Agentic Commerce Simulator

Objective

Test the main claims of When AI Becomes the Consumer in a working market model, rather than asserting them. The question is which outcomes hold across a wide range of assumptions and which depend on particular conditions such as platform incentives or mandate design.

Technical architecture

The simulator is a self-contained agent-based model written in TypeScript. It runs entirely in the browser: every slider change reruns the model, with no server and no external data.

• Engine. A seeded random number generator makes every run reproducible. Each round, 400 buyers arrive, choose a merchant or walk away, and merchants then adjust their prices.
• Aggregation. Headline figures average the last 20 of 80 rounds across five random markets. The adoption curve repeats this for eleven adoption levels.
• Interface. React components with SVG charts. The adoption curve is recomputed only when assumptions other than adoption change, which keeps the sliders responsive.

Methodology

Merchants. Twelve merchants each have a unit cost ($40 to $60), a quality score and a brand score, drawn at random. Merchant 1 is owned by the agent platform. Prices start 25 to 35 above cost.

Human buyers. A human compares only a few merchants (two by default), and better-known brands are more likely to be among them, which reflects costly search. Humans value brand strongly, perceive only part of real quality, and choose with a lot of noise.

Agent buyers. An agent compares every merchant, sees quality fully, ignores brand and chooses with little noise. Three settings change how an agent ranks options:

• Self-preferencing adds a bonus to the platform-owned merchant.
• Paid placement adds a bonus to a sponsored merchant, which pays a fee on each agent sale. The placement goes to the highest-margin merchant, since it can afford the top bid.
• A mandate price cap removes any offer above the cap.

Pricing. Each merchant keeps moving its price in the same direction while profit rises, reverses when profit falls, and cuts its price when it sells nothing.

Real value. Each buyer's surplus is measured as real value (base value plus quality) minus the price paid. Brand moves human choices but adds no real value, so persuasion that wins a sale can still leave the buyer worse off.

Model constant	Value
Base value of the product	$100
Value of top versus bottom quality	$20
Weight of brand in human choice	20
Quality perceived by humans	30% of real quality
Self-preferencing and paid-placement bonus at 100%	15
Paid-placement fee	8% of price per agent sale

Evaluation

The model was checked against results that economic theory predicts. It reproduces each of them:

1. Search costs sustain high prices. With one merchant compared per buyer, prices settle at $90 and margins at 45%, close to the monopoly-like outcome Diamond described. Each additional merchant compared lowers prices, falling to $60 when buyers compare four.
2. Informed buyers discipline prices. Average price falls steadily as more buyers delegate to agents.
3. Mixed markets sustain price dispersion. Dispersion peaks when informed and uninformed buyers coexist and falls when nearly everyone is one or the other. This is the pattern in classic models of sales with informed and uninformed consumers.
4. Self-preferencing shifts share without lowering prices. The platform's own merchant gains share while buyers receive less value.

All figures below average 10 random markets per scenario. To reproduce them, set the same values in the tool above (seed 7).

Results

1. Agents lower prices and margins, and concentrate the market.

Buyers using agents	Avg. price	Margin	Price dispersion	Concentration (HHI)	Platform share	Value per human buyer	Value per agent buyer
0%	$71.1	31.1%	9.3%	1,125	8.6%	$38.5	-
25%	$67.8	28.9%	12.3%	1,517	7.1%	$39.6	$54.5
50%	$65.1	27.0%	13.0%	2,445	10.6%	$39.5	$55.6
75%	$59.7	22.3%	14.8%	4,013	8.1%	$42.0	$58.0
100%	$49.0	9.1%	9.4%	4,826	11.3%	-	$65.9

Delegation transfers value from merchants to buyers, as the report argues. It also concentrates sales: when every buyer ranks offers the same way, the best value merchants capture most of the market. Concentration passes the threshold competition authorities treat as highly concentrated (HHI 2,500) once just over half of buyers delegate.

2. Self-preferencing moves share to the platform's own merchant. With 75% of buyers using agents:

Self-preferencing	Avg. price	Platform share	Value per agent buyer
0%	$59.7	8.1%	$58.0
50%	$59.0	12.8%	$58.2
100%	$62.9	33.9%	$52.8

A fully self-preferencing agent quadruples its owner's share. Agent buyers get about 9% less value, and average prices rise, because the favoured merchant faces less pressure to compete.

3. Paid placement had a modest effect in this model. At full weight, average prices rose by about $1 and the value each agent buyer received fell by about 2% ($58.0 to $56.7). The effect is limited because the placement goes to the highest-margin merchant, whose higher price already makes it less attractive to agents.

4. Tight mandate price caps backfire.

Mandate price cap	Avg. price	Margin	Value per agent buyer
None	$59.7	22.3%	$58.0
$80	$61.5	24.6%	$55.6
$70	$61.7	25.0%	$53.7
$65	$64.0	27.4%	$45.8

This result was not expected. A tight cap removes good-value, higher-quality offers, and merchants converge on prices just under the cap, so the cap becomes a focal point. This is consistent with empirical research on price ceilings acting as focal points for tacit collusion (Knittel and Stango, 2003). The implication for mandate design is that a protection consumers set as a ceiling can work as a coordination device for sellers.

Limitations

• Illustrative, not calibrated. The constants are chosen to be plausible rather than estimated from market data, so the direction and relative size of the effects are the findings, not the levels.
• Simple pricing rule. Merchants learn prices by trial and error. They do not anticipate rivals, collude explicitly or use the reinforcement-learning pricing studied in the algorithmic collusion literature.
• No entry or exit. Merchants that lose all their sales stay in the market.
• A single product category, with fixed quality and brand.
• A simplified platform. The agent platform has no fees, data advantages or switching costs beyond the self-preferencing and placement settings.
• Price dispersion measures posted prices, including those of merchants with few sales.

Code and resources

The full model runs in your browser on this page. A public repository will be linked here once it is published.

• Reference: Knittel, C. R., & Stango, V. (2003). Price ceilings as focal points for tacit collusion: Evidence from credit cards. American Economic Review, 93(5), 1703–1729. Source ↗
• Reference: Varian, H. R. (1980). A model of sales. American Economic Review, 70(4), 651–659.