Footfall Is Not a Receipt

The first receipts from AI advertising are in, and a paid ChatGPT visit is earning about a third of a Google search visit while the earned recommendation beside it pays best of all. Price the new channels by what rings up, not by how many people walk past.

Published
Updated
Version
v1.0
Read
6 min
By

Version

v1.0 / current

Every shopping mall has a leasing office, and every leasing office has a number it loves: footfall. Counters over the entrances tally each person who walks through, and the brochure for a vacant unit leads with the total, millions of visitors a year, a figure large enough to make almost any rent sound reasonable. The tenants who have been in the building for a while read the brochure differently. They know the counter cannot tell a shopper from a teenager cutting through to the cinema, and they judge the mall by the one number the leasing office never prints, which is what rings up at their own till on a wet Tuesday. Footfall is what the landlord sells. The receipt is what the tenant keeps.

The new AI channels are leasing space now, and their brochures lead with footfall.

The First Receipts Are In

OpenAI has the largest crowd in the building. ChatGPT counted 900 million weekly users in February, and this year it began selling ads beside its answers in a growing list of countries. For most of the year advertisers had only the size of the crowd to go on. In October they got a receipt. A working paper by researchers at WU Vienna and the Frankfurt School of Finance and Management, drawing on nearly 10 billion sessions across 1,898 online shops, measured what a paid ChatGPT visit actually brings in. In the United States it was $0.89 per session, against $2.45 for a Google search ad. Paid ChatGPT visitors converted at 0.37%, against 1.17% for search. Of the 87 advertisers buying both, 97% earned less per ChatGPT visit, and the researchers estimate that the click would need to cost about $0.26 to match the return search delivers, where one advertiser has publicly reported paying $1.72.

The same paper found something more interesting in the next aisle. The unpaid visits ChatGPT sends to a shop, the ones that arrive because the answer itself recommended the shop, kept converting after the ads arrived, and in the United States they earned more per session than any of the five channels the authors compared. The working paper is observational, it counts revenue attributed to a visit and may miss later purchases, and one of its authors advises the company that supplied the data, so it is a first receipt rather than a final audit. It is still a receipt, and it says that the recommendation is worth a great deal while the space rented beside it, so far, is worth much less.

The Oldest Tenant Starts Counting

What makes this matter now is that the oldest tenant in the building has started reading its own till. Digiday spoke with 18 media buyers and found clients questioning 10% to 30% of their Google search budgets, the part that used to renew without a meeting. Branded search clicks cost 5% more in the third quarter, according to Tinuiti, and the question being asked about them is whether those customers would have arrived anyway. "The incremental 10% to 20% is becoming less automatic," one buyer told the publication. Search took $114.2 billion of American digital advertising last year, 38.8% of the total, and it earned that share by being the place where intent showed up carrying a wallet.

Google's own numbers show the same strain from the landlord's side of the counter. Search revenue grew 12% to $63.4 billion in the third quarter, down from 17% the quarter before, as the company moved more queries into AI answers that it has not yet learned to monetize as well as the old page of links.

So the old channel is being asked to prove its increment, and the new one is being sold on its crowd. A budget committee cannot hold those two standards at once without paying for the difference.

Who Has Shown the Till

Jim Cook of Agentic Landmark did the useful sort last week. He ranked the shopping agents by published purchase results rather than by audience, and the order flipped. The agents with evidence all belong to retailers running on their own catalogs and their own customers. Amazon says its shopping assistant delivered nearly $12 billion in incremental annualized sales in 2025, and that shoppers who used it were 60% more likely to buy. Walmart says customers who use its assistant, Sparky, spend 40% more per order. Lowe's says conversion more than doubles when shoppers engage its assistant online. The general platforms, which hold the largest crowds, have published audience figures and nothing yet about what those audiences bought through them.

None of that makes reach worthless. A crowd is the precondition for a sale, and the evidence that visitors referred by AI convert well on brands' own sites is real. It means the value of these channels has so far shown up in one place, which is the earned recommendation that sends a ready buyer to a store able to answer the question, and not in the rented unit next door to it.

Lease by the Till

For a chief marketing officer that suggests an order of operations. Earn the recommendation before you rent the space beside it. What gets a company named in an answer is the same thing that gets it chosen by a retailer's assistant: product and service information specific enough for a machine to compare, prices and terms stated plainly, pages that answer the question the buyer actually asked. That work compounds in every channel at once, and it is the part of the new building that the first receipts say is paying.

Then, when the space next to the answer is offered, price it the way a tenant prices a lease. Ask for revenue per visit rather than impressions, set the bid against the click price that would match what search returns today, and hold out a share of the audience so the channel has to show what it adds rather than what it touches. Apply the same test to the search budget you already have, because the buyers in Digiday's reporting are right that a channel which cannot show its increment should not renew out of habit. A test with a ceiling is cheap. A line item that renews because the crowd is large is how money quietly leaves the building.

The leasing office will keep printing footfall, because footfall is the number it controls. The tenant's job is to keep the receipts, and to remember that a crowd walking past the window has never once paid the rent.

paid-mediasearchaimeasurementcmo
Delivered every weekday

A daily read on AI and go-to-market systems from Magnet, a national agency. Unsubscribe anytime.

The Briefing

A daily read on AI and go-to-market systems.

Ranked signals. Clear implications. Original sources.

Five to eight developments that matter to marketing leaders, with the context to act on each one.

All postsv1.0 / footfall-is-not-a-receipt