Your AI campaign won the click. Did it lose the customer?
AI-generated ads can win attention and still cost a brand the sale, the return and the trust. What the 2026 evidence says about the distance between a click and a customer.

The meeting room has been cooled to the temperature of an airport lounge, eighteen degrees, calibrated for the two men in unstructured navy blazers rather than the rest of us in cotton. On the wall screen, Variant 19 glows pale mint at the top of the table. Click-through rate up. Cost per click down by nearly a fifth. Forty-two versions generated between Tuesday’s stand-up and Wednesday’s lunch.
Everyone around the laminate table relaxes in the specific, brittle way salaried people relax when a number arrives to absolve them of having to argue about taste.
Ten minutes later, downstairs by the boom barrier where the diesel heat hits your collarbone like a damp towel, a junior art director and a production runner pass a cigarette back and forth before calling their autos. Neither of them mentions the click-through rate. The art director is still annoyed about the woman’s hand in Variant 19—how she held a steel tumbler of tea by the rim with three weightless fingers, as though steel does not scald, as though milk has never boiled over on a two-burner stove while somebody yelled from the bathroom, as though a kitchen were merely a beige rumour assembled from ten thousand European stock kitchens. "Nobody holds hot steel like that," she says, flicking ash toward the drain. "Anyone who actually drinks at home is going to clock it in three seconds—or worse, they won’t clock it till the parcel arrives and feels cheap."
I have spent the better part of two decades in rooms near brands—writing for them, editing their institutional prose, watching what happens when an organisation falls in love with its own shorthand—and I cannot bring myself to sneer at advertising’s old hunger to seduce. If you grew up queer, or watchful about money, or slightly sideways to the rooms you hoped one day to enter unchallenged, commercials were often your earliest school of social anthropology. Long before I knew phrases like brand equity or funnel progression, I knew how to study an actor leaning against a doorframe in a magazine tear-sheet: the precise looseness of a wristwatch strap, the cut of a linen collar that said old ceiling fans rather than new credit, the half-averted gaze that promised you could step into another class, another skin, and pass without explaining your surname.
Queer people know a great deal about passing, and also about its cousins: camp, glamour, and the counterfeit. Camp winks; it lets you see the sequins stitched onto cheap net and invites you into the conspiracy of wanting beauty anyway. A counterfeit does the opposite. It hopes you are moving too fast in poor light to check the lining.
What unsettles me about the current wave of synthetic advertising is not that machines are making pictures. Advertising has never been a monastery of documentary truth; I remember retouchers bent over glowing monitors at midnight, smoothing collarbones into porcelain and cloning hibiscus flowers onto wintry lawns. What has changed is the severance of desire from consequence. In a study of more than three hundred thousand live Taboola advertisements, researchers from Munich, Columbia, and Harvard accidentally named the whole psychological economy in three words: "AI in Disguise." Across tight quasi-experiments, synthetic display images matched human photography on clicks only on one fragile condition—that viewers mistook them for human work, pulling a 0.79 per cent click-through rate when they passed and sinking to 0.62 per cent once clocked. Stranger still, real photographs made by human hands that happened to look slightly too warm, too saturated, too frictionless—pictures that accidentally resembled a prompt—fared worst of all, dropping to 0.55 per cent.
And one of the visual cues that helped a machine pass in the crowd? A large human face looking back out of the frame.
Even inside a spreadsheet of sixteen billion impressions, people are still scanning the street for a witness. In an October 2026 consumer audit by Cint and Advertising Week, forty per cent of respondents picked realistic landscapes as their favourite AI-generated element, yet forty-six per cent disliked seeing AI-generated people in ads. Paint the backdrop all you like; we know backdrops are painted. Just do not send a mannequin across the footlights to swear the medicine works or the coat keeps out the rain.
An afternoon matinee in the nineteen-nineties, velvet seats worn down to grey hessian at the knees, the projector beam thick with dust. Two minutes of trailer cut so breathlessly—every slap, every brass cue, every drenched embrace—that when the actual film unspooled, slow and badly lit, half the hall grew restless before the interval.
A click is a matinee trailer. It measures the turn of a head, nothing more.
I keep thinking about what an advertisement actually does to a product long before a customer touches the box. Marketing is not wrapping paper added after the factory shuts; it sets the sensory temperature at which the product will be judged. Promise effortless radiance in four seconds of frictionless light, and an honest moisturiser—good enough on its own modest terms—becomes a disappointment on the bathroom shelf by Tuesday morning. Overpromise certainty in a B2B software hook, and six weeks later eight tired people on a buying committee are arguing with your implementation lead about a feature nobody built.
