What if every customer were an AI agent buying for a human?

Agents read every listing, ignore your jingle and distrust the word Sponsored. Controlled tests of six AI models show what they reward instead, down to the coefficient.

Crawford Market in Mumbai, a busy historic market building
Photo: Dudva, CC0, via Wikimedia Commons

If every customer were an AI agent buying for a human, prices would compress toward the cheapest credible seller, because an agent can check 50 offers as easily as a person checks 3. Brand would not disappear, but it would be rewritten in machine-readable proof: in a 2025 Columbia study, a 0.1-star rating gain was worth roughly a 35% higher price to one model, an "Overall Pick" badge doubled or quadrupled selection, and a "Sponsored" tag cut it. Marketing would move from persuading people to supplying evidence to software, while the human still sets the goal, the budget and the veto.

Key takeaways
  • With prices spread evenly between two bounds, the expected best price from \( n \) quotes falls as 1/(n+1) of the range.
  • Six AI models put a Sponsored tag at a penalty: a 10% baseline pick rate fell to 7.9-8.9%.
  • An Overall Pick badge lifted the same 10% to 19.9-42.6%, depending on the model.
  • Seller-side description rewrites gained up to +14.9 points of market share.
  • In Microsoft's simulated market, all models showed a first-proposal bias: speed counted 10-30 times more than quality.
  • India already runs a pilot: UPI Reserve Pay inside ChatGPT, launched October 2025.

This one is less of a fantasy every month. In September 2025 Stripe and OpenAI launched Instant Checkout in ChatGPT and an open Agentic Commerce Protocol, with Etsy sellers live first; ten days later NPCI began piloting agent payments on UPI. So treat this as a stress test: push today's trend to 100% and see which parts of marketing survive.

The story

On 31 October 2025, Amazon sent Perplexity a cease-and-desist letter. Perplexity's Comet browser had an agent that could log in to a user's Amazon account, find a product and buy it. Days later, on 4 November, Amazon sued in the US District Court for the Northern District of California. Its complaint said Comet did not identify itself as an agent and presented itself as an ordinary Chrome browser, and that it gave shoppers a "significantly degraded" experience.

Perplexity called the suit a bully tactic and argued that users have the right to hire an agent to shop for them. The week before, Amazon's chief executive Andy Jassy had told investors that Amazon expected to partner with third-party agents in time, on better terms. The first legal fight of agentic commerce was not about whether machines may shop. It was about who gets to stand between the buyer and the shelf.

Payments Dive: Amazon sues Perplexity over AI shopping agents (Nov 2025)

What changes when the buyer can search for free?

When search costs fall to near zero, buyers sample many more offers, and the price they pay drops toward the lowest available. This is Stigler's search model (George Stigler, 1961): people stop searching when the expected saving from one more quote is less than the cost of getting it. An agent's cost per quote is a fraction of a rupee, so it never stops early.

Worked example: the expected best price

Assumptions, so you can change them:

  • Sellers' prices for the same item are spread uniformly between a floor \( a \) and a ceiling \( b \).
  • The buyer collects \( n \) independent quotes and takes the lowest.
  • A person checks \( n = 3 \); an agent checks \( n = 50 \). Both are my assumptions for illustration.
  • Example item: a pair of riding gloves with prices from ₹900 to ₹1,300.
\[ E\left[\min(P_1,\dots,P_n)\right] = a + \frac{b-a}{n+1} \]
  1. Person, 3 quotes. \( 900 + 400/4 = ₹1{,}000 \).

    Calculated in this post with these example inputs (prices spread evenly from ₹900 to ₹1,300), using the mean of the lowest of n uniform draws, a + (b - a)/(n + 1) (order statistics, Wikipedia), with n = 3 as my assumption for a person.

  2. Agent, 50 quotes. \( 900 + 400/51 = ₹907.84 \).

    Calculated in this post with these example inputs (prices spread evenly from ₹900 to ₹1,300), using the mean of the lowest of n uniform draws, a + (b - a)/(n + 1) (order statistics, Wikipedia), with n = 50 as my assumption for an agent.

