AI: 54 findings, with sources

The Four Modes Framework (PHD) maps categories on two axes - who chooses (people vs. agents) and brand control vs. brand influence - yielding four coexisting marketing modes: Brand→Consumer, Consumer→Consumer, Agent→Consumer and Agent→Agent. Each requires different strategy and capability design.

Categories most vulnerable to agentic disruption are those highly mediated by information - travel, beauty, electronics, CPG categories relying on reviews and recommendations - where discovery, comparison and purchase collapse into one step.

Trust signals become ranking signals in agentic commerce: reviews, complaints data, regulator ratings, customer service scores and transparency all influence whether an agent recommends one provider over another.

AI agents optimize for the best price over time and may not be swayed by temporary deals or complicated bundles - promotional pricing loses power in agent-mediated categories.

Dove's campaign publishing unedited Reddit reviews drove sales up 108.4% year-over-year and made it the most ChatGPT-recommended hair mask, appearing 58% of the time and being recommended 89% of the time.

Dove invited Reddit's brutally honest community to review its hair mask, publishing the first 50 reviews unedited as the campaign: sales +108.4% YoY, #1 presence on ChatGPT (58%), recommended 89% of the time. Authentic community conversations feed AI visibility.

Share of Model is not share of mind: in the US EV category, Rivian scores high on AI visibility but low on consumer brand awareness. Models and human minds hold different brand rankings, so both need measuring. Share of Model is Jellyfish and Brandtech's concept, metric and platform.

A ski brand proudly led with 'rigidity' - a prized expert attribute - but the LLM, trained largely on Reddit and YouTube where rigid reads as stiff and negative, treated it as a con and suppressed the brand's visibility. Audit whether your category's expert language means something else in the model's training corpus.

Make your brand legible to models by flattening it - plain language, descriptive, functional tones - without losing its texture. Models generate answers by assembling brands 'in the moment' from a cross-section of sources; be an influential source in that mix.

People love disruption in content; models don't - they want direction and directional cues. Unlike human creativity, model receptivity is testable pre-launch, so message structure for machines can be verified before spend.

Almost every AI Overview appears on an informational query, not a commercial or transactional one

Google AI Overviews skew overwhelmingly toward informational queries. Ahrefs distribution: 97.70% of AI Overview appearances are on informational intent against 74.15% of normal search, while navigational drops from 18.87% to 1.23%, commercial from 37.00% to 12.86% and transactional from 19.78% to 2.85%.

When an AI agent writes the shortlist, the fight moves up the funnel: brands must influence the criteria the agent uses - value, trust, suitability, service, transparency and prior customer preference.

People may not ask an agent for 'a credit card'; they may ask for 'the safest card for travel with no hidden fees.' Winning brands will have propositions clear enough for both people and agents to understand.

Marketing to machines requires three foundations: structured machine-readable data, distinctive differentiated brand assets, and a unified brand story that resonates with both humans and AI interfaces.

Human and AI effectiveness scores of 480 Cannes ads show zero correlation.

480 Cannes-entered ads scored on effectiveness by humans and by Gemini show zero correlation - 'the regression line is totally flat.' AI does not judge creativity like a human and must not be treated as a surrogate human judge.

Human perception of advertising is wide beam (gestalt, story, emotion, semiotics, the gaps between information); LLM perception is narrow beam (features, claims, structure, description, detail). Humans are barely interested in information - LLMs are interested in little else.

The slow-motion experiment: humans repeatedly prefer slow-motion product videos (immersion, pleasantness); the LLM prefers regular motion - its slice-by-slice analysis reads slow-mo as carrying less information. The same asset gets opposite verdicts from the two audiences.

Do not over-optimize creative for machines: packing communication full of information to please the models risks 'a second crisis of creative effectiveness' - repeating the performance-marketing over-rotation of the first digital revolution. The challenge is improving model receptivity without damaging creativity.

Map where LLMs sit in your customer journey: they are systematic, product-centric deciders, so structure evaluation-stage messaging for machine comparison while keeping brand-building creative human-first.

The shared ground between human and machine audiences: distinctive brand assets and category entry points work for both. If your brand is indistinguishable from a competitor in an evaluation, the model will not choose you.

For AI-era prompts, consider whether your category is queried by product or by context: a high-end jeweler may win more model visibility framed around occasions (weddings, engagement) than around product heritage - match your content to how people actually prompt.

NatWest converted 10,000 real consumer voices into schema-tagged data seeded across AI platforms, lifting AI-visible answers by 508% and driving 500,000 new savings accounts, up 168% year-over-year.

NatWest captured 10,000 real consumer voices about savings, transformed them into schema-tagged semantic datasets and seeded them across high AI-authority platforms: +508% growth in AI-visible answers, #1 bank for machine-cited savings advice, and 500k new savings accounts (+168% YoY).

Magalu's AI-powered virtual influencer Lu reached 4.5 million WhatsApp users in five months, generating R$28 million in sales at 3x the conversion rate of Magalu's own app.

Magalu gave its 23-year-old virtual influencer Lu an agentic AI brain on WhatsApp: 4.5 million users in five months, R$28 million in sales, conversion 3x higher than Magalu's own app. A beloved brand asset can become the AI interface itself, turning agentic commerce from cold automation into cultural connection.

AI agents reward genuine advocacy - brands should earn recommendations, not manufacture them.

Do not over-optimise for machines at the expense of humans: agents will influence behaviour, but people still value their own decision-making and people, not agents, ultimately buy. Brand building remains core even in agent-heavy categories.

Humans are heuristic machines (35,000 decisions a day, symbolic interpretation, implicit cultural association); AI thrives on complexity, description, hierarchy and rankings. Marketing to machines means feeding structure, not stories.