Look closely at the May 2026 test run by Ipsos and Syracuse University across three thousand viewers and ten household brands. Researchers used Gemini to extract the creative briefs behind ten real thirty-second commercials and handed those shot lists to Sora 2 with no human creative intervention. Three out of four viewers watching the AI spots could not say with confidence that a machine had made them, yet the human originals still scored fourteen per cent stronger on immediate sales effectiveness and seventeen per cent stronger on long-term brand health. What gives the game away is where the machine held steady and where it went dead behind the eyes. On Febreze, Herbal Essences, and Cheerios—a smell in the upholstery, clean hair, oats for a family heart—the synthetic versions kept pace. Functional errands survive automation because the bargain is plain. Where the brief asked Fiat for playfulness and defiance, or asked Chewy for the small domestic ache of pets and holidays, the machine produced competence without pulse.
Defiance requires knowing what you might lose if the room turns against you. Humour requires timing learned in bodies that have known awkwardness. An algorithm trained to hug the middle of the distribution cannot risk embarrassment—and creative work that never risks embarrassment rarely survives a purchase someone has to think hard about.
9:14 p.m. A phone lights up face-up on an oilcloth dining table, wedged between a steel tiffin box and a strip of blood-pressure tablets. On WhatsApp, nestled between a cousin’s forwarded prayer graphic and an electricity bill alert, a brightly lit face speaks the shopper’s past purchases back to them in unhurried, synchronised syllables.
In a field experiment with more than twenty-one thousand customers of an Indian eco-friendly online retailer, researchers Madhav Kumar and Anuj Kapoor found that these AI-avatar WhatsApp videos lifted click-through rates by 9.4 per cent over personalised image cards and 6.5 per cent over standard video. It is an arresting engineering feat. It is also an unfinished sentence: as the MIT Initiative on the Digital Economy summary is careful to note, this was a single-exposure experiment, and it reports clicks, not whether anyone bought, kept the goods, came back a second time, or felt a quiet chill at having their shopping history recited by a face that has never breathed monsoon air.
Follow that unfinished sentence out of the marketing deck and into the street, and class re-enters the frame immediately.
Automated delivery engines—Meta’s Advantage+ suite alone passed a seventy-five-billion-dollar annual revenue run rate in mid-2026—do not ask whether a promise was honourable; they hunt the cheapest event you tell them to count. Meanwhile, the seven-day last-click attribution column that media buyers paste into Friday reports predicts only nineteen per cent of the variance in true incremental conversions per dollar across nearly a thousand Meta experiments, overstating real causal lift by roughly a third. When a campaign buys cheap curiosity with synthetic polish, the strategist in the chilled conference room collects the green arrow in Week One. The bill arrives in Week Three, and it is almost never paid by the person who wrote the prompt.
It is paid by the delivery rider climbing four flights without a lift at two in the afternoon, carrying a cash-on-delivery parcel that gets refused on the landing because once the buyer turned the kurta inside out in daylight, the cotton and the overlocked seam belonged to a poorer world than the glowing video implied. It is paid by the twenty-two-year-old customer-support associate in a suburban tech park, swallowing sharper insults on chat queues because customers who suspect they were worked over by a machine treat the first human voice they reach without mercy. It is paid by the junior sales representative ringing two hundred form-fills who clicked an extravagant AI promise and never had the budget or the problem the software actually solves.
And now the disguise itself is running out of legal road. In experiments by researchers at NYU Stern and Emory, fully AI-generated display ads beat human-expert ads on clicks by up to nineteen per cent while unlabelled—then lost 31.5 per cent of their clicks the moment viewers were told generative AI made the ad. At the Nuremberg Institute for Market Decisions, stamping "AI-generated" onto the exact same product advertisement earlier credited to a human photographer made people rate the image less credible, less emotional, and harder to remember. Yet sneaking past the audience is a worse bet still: forty-three per cent of consumers in Cint’s 2026 survey named undisclosed AI or synthetic media as a major reason for losing trust in a brand, ahead of influencer deals and disguised sponsored content. In India, under the Advertising Standards Council of India’s labelling guidelines, issued in September 2026 and due to apply from late December, synthetic influencers, cloned voices, and fabricated product scenes that shape buyer choices are expected to wear their provenance openly on screen.
A label sewn onto the outside of a collar does not create trust on its own. It simply asks whether the garment underneath can bear being seen in full light.
Red lorries sliding through blue holiday snow on a living-room television, tyres forgetting to revolve for half a heartbeat while the old chorus swells.