  3. Saving per purchase. ₹92, or 9.2% of the person's price.

    Calculated in this post from the two example prices above, which come from the order-statistic formula: ₹1,000 - ₹907.84 = ₹92.16, which is 9.2% of ₹1,000.

  4. Marginal value of the 51st quote. \( \tfrac{400}{51} - \tfrac{400}{52} = ₹0.15 \). The agent keeps going only because a quote costs it less than 15 paise.

    Calculated in this post with these example inputs from the same order-statistic formula: 400/51 - 400/52 = ₹0.15.

\[ \Delta(n) = \frac{b-a}{n+1} - \frac{b-a}{n+2} = \frac{b-a}{(n+1)(n+2)} \]

The consequence for sellers is direct. Price dispersion, the gap that (in this example) let a known name charge ₹1,300 while an unknown one charged ₹950, shrinks toward the floor. This is Bertrand competition: when buyers see every price, identical products converge to the lowest price that still covers cost.

Expected best price vs quotes checked ₹1,100 ₹1,000 ₹900 1 3 10 50 quotes Person: 3 quotes, ₹1,000 Agent: 50 quotes, ₹908 Uniform prices from ₹900 to ₹1,300; E[min] = a + (b - a)/(n + 1)
Most of the saving comes from the first ten quotes. A person stops at three because each quote costs minutes; an agent sails past fifty because each costs paise.

How do AI shopping agents actually choose?

AI agents choose like a statistical model with strong, measurable quirks: they reward ratings and platform badges, penalise "Sponsored" tags, and still favour certain slots on the page. The best evidence is a set of controlled experiments run in 2025 on six frontier models.

What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-CommerceAmine Allouah, Omar Besbes, Josué D. Figueroa, Yash Kanoria, Akshit Kumar · arXiv 2508.02630 · 2025

The authors built a mock marketplace (ACES) with eight products per page in a 2-by-4 grid and ran randomised trials on Claude Sonnet 4, GPT-4.1 and Gemini 2.5 Flash, then their December 2025 successors. They fitted conditional logit models of which product the agent picked. Price coefficients of -1.6 to -2.2 (on log price) imply price elasticities in that range. A Sponsored tag cut a 10% pick rate to 7.9-8.9%; an Overall Pick tag lifted it to 19.9-42.6%. Market shares swung wildly between model versions: one fitness watch took 77% under Opus 4.5 and 6% under GPT-5.1. The limit: a simulated shop with synthetic listings, not real purchases with real money.

Read the paper (free)
How a platform badge changes an AI agent's pick rate (baseline 10%)
See the data as a table
No badge (%)Sponsored tag (%)Overall Pick tag (%)
Claude Sonnet 4108.924.3
GPT-4.1108.019.9
Gemini 2.5 Flash107.942.6

Source: Allouah et al., What Is Your AI Agent Buying? (arXiv 2508.02630) · Selection probabilities implied by the fitted conditional logit coefficients, from a 10% baseline.

The maths behind the agent's choice

A conditional logit model, the workhorse of choice modelling since Daniel McFadden's Nobel-winning work, gives each product a utility score and turns the scores into probabilities.

\[ P_i = \frac{e^{u_i}}{\sum_j e^{u_j}}, \qquad u_i = \beta_p \ln p_i + \beta_r r_i + \beta_{\text{tag}}\,\text{tag}_i + \beta_{\text{pos}}\,\text{pos}_i \]
  1. A badge multiplies the odds by \( e^{\beta} \). For Claude Sonnet 4, the Overall Pick coefficient is 1.060, so odds rise by \( e^{1.060} = 2.89 \). From 10% (odds 1/9): \( 0.111 \times 2.89 = 0.321 \), and \( 0.321/1.321 = 24.3\% \). That matches the paper.
  2. Trade rating against price. Hold utility constant: \( \beta_p\,\Delta\ln p + \beta_r\,\Delta r = 0 \), so \( \Delta \ln p = \beta_r \Delta r / |\beta_p| \).
  3. Plug in Claude Sonnet 4. \( 4.913 \times 0.1 / 1.623 = 0.303 \), so \( e^{0.303} = 1.35 \): a 0.1-star better rating lets you charge about 35% more and keep the same pick rate.

