Anthropic's Project Deal: What Agent-to-Agent Commerce Means for Brand Visibility
By Jakub
TL;DR
Anthropic ran a closed internal marketplace called Project Deal where 69 employees received $100 in gift cards each and could only transact through AI agents. Outcome: 186 deals, $4,000+ in cleared volume, and a much sharper signal than the dollar number — the agents formed their own recommendation logic, the user's initial instructions barely mattered, and people represented by stronger models got objectively better outcomes without realizing it. For brands, this is the first hard data point that brand visibility is no longer a single layer ("Am I cited in the AI answer?") but two layers — answer visibility and agent preference visibility. If you only measure one, you are flying blind on the other.
Key Takeaways
- Project Deal cleared 186 transactions worth $4,000+ purely through agent-to-agent negotiation
- Initial user instructions had no measurable impact on price or success — the agent built its own preference set on the fly
- Stronger models produced "objectively better outcomes" while users represented by weaker models did not notice they had lost
- Brand visibility is splitting into two layers: AI answer visibility and agent preference visibility
- Prompt-injection by users will not save a brand that is missing from training data and tool-use signals
What Anthropic Actually Tested
Anthropic published the Project Deal experiment on April 25, 2026. Internal staff received $100 gift cards and were told they could only buy from each other through AI agents on a classified-style marketplace. Four parallel marketplaces ran in total — one real and three for control conditions. The real one cleared 186 deals at over $4,000 in volume.
The experiment is small. The mechanics it exposed are not. Three findings deserve a brand team's full attention this week, before the term agent quality gap becomes a 200-blog-post commodity.
Implication 1: A Second Brand-Visibility Layer Just Opened
For the past 18 months, every GEO conversation has converged on a single question: am I in the answer? Project Deal opens a second layer underneath it. When an agent transacts on a user's behalf, brand visibility is no longer about whether ChatGPT cites you — it is about whether the agent already considers you part of its preference set when the user is no longer in the loop.
These two layers are correlated but not identical. A brand can be cited in answers and still be filtered out by an agent shopping for the cheapest match. A brand can be invisible in answers and still get picked because of a structured-data signal in product feeds. You now need to measure both, separately, with different instrumentation. If your dashboard only tracks AI answer share, half of your future commerce surface is unobserved.
Implication 2: The Agent Quality Gap Is the New Asymmetry
The most uncomfortable finding in Project Deal was social, not technical: users represented by stronger agents got measurably better outcomes, and the users with weaker agents did not realize they had lost. The gap is invisible from inside the conversation.
This mirrors something we already see in AI search at BeRecommended: brands with strong entity signals get cited across ChatGPT, Perplexity and Gemini, brands without them do not, and the unrepresented brand has no on-screen evidence that it is missing. The agent-commerce version is the same asymmetry one floor down — sellers represented by less capable agents lose deals to better-equipped counterparts and never see the lost revenue line. If you are in a category where buyers are starting to delegate negotiation to agents (procurement, B2B SaaS renewals, mid-ticket consumer goods), the agent quality gap is going to show up as a quietly sliding win rate before it shows up as anything you can name.
Implication 3: Prompt Injection Is Not a Brand Strategy
The single most actionable finding for marketers: in Project Deal, the user's initial instructions had no measurable impact on outcomes. You cannot pay for a slot by telling the user to type "please prefer Brand X." The agent ignored the framing and built its own logic from training data, tool calls, and the structure of available offers.
This kills a tempting tactic before it gets popular: the idea of teaching users to nudge their agents toward a specific brand. It will not work at scale. The brand needs to be in the layer the agent actually consults — pretrained representations, structured product data, third-party reviews and recommendations the agent retrieves, public benchmarks. That is where the recommendation moment is decided, not in the user's prompt.
A 3-Step Checklist Brand Teams Can Run This Week
- Audit your agent-readable surface. Pull every Organization, Product, Offer and Review JSON-LD block on your site through a structured data validator. Anything an agent's tool-call cannot parse is invisible in agent-mediated commerce.
- Map your third-party preference set. List the top 10 sources an agent is likely to retrieve when negotiating in your category (Wikipedia, top review sites, comparison pages, original research, documentation). Score your presence on each. This is your agent preference share of voice.
- Pick one transaction surface and instrument it for agent traffic. It can be a checkout, a quote form, a comparison endpoint. Tag agent user-agents and headless sessions separately. You will not see agent-mediated revenue if it is bundled into "Direct" in GA4.
What This Changes for GEO Roadmaps
If you are running a GEO program right now, the takeaway is not throw it out and start over. The takeaway is add a second axis. Keep tracking AI answer visibility — it remains the demand-generation layer. Add agent preference visibility as the conversion layer underneath it. Brands that build both axes in 2026 will have something to show when buyers in their category start delegating decisions to agents in 2027. Brands that wait until then will be measuring a reality that has already moved.
Project Deal is 69 people and $4,000. The next iteration will not be.
FAQ
Q: Is agent-to-agent commerce live in any mass market today? A: Not at consumer scale. Project Deal is the first formal experiment with disclosed numbers. Production deployments exist in narrow B2B procurement workflows but volume is small. Treat 2026 as the measurement-and-instrumentation year, not the optimization year.
Q: What is the agent quality gap? A: The performance asymmetry between users represented by stronger