides a useful comparison. The company says its platform manages about $3 billion in annual advertising spend, while its architecture separates AI-generated recommendations and creative work from deterministic execution against live budgets.⁵ Campaigns still launch after human review. Exploration can remain probabilistic while the parts capable of spending money follow tighter rules.

That approach reflects the fact that marketing mistakes have different costs. A poor suggested headline can be rejected before anybody sees it, while a poorly controlled budget action can spend real money across accounts before a marketer understands what happened. A system should earn more freedom where the errors are cheap, visible and reversible. The boundaries should get tighter as the financial or reputational cost rises.

Permissions encode marketing judgement. The person setting spend limits, approval thresholds, market exclusions and brand constraints is deciding where experimentation is acceptable and where predictability matters more. That makes governance part of the marketing design rather than a technical setting added after the workflow is built. Teams that make those decisions early can automate more confidently because the machine is operating inside a boundary somebody understands.

This is a more useful way to think about agentic AI than asking whether an agent can run a campaign by itself. The better question is which parts of the campaign can safely become autonomous first. Analysis may need very little supervision, recommendations may need review, and live budget changes may need deterministic rules. Autonomy works better as a staircase than as a switch.

The brand still has to sound like itself

The efficiency gains become less useful if customers can see the machinery. Rival Technologies reported that 72% of Gen Z respondents in its research had taken some negative action against a brand after encountering AI-generated marketing, including unfollowing, unsubscribing, complaining or abandoning a purchase.⁶ It is a vendor study based on reported behaviour, so the percentage should not be treated as a universal rejection rate. It still gives marketers a reason to notice when automation makes customer-facing work feel interchangeable.

A model cannot preserve details it has never been given. If the instruction is simply to write an engaging caption or advert, the output will lean on patterns that could belong to thousands of businesses. Better inputs are less glamorous: approved claims, real product photography, previous work that still sounds right, customer language, examples of strong creative and a clear sense of what the brand would never say. Those constraints give the model less room to retreat into generic marketing copy.

This is where AI becomes more useful as an amplifier than as a substitute. A genuine customer interaction, a founder explaining a product or a photograph from a real shop floor already contains information that generic generation cannot recreate convincingly. The role of the system is to help that material travel further without sanding away the details that made it worth using. The more recognisable the source, the more useful automation can become around it.

Boticário's in-store pilot makes that point unusually clearly. Beauty advisers used AI skin analysis on the same mobile devices they already used during a sale, and the system produced personalised recommendations while the adviser remained part of the interaction.⁷ The company did not ask the customer to replace a human consultation with a machine-generated experience. It gave the adviser another source of information to work with.

That distinction matters because marketing teams already possess useful source material in product photography, customer reviews, sales conversations, campaign briefs and the accumulated judgement of people who know the audience. AI can make those inputs easier to reuse, adapt and distribute. It becomes less useful when the volume of output grows faster than anyone's ability to recognise whether the work still belongs to the brand. More content does not compensate for weaker identity.

AI discovery adds another reason to care about source quality. Demandbase measured ChatGPT-referred visits to B2B sites in its dataset rising from roughly 645,000 in June 2025 to 2.6 million in June 2026.⁸ PartnerCentric's research found that many AI-influenced shoppers still wanted reviews before acting on a recommendation and that generic recommendations were often ignored.⁹ Assistants may become a larger route into discovery, but they still depend on evidence that exists somewhere outside the answer.

For marketers, that creates two connected responsibilities. Product information, pricing, policies and claims need to be explicit enough for machines to interpret accurately, while customer-facing work still needs judgement, photography, opinion and recognisable language. Filling both sides with generic generated material weakens the evidence available to machines and the identity available to people. A brand becomes easier to produce and harder to recognise at the same time.

Measurement needs a denominator

Scaling a workflow also requires knowing what the number on the dashboard actually represents. XstraStar published an example showing how the same brand could receive very different AI visibility percentages depending on which prompts, engines and opportunities sit underneath the calculation.¹⁰ A score of 38% and a score of 11% can both be mathematically defensible when they use different denominators. That is a poor basis for changing budget unless the marketer can inspect what was counted.