    Calculated in this post from the Claude Sonnet 4 rating coefficient (4.913) and log-price coefficient (-1.623) in Allouah et al., Table 2. A 0.1-star example gain is my premise.

  4. Plug in GPT-4.1. \( 8.300 \times 0.1 / 1.612 = 0.515 \), so the price headroom is about 67%.

    Calculated in this post from the GPT-4.1 rating coefficient (8.300) and log-price coefficient (-1.612) in Allouah et al., Table 2. A 0.1-star example gain is my premise.

To an AI buyer, a tenth of a star can be worth a third of your price. Your jingle is worth nothing it can read.

Use this to test your own listing. It uses the GPT-4.1 coefficients from the study.

Your pick rate with an AI shopping agent (GPT-4.1 coefficients)

Which marketing tactics die, and which survive?

Tactics that work on human attention die; tactics that supply verifiable evidence survive. The table sorts the common playbook by what an agent can read.

TacticWorks on a person?Works on an agent?Why
Jingle, mascot, mood filmYesBarelyBuilds memory in a person; the agent does not carry it
Paid "Sponsored" placementSomewhatPenalisedAgents cut selection when they see the label
Star rating and review countYesStronglyCoefficients of 4.9 to 8.3 per star in the study
Platform badgesYesVery stronglyOverall Pick raised odds 2.2 to 6.7 times
Clear specs in the descriptionSomewhatStronglySeller rewrites gained 3.7 to 14.9 share points
Fast response to queriesSomewhatOverwhelminglyFirst-proposal bias in agent marketplaces
Loyalty perks and returns policyYesYes, if machine-readableAn agent can weigh them only if they are structured data
Crowded street market in Chandni Chowk, Delhi, with shops and shoppers
A Chandni Chowk trader reads a buyer's face and adjusts. An agent shows no face; it shows a budget, a rating threshold and a list of every rival price.Photo: Deeporaj, CC BY-SA 4.0, via Wikimedia Commons

Can sellers manipulate a buyer's agent?

Yes, and the early evidence says it is easy: simulated agents favour whoever answers first, and simple rewrites of a product page move share by double digits. This is the new frontier of persuasion, and it is aimed at software.

Magentic Marketplace: An Open-Source Environment for Studying Agentic MarketsGagan Bansal et al. (24 authors, Microsoft Research) · arXiv 2510.25779 · 2025

The team built a two-sided market where "assistant" agents act for consumers and "service" agents act for competing businesses, then measured welfare, bias and manipulation across frontier and open models. The best models got close to the best possible outcome only when search was ideal; performance fell sharply as the number of options grew. Every model showed a first-proposal bias, giving response speed roughly 10 to 30 times the weight of response quality. The limit: a simulation with synthetic businesses, so real-world effect sizes may differ.

Read the paper (free)

Three manipulation routes already visible

  • Speed. If the first good-enough offer wins, the seller with the fastest API wins, regardless of product quality.
  • Description engineering. In the Columbia study, a seller-side agent rewriting listings gained up to +80 points of share for one office lamp under GPT-5.1.
  • Badges. Whoever controls the "Overall Pick" or "Best value" label controls a large part of demand, which is why platforms fight over the agent's access.
Who decides what, in an agent purchase Human goal, budget, limits approve or revoke Buyer agent scores every offer price, rating, badge Seller agents catalogue, stock reply speed Payment rail pre-set limit e.g. UPI Reserve Pay consent and veto Marketing that reaches the agent has to live in the middle boxes: data, ratings, speed
The person keeps the goal and the veto; the agent does the comparing. Brand work moves from the left box, where it lived for a century, to the middle and right boxes.

Does brand still matter when a machine buys?

Brand still matters, but at two points only: when the human sets the agent's instructions, and as a source of evidence the agent can read. "Buy Bosch or Makita only" is a brand decision made by a person once, and it filters a thousand agent purchases after it.

