On September 10, 2026, Amazon Ads and OpenAI opened a pilot letting select US advertisers extend Amazon DSP campaigns into ChatGPT. The demand side of AI Search advertising just acquired the largest buying pipe in retail media. The AI Search ad supply chain, meaning the publishers whose pages the answer is assembled from, acquired nothing. In twenty-five years of watching this industry, I have never seen a DSP wired into a surface whose supply side does not exist as a market.

  • Amazon's ChatGPT pilot repeats its supply-chasing pattern (Netflix, Roku, Spotify, Disney), except every previous pipe led to inventory the seller owned; ChatGPT's inventory is an answer built from other people's content.
  • Demand is consolidating fast: Madison and Wall puts AI-directed spend at 12% of the US ad market in 2026, heading to 27% by 2030, and Amazon Ads runs at roughly $76bn a year.
  • Supply is volatile and unpriced: Reddit's share of ChatGPT citations fell from 3.83% to 0.52% in mid-August, and the sources that absorbed it were paid nothing for it.
  • The open web built the DSP and the SSP together. The agentic web is building only the buy side. Whoever standardizes the sell side of the AI answer sets the price.

What the AI Search ad supply chain actually is

The AI Search ad supply chain is the set of sources an AI engine retrieves, weighs and cites to compose an answer, plus the commercial layer that decides who gets paid when a brand appears in or beside that answer. In classic programmatic, the chain was explicit: publisher, SSP, exchange, DSP, advertiser. In AI Search, the first link is a publisher page and the last link is now Amazon DSP. Everything in between is a black box owned by the engine.

The programmatic supply chain vs the AI answer supply chain

The open web priced supply before it scaled demand. Ad exchanges and SSPs gave publishers a seat, a floor price and a share of every impression before DSPs aggregated buyers at scale. The order mattered: DSPs bought what SSPs had already made legible. The agentic web is running the sequence backwards. Amazon now aggregates buyers into ChatGPT while the sources feeding the answer are retrieved for free, have no exchange, and cannot set a floor. A DSP without an SSP is not a marketplace. It is a distribution channel for a single seller.

What PPC Land reported, and what we verified

PPC Land's September 11 briefing frames the deal as a change in who controls the demand path into ChatGPT, and pairs it with the IAB lifting its 2026 US ad growth forecast to 12.3% and Madison and Wall's estimate that 12% of US spend is now AI-directed. Both figures check out against the IAB's September 10 release and Digiday's September 10 report. Two corrections: PPC Land cites Amazon's ad business as passing $70bn on a trailing basis, while summing Amazon's reported quarters through Q2 2026 ($19.8bn, up 26%) gives closer to $76bn, which we use here. PPC Land also states only 16% of brands can measure AI visibility systematically; the IAB release reports instead that 48% of buyers now intend to measure brand visibility and citations inside AI tools. We use the IAB figure.

Why a bigger demand pipe does not fix the answer's economics

The distribution case: pipes won the open web, so pipes win here

The strongest argument for the deal is operational, and Digiday and Search Engine Land both made it on September 10: budgets move through the systems buyers already sit in, not through abstract intent. eMarketer's Nate Elliott noted OpenAI is still building the basics of an ad sales operation; Amazon brings one ready-made. Analyst Robert Hu adds that Amazon is likely one demand partner among several, that OpenAI keeps control of placement, and that the pilot is small, US-only, and led by a travel brand with no ASIN. On that reading this is plumbing, and plumbing is how every ad market has scaled.

Why plumbing into borrowed inventory is not a market

Plumbing worked on the open web because the seller owned what it sold. Netflix, Roku and Spotify own their audiences and their content; wiring Amazon DSP into them connected two parties who both held title to something. ChatGPT holds title to the slot beneath the answer. It does not hold title to the answer. Marketing Dive confirmed on September 10 that the ads run as labeled text and image units under the organic response, and that testers receive only aggregated performance data. So the brand buys adjacency to an answer it cannot see into, built from sources that are not party to the transaction. The supply side has not been priced. It has been skipped.

The Reddit episode shows how thin that skipped layer is. Promptwatch measured Reddit's share of ChatGPT Search citations falling from a steady 3.83% between July 18 and August 7 to 0.52% by August 14 to 17, an 86% relative drop. PPC Land reads it as the model detecting manipulation; Promptwatch itself says the data shows when, not why, and other trackers point to a change in ChatGPT's query fan-out. The cause is disputed. The consequence is not: Qwairy's data shows institutional and official brand sources absorbed the share. A supply layer that can be reweighted overnight, with no notice and no compensation, is exactly the layer no DSP can price.

What this means for brands and for publishers

For CMOs, media buyers and agencies: convenience is not evaluability

  • Treat the Amazon route as a buying convenience, not a measurement solution. Aggregated performance data cannot be deduplicated against your other DSP placements or tied to a purchase. Ask for the attribution spec before scaling.
  • Run two lines in the plan: the labeled unit beneath the answer, and presence on the sources the engine cites. The first is now bookable through a DSP. The second decides whether the answer recommends you at all.
  • Stop funding tactics that game a source pool. The IAB reports 86% of buyers will change how they measure media within twelve months because of conversational AI. Build measurement of citation presence now, before the DSP route makes spending easier than proving.

For publishers: you are the supply side that has not been packaged yet

  • The IAB's advice to sellers this week was to package audiences, not impressions. In AI Search your audience is the engine, and your package is being the cited source. Inventory that reweights weekly is exactly why a priced, stable, brand-funded placement on your pages matters.
  • Do not wait for OpenAI or Amazon to add a publisher line to this deal. Every previous supply Amazon wired in was owned media that negotiated terms. Cited content negotiated nothing.
  • Price your role now. A DSP is buying adjacency to answers your reporting builds. The only durable revenue model is one where the brand pays to appear on the pages that form the answer, and you take a share.

Three moves that decide the next 18 months

First, more pipes. OpenAI has already worked with Adform and with retail media networks at Target and Albertsons; every major DSP will be wired into an answer engine by 2027. Second, more agents on the buy side of commerce: Meta shipped Muse on September 8, an agent that completes purchases autonomously, and Originality.ai found only 26 public sites carrying Universal Commerce Protocol files out of three million checked. Machines are arriving to buy before merchants are ready to sell. Third, the measurement gap closes from the buyer's end only: Google's Meridian GeoX left beta on September 9, and John Mueller conceded on September 10 that position one to ten no longer maps onto a page with AI Overviews. Nothing in that list gives the publisher a price. That is the open slot.

Conclusion

Hold on to this: the agentic web now has a demand-side platform and still has no supply-side market, and a DSP pointed at borrowed inventory is a channel, not an economy. Smalk AI is built for the missing half: Generative Engine Advertising, native ads for AI agents placed on the publisher pages AI engines cite, giving brands presence inside the answer rather than beneath it and giving publishers the revenue share the AI Search ad supply chain has skipped. Watch which party first publishes a price for appearing on cited sources; that is the moment the sell side exists.