When System1 tested Coca-Cola’s AI-remade festive commercial across a broad public panel, leaning on a century of accumulated red-and-white nostalgia, it pulled the top 5.9 Star score with mass audiences. Then the researchers separated out younger viewers, and the floor gave way: 2.2 Stars among Gen Z men, 2.5 among Gen Z women—cutting predicted long-term effectiveness roughly in half among the cohort most fluent in synthetic texture. Statisticians studying online experiments have a dry technical name for the risk—segment-level fragility, where an encouraging aggregate hides a reversal inside a subgroup. Anyone who has ever watched a wealthy host charm the head table while the twenty-somethings in the corner quietly put their coats on knows what it looks like without the algebra. Viewers were not offended by pixels as a matter of theology. They looked at one of the richest consumer companies on earth automating its own flagship memory, and what registered was stinginess dressed up as wonder.
A giant with a hundred years of muscle memory in every supermarket aisle can survive a season of stinginess. Most of the brands rushing into generative volume have no such inheritance in the bank. They are still asking strangers to believe they will answer the phone next month.
There is no virtue in waxing nostalgic about the old agency floor. That world wasted money with theatrical vanity, confused exhaustion with genius, and cropped whole classes and complexions out of desire unless we were useful as background colour. AI, handled without pretence—clearing away tedious background variations, testing plain functional explanations, letting a small firm without a Mumbai studio budget show how a hinge actually shuts—has real dignity in it.
What cannot be automated without hollowing out the brand is the argument before the picture is made. When a writer, a product manager, and an art director sit in a room long enough to quarrel over one adjective—when someone says, "No, don’t show that, our onboarding takes three weeks, not ten minutes," or "Give her a real kitchen counter with a stain on the grout so the bottle looks like it belongs in a life"—that friction is not creative indulgence. It is the moment an organisation decides what promise it is willing to stand behind thirty days after the credit card clears, when refunds, cancellations, and word of mouth come home to roost, and when only a chain of honest intermediate signals rather than a single opening click can tell you whether a customer actually stayed.
Downstairs by the security gate, the cigarette is out. Upstairs, Variant 19 is still winning the evening in pale mint green, gathering cheap glances by the thousand from people scrolling on night buses and half-lit balconies.
A click is only someone turning their head across a crowded room because the lighting flattered your coat. Everything that makes a brand—or quietly ruins a product—begins ten minutes later, when the music drops, the coat comes off, and your stranger finds out whether anybody is really there.
Notes and sources
- Yannick Exner, Jochen Hartmann, Oded Netzer, and Shunyuan Zhang (2025), “AI in Disguise – How AI-Generated Ads’ Visual Cues Shape Consumer Perception and Performance”, Columbia Business School, Technical University of Munich, and Harvard Business School
- Cint and Advertising Week (October 2026), “The Authenticity Audit: AI, Advertising, and the New Rules of Consumer Trust”
- Adam Peruta, Carrie Riby, Nate Cummings, and Nate Pascale (May 2026), “AI Ads Are Good Enough — And That’s the Problem”, Ipsos and Syracuse University S.I. Newhouse School of Public Communications
- MIT Initiative on the Digital Economy / Madhav Kumar and Anuj Kapoor (March 2026), “AI-Generated Video Ads Are Getting Personal. Are Consumers Buying It?”
- Meta Platforms, second-quarter 2026 earnings call, as reported by Nasdaq, “Meta Platforms Q2 earnings call highlights”, 29 July 2026.
- Brett R. Gordon, Robert Moakler, and Florian Zettelmeyer (April 2026), “Predicted Incrementality by Experimentation (PIE) for Ad Measurement”, Kellogg School of Management and Meta
- Anindya Ghose, Hyesoo Lee, Vilma Todri, and Panagiotis Adamopoulos (November 2025), “The Impact of Visual Generative AI on Advertising Effectiveness”, NYU Stern
- Nuremberg Institute for Market Decisions (December 2025), “Transparency Alone Does Not Create Trust”
- Legal 500 India (November 2026), “Advertising Law and AI Regulation in India”
- Data Security Council of India (2026), “Insight Brief on ASCI Guidelines for Responsible Labelling of Synthetically Generated Content in Advertising”
- System1 Group (December 2025), “4 Festive Themes from 2025”
- Avinash Amudala (April 2026), “PROXIMA: A Reliability Scoring Framework for Proxy Metrics in Online Controlled Experiments”, arXiv preprint
- Susan Athey, Raj Chetty, Guido W. Imbens, and Hyunseung Kang (2026), “The Surrogate Index: Combining Short-Term Proxies to Estimate Long-Term Treatment Effects More Rapidly and Precisely”, The Review of Economic Studies