Two Promises That Don’t Travel Together
Transparency and explainability sound like the same virtue wearing different coats. They are not. A system can tell you exactly what it did and still leave you unable to say why. It can also explain its reasoning in fluent, confident prose while the explanation drifts free of the mechanism that produced the answer. Marketing has just walked into that gap, and the walk is not hypothetical.
The cosmetics giant L’Oréal is now treating AI as a channel in its own right, not a back-office convenience. Its chief digital and marketing officer, Asmita Dubey, framed the shift around a billion-plus consumers already using the technology, which means brands have to work out how to reach people through it. [1] That is the visible side: a brand buying presence where the customer now looks.
The invisible side is where the deception lives. When an AI assistant answers a shopper’s question about a serum or a shampoo, it assembles an answer from product descriptions, reviews, research and whatever else it can retrieve. The answer reads as settled fact. Nothing in its tone signals that the underlying data may be stale, incomplete or simply absent. The consumer sees authority; the brand may be seeing a garbled version of itself, and neither party gets a warning label.
What the Referral Numbers Actually Say
There is a number circulating that makes the case for spending more on AI visibility, and it deserves careful handling. Adobe’s Q2 2026 report found that in March 2026 retail visitors referred by AI converted at a 42% higher rate and generated 37% more revenue per visit than non-AI traffic. [2] On its face, that is a compelling argument for pouring budget into the new channel.
Read it again with the middleman in mind. Those figures describe visitors arriving through AI referrals, not the actual performance of paid advertising placed inside AI search experiences. [2]. Higher conversion among that group does not establish that buying ads in an AI answer produces extra sales. The gap between what a statistic appears to prove and what it actually measures is the same gap that runs through the whole AI marketing conversation, and it is wide enough to drive a budget through.
The honest reading is narrower and more useful: visitors who come through AI arrive with a question already half-answered, which may make them readier to buy. Whether the brand controlled that half-answer is a separate question entirely. Confusing the two is how a measurement artifact becomes a strategy.
The Infrastructure Nobody Audits
Beneath every campaign sits plumbing: the SaaS tools that manage customer relationships, websites and media buying. AI is being folded into that layer, and the folding is mostly invisible to the marketer who clicks the dashboard. What changes is how those tools share data and how their decisions get checked — or don’t.
This is the quiet version of the deception problem. A recommendation engine inside a marketing stack does not announce its reasoning. It returns a segment, a bid, a piece of creative. The marketer sees an output and assumes a process. The process may be drawing on fields that were never cleaned, joining records that were never meant to meet, or weighting a signal that made sense in a different context. The tool is not lying. It is simply not volunteering the part where things went sideways.
The logic is sound: Google’s guidance states that established SEO practices remain relevant to AI search features, and that up-to-date, accurate product information and website content can help products appear in paid ads, unpaid results and AI answers alike. [3]. That is also the premise behind generative engine optimisation, the practice of making scientific research, brand content and product data legible enough for AI systems to interpret and reuse in consumer answers.
Notice what GEO concedes. The brand is no longer writing for a human reader who will judge tone and trust. It is writing for a retrieval system that will slice the text into fragments and reassemble them elsewhere. The original context — the caveat, the qualifier, the “results may vary” — is exactly the material most likely to be dropped in reassembly. What survives is the claim without its conditions.
The Confidence Problem
Ask an AI assistant about a moisturiser and it will often answer with the poise of a dermatologist. It may cite mechanisms, compare formulations, recommend a routine. Some of that will be right. The trouble is that the fluent register is identical whether the underlying evidence is a peer-reviewed study or a single enthusiastic review that happened to rank well.
For a brand, this cuts two ways. A well-structured product page can be lifted into an answer and presented as fact, which flatters the brand. A misread spec sheet or an outdated ingredient list can be lifted with the same confidence and presented with the same authority. The system has no mechanism for signalling “I am not sure about this particular attribute” in a way the shopper will notice. Uncertainty gets smoothed into prose.
