Fashion and beauty brands are replacing an advertising budget they can switch off with rent they cannot. The trade was laid out at a brand leaders’ dinner hosted by Glossy and Modern Retail, where founders said AI-mediated product discovery had made search and display advertising too unreliable to plan around, and that they were moving spend into stores, events and staffed service instead. That is a fixed-cost answer to a variable-cost problem. The two behave very differently when demand disappoints.

The evidence for the swap is thinner than the swap itself. The dinner ran under Chatham House rules, so the founders are unnamed and the only figure produced in the room was one brand’s estimate that “about 40% of our business is really done with wholesalers.” One founder said the company had relied on “search engine [ads], display” and now finds “it’s very difficult to advertise on those channels,” and so was “starting to rely on in-person and personal connections again.” Another had “just started throwing parties because we have a big space.” A dinner is not a dataset, and unnamed founders describing their own budgets rank low on any evidence scale. The mechanism they are describing, though, is now measurable elsewhere.

The advertising arithmetic broke because the surface replacing it has no auction. A September paper by Benjamin Tannenbaum measured 34,960 unbranded prompt runs across GPT and Gemini between June and September 2026, and found that when neither the brand nor its own domain appeared in the retrieval path the engine actually used, the engine named that brand in under 4% of runs. Raising a bid does not put a domain into a retrieval path; nothing in that condition carries a rate card. Where the brand’s own domain was cited, mention rates reached 49.0% on GPT and 58.4% on Gemini, the same finding restated as an opportunity. The data comes from 75 anonymised projects run through a commercial AI-visibility tool rather than a fashion sample, and the paper has not been peer reviewed.

The effect persists across runs, which is what turns a bad quarter into a cost-structure problem. A brand that went unmentioned once, with no domain exposure in the following run, was named again under 2% of the time on both engines. Paid media does not behave that way: a paused campaign resumes on the day the budget resumes. Absence from a retrieval path is not a spend decision and cannot be reversed by one, which is exactly why it cannot be forecast. CPMs can be modelled a quarter ahead. There is no equivalent instrument for a retrieval path you sit outside of.

The click loss underneath all this was measured before the engines got good. Pew Research Center tracked around 900 US adults across 68,879 Google queries in March 2025 and found a traditional search result was clicked on 8% of visits where an AI summary appeared, against 15% of visits where none did. That was general search behaviour among US adults, not fashion shopping, and it is eighteen months old. Whether it has worsened for any particular retailer is unprovable, because the meter that would settle it does not report the number: Google’s generative performance report gives site owners impressions inside AI Overviews and AI Mode and withholds clicks, as we covered in August. A brand cannot price a channel whose only public gauge counts the times it was seen.

Faced with a number they could not forecast, brands went out and bought one they could.

Rent forecasts beautifully, which is its appeal and its trap. CBRE put average US retail asking rent at $24.79 per square foot in the second quarter, up 2.4% year over year. The rise is not the problem; the scarcity behind it is. Availability held at 4.9% on construction completions CBRE calls historically low, and at that level the landlord writes the term and the tenant signs it. Brands are taking multi-year fixed obligations in the tightest retail leasing market in years, at the moment their own demand forecast is least reliable. A campaign can be switched off before lunch.

That data has a wrinkle the framing above skips. The same CBRE release credits net absorption leadership in Boston and Los Angeles to “aggressive expansion by grocery and apparel retailers,” which describes brands taking space because business is good, not because a channel failed them. Fear-driven leasing and opportunity-driven leasing produce the same signature on a lease; only the balance sheet behind it tells you which one you’re looking at.

The strongest case against this argument is that paid acquisition is not broken, and the sell side has the receipts. Meta’s second-quarter results show ad impressions up 14% year over year. That is more inventory sold, not spillover into cheaper unsold slots. The average price per ad rose 12% over the same period. A channel in structural decline does not raise prices into rising volume.

The buy-side data says the same thing. Common Thread Collective’s index of 200 to 300 direct-to-consumer stores found Meta spend up 25.28% in the first quarter of 2026. Brands did not pull back; they spent a quarter more into the same channel. Return on that spend slipped only 3%, which the index’s authors read as proof that “the platforms are more elastic” than a year ago. Stated at full strength, the counter-argument runs like this: if Meta still absorbs a quarter more spend for a rounding error in efficiency, then a brand signing a five-year lease is insuring against a risk that has not arrived, at a premium it has not calculated.

Two features of that index answer it. Inclusion requires two unbroken years of Meta and Google data, which admits the brands that kept paying and excludes, by construction, the ones that stopped. The index also attributes its 13.6% revenue growth primarily to returning customers, which is retention rather than acquisition. ROAS is a ratio between the spend a brand chose to make and the return it got, and it is silent on demand that never entered the auction to be bid on. Elasticity in a channel and health in a funnel are different measurements, and only one of them was taken.

The most credible advocate for the staffed store is also its cautionary tale. Ron Johnson, who built Apple’s retail operation, told TechCrunch this month that “the secret sauce for Apple has always been its people” and that AI “will never be able to have you physically experience a product.” He is right about the product and about the people. He also founded Enjoy Technology, an e-commerce business that brought technology products and setup services directly into customers’ homes, and which filed for bankruptcy in 2022 and sold its assets to Asurion. Human service is a genuine advantage and a fixed cost at the same time, and the second fact does not wait politely for the first to fail.

None of this makes the store the wrong purchase. It makes the reason for the purchase the thing worth auditing. A store that earns its rent is a business decision, and it survives a bad discovery quarter. A store bought because the CPM chart became frightening is an insurance policy with a fixed payout and a premium that renews on the first of every month. If AI-mediated discovery keeps routing demand around brands that sit outside the retrieval path, no lease puts them back inside it; the lease changes only the form in which the shortfall arrives, from a campaign a marketer paused to rent a CFO owes. The brands that come through this will be the ones that can say what the store earns without mentioning the advertising it replaced.