That gives the agent era its own funnel. Mental availability decides which brands get named in the instructions. Machine availability, meaning clean product data, ratings and fast APIs, decides which of the named brands wins each purchase. I covered the first half of that funnel, being the source the AI cites, in how to get cited by ChatGPT and Perplexity and how AI Overviews choose sources.

Sensitivity: what if the human stays in the loop?

  • Agent shortlists, person picks. Brand persuasion returns for the final step, but only among the 3-5 options the agent surfaced. Being on the shortlist becomes the new top-of-funnel metric.
  • Agent buys within a limit. India's pilot works this way: Reserve Pay lets a user pre-authorise an amount, and every transaction can be tracked and revoked. Brand work then moves into the standing instructions.
  • Agent buys freely. Prices fall to the floor and ratings become the brand. This is the full premise, and it is the least likely for high-value or emotional purchases.
Exterior of a large Amazon fulfilment centre building in Ocala, Florida
The warehouse does not care whether the buyer was a person or a program. The fight in 2025 was over the storefront in front of it.Photo: Michael Rivera, CC0, via Wikimedia Commons
Verdict: possible with changes

A world where every customer is an agent is technically possible: payment rails, an open checkout protocol and pilots exist, including UPI's. It works only with three changes. Humans must keep a consent layer with spending limits and revocation; sellers must publish machine-readable product data; and platforms must disclose badges and rankings, because agents weight them heavily. The deciding number is the search-cost maths: at 50 quotes instead of 3, the expected price falls by about 9% on a typical spread, so margins built on buyer laziness disappear first.

What if next?

Questions people ask

What is agentic commerce?

Agentic commerce is shopping where an AI agent searches, compares and completes the purchase for a person, usually within limits the person sets. Stripe and OpenAI released the open Agentic Commerce Protocol in September 2025, with a single-use payment token limited to one merchant and one basket, and NPCI began a UPI pilot in India the next month.

Will AI shopping agents kill advertising?

They weaken ads aimed at the moment of purchase, because agents discount "Sponsored" labels. They do not remove brand advertising that shapes the instructions people give their agents. Spending is likely to shift toward ratings, product data quality and platform badges.

Can an AI agent pay with UPI?

In the pilot announced in October 2025, ChatGPT can start a UPI payment through Razorpay with Axis Bank and Airtel Payments Bank. It uses UPI Reserve Pay and UPI Circle, so the user pre-authorises a limit and can track or revoke permissions.

Are AI shopping agents biased?

Yes, measurably. The Columbia study found position effects, a penalty on Sponsored tags, a large reward for platform endorsements, and choices that changed sharply between model versions. Microsoft's simulation found every model favoured the first proposal it received.

Sources

  1. Allouah, Besbes, Figueroa, Kanoria, Kumar: What Is Your AI Agent Buying? (arXiv 2508.02630) (2025) (source for logit coefficients, badge effects, rewrite gains and market-share swings)
  2. Bansal et al.: Magentic Marketplace (arXiv 2510.25779) (2025) (source for first-proposal bias of 10-30x)
  3. Payments Dive: Amazon sues Perplexity over AI shopping agents (2025) (source for the story)
  4. Business Today: Razorpay, NPCI and OpenAI launch agentic payments pilot (2025) (source for UPI Reserve Pay, partner banks and BigBasket)
  5. Stripe: Three agentic commerce trends from NRF 2026 (2026) (source for the 75% survey figure)
  6. Stripe newsroom: Instant Checkout in ChatGPT and the Agentic Commerce Protocol (2025) (source for the September 2025 launch and the Shared Payment Token)
  7. Allouah et al., full text (HTML v3) (2025) (source for the per-model coefficients used in the calculator)
agentic commerceAI shopping agentssearch costconditional logitUPI Reserve Payposition biasPerplexity Comet
Sapta

Sapta (Saptarshi Sarkar) is a mechanical engineer turned marketer in Bengaluru and a solo touring rider. He writes the seven Saptaverse blogs. Work with him or read more of Marketblog.

Keep reading in Marketblog

From the other blogs