That may be the deception in its purest form: not a lie, but a flattening. The model may not know which of its retrieved fragments is load-bearing and which is decorative, and it may have no incentive to say so. The shopper reads a confident paragraph and treats it as a settled answer. The brand, meanwhile, has no direct line to correct the record in real time.
That is a different competence from writing a brief or buying media. It is closer to reverse-engineering a referee.
Where the Responsibility Lands
AI typically draws information from several sources, including product descriptions and customer reviews, so ecommerce, PR, content and legal all feed the same pool. [4]. If those teams describe the brand differently, the AI will reconcile the differences on its own terms, and its terms are not the brand’s.
That is true and also slightly beside the point. The practices are necessary; they are not sufficient. A brand can follow every rule and still find itself summarised inaccurately by a system that had to guess.
What makes this a governance problem rather than a content problem is the asymmetry of correction. A wrong web page can be fixed in an afternoon. A wrong impression lodged in a model’s retrieval layer may persist across thousands of conversations before anyone notices, and there is no complaint form. The feedback loop that normally disciplines marketing — see the error, fix the error, watch the metric move — is broken at exactly the point where it matters most.
The Experiment That Teaches the Lesson
L’Oréal ran an internal exercise that is more instructive than any framework: twenty of its most senior executives were asked to build an advertisement with AI, start to finish. [1] Dubey’s conclusion from the exercise was blunt: unless you have done it, you cannot imagine how AI works, and you have to be ready to learn, be inquisitive and be hands on rather than look at it on paper. [1].
There is a deception lesson buried in that exercise, and it is not the one about productivity. Executives who build something with a generative tool quickly discover where it fabricates, where it confabulates a plausible detail, where it produces something that looks finished and is quietly wrong. That discovery is hard to transmit secondhand. It arrives through the hands, not the briefing deck.
The company now directs almost 20% of its marketing investment into AI-led work, with return-on-investment measurement already running on AI tools and machine learning. [1] Dubey described the result as a dual muscle of creativity and technology. The phrase is worth taking seriously because it implies a check: creative judgment on one side, technical verification on the other. Neither alone catches the failure mode where a system produces something confident and incorrect.
What the Shopper Never Sees
Step back from the brand’s perspective and the consumer’s experience of all this is stranger than the industry conversation admits. A shopper asks a question, receives a recommendation, asks a follow-up, compares two options, and buys. The exchange feels like advice. It is in fact an assembly of retrieved fragments, ranked by a system whose selection criteria are not disclosed and may not be stable week to week.
The shopper has no way to know whether the product they were steered toward appeared because it was well-suited, because its data was unusually clean, or because it was advertised. Those three routes produce identical-looking answers. That is the gap in its most consequential form: the appearance of neutral guidance sitting on top of a commercial and technical process that is anything but neutral.
Brands are not passive in this. They can shape content and advertising investment. But the selection and presentation of information is increasingly decided elsewhere, and independent reviews and media coverage feed the same pool. Control over brand discovery is now shared among parties who do not coordinate and cannot see each other’s inputs. The word “discovery” flatters what is really a negotiation conducted in the dark.
The Variable That Decides It
None of this argues for retreat. It argues for a specific discipline: treating AI visibility as a claim to be tested rather than a channel to be funded. Before scaling spend, the question is whether target customers actually use the technology to research the relevant products, and whether the investment produces profit rather than merely traffic. The comparison set matters too — money spent on AI search should be weighed against money spent on product-data quality and content, and the results tracked long enough to see whether they hold.
The open variable is not technological. It is whether brands can build the habit of verifying what the middleman says about them, at the speed the middleman operates. L’Oréal’s standardised data and internal experimentation are early moves in that direction, and the wider industry is still working out its own. The tools will keep changing; Dubey’s point that the underlying goal remains meeting consumer needs is the stable part.
But meeting a need requires knowing what the customer was actually told. Right now, the answer arrives in confident prose, and nobody on either side of the transaction can fully audit it. That is the deception worth naming — not that AI lies, but that it presents assembly as understanding, and does so in a voice that sounds like it knows.
The brands that get through this will be the ones that stop asking whether the AI is talking about them and start asking what, precisely, it is saying, and that can show the difference with evidence rather than assertion.