AI discovery metrics are arriving faster than common measurement conventions. Brands are being offered visibility scores, citation rates, prompt coverage, share of voice and other ways of describing how often an assistant encounters or recommends them. Some of those measures will become useful, but a single percentage with no buyer persona, prompt set, source trail or time period underneath it gives a team very little to act on. Precision in the interface does not guarantee precision in the underlying measurement.

Attribution is messy too. An analysis of more than 51,000 Google AI Overview events found that an average 22.4% of the tracked events appeared as Direct traffic in GA4 rather than Organic Search.¹¹ A customer can begin a discovery journey through an AI surface while the analytics system assigns the eventual visit to another bucket. That makes confident channel conclusions harder, especially when teams are eager to prove that a new source of traffic is growing.

Marketers should resist turning every new surface into another volume target. A ChatGPT impression, an AI citation or a visibility percentage is useful when it explains something a team can change. If a score shows that finance buyers rarely encounter the product while developers do, somebody has a place to investigate. If it simply produces a larger number next month, it risks becoming another activity metric.

The same standard should apply to production automation. Generating ten times more material is easy to count, but the number says little about whether the business is better off. Automation should make useful outcomes clearer, not give teams more activity to report. Enquiries, qualified traffic, bookings, repeat purchases and profitable sales remain much closer to the reason the work exists.

Prove it before scale

Boticário's rollout is compelling because the company did not begin with 4,000 stores. It began with 24.⁷ The company says the pilot increased average skincare order value by about 80%, giving it a commercial reason to expand. Even allowing for the fact that this is company-reported performance rather than an independent experiment, the sequence is useful.

The pilot also ran inside the work employees were already doing. Beauty advisers used familiar devices during a familiar customer interaction, while the system added another input to the recommendation process. Adoption did not require thousands of employees to abandon their workflow and organise themselves around a new piece of software. The tool had to fit the job before the job was expected to fit the tool.

Alchemy Cloud's ZoomInfo case provides a B2B version of the same discipline. The company reported roughly fivefold growth in average clickthrough rate and a 24% reduction in cost per click after improving its account data and buying signals.¹² The useful change happened before another layer of creative generation. Better audience information gave subsequent activity a better chance of reaching the right people.

That order matters because AI is exceptionally good at multiplying whatever it receives. Give it a strong product photograph and it can produce more useful versions, while clear brand rules can be applied repeatedly across more work. Give it inaccurate contact data, confused positioning or vague audience definitions and it can distribute those weaknesses much more efficiently. Scale magnifies the quality of the inputs in both directions.

For a marketing team, the practical starting point is one bounded workflow where the source material is strong, the expected result is clear and somebody knows what failure would look like. A team might test whether adapting existing video improves placement coverage, allow an agent to recommend bid changes while requiring approval before money moves, or use AI to improve a customer interaction without removing the person responsible for it. Each creates evidence before the permission or production volume expands.

Earn the 4,000-store rollout

There is a tendency to judge AI adoption by how much of a workflow it touches. That rewards broad deployment even when nobody can say whether the work became better. Boticário offers a more demanding benchmark: the system had to produce a useful result in 24 stores before it earned access to thousands more. Scale came after evidence rather than standing in for it.

That standard works beyond retail. Before a team gives an agent a larger budget, let it demonstrate that its recommendations improve the economics. Before automating publishing, test whether the content still sounds recognisably like the business. Before spending heavily on AI visibility, establish which score connects to customer behaviour rather than to a new reporting category.

The companies in this week's stories are increasingly giving AI more reach. Google can alter existing creative, Xnurta is preparing assistants to act inside advertising accounts, and video platforms can create more assets at rapidly falling cost. Greater capability makes the test-before-scale discipline more valuable because mistakes can now travel further and faster too. The cost of producing an error may fall while the cost of distributing it rises.

Marketers still control the inputs that matter most. They decide which customer evidence is credible, which creative deserves another life, what the brand is willing to claim, how much money an agent can touch and what outcome is good enough to justify expansion. Scale should be the reward for a workflow that has already proved itself. That principle gives teams room to automate aggressively without pretending every new capability deserves immediate trust.

Twenty-four stores before 4,000 is not a timid approach to AI. It is an ambitious approach with evidence attached, and it puts the burden of proof in the right place. Marketing teams that adopt that habit can give machines more responsibility where the results support it while keeping human judgement exactly where the consequences are hardest to reverse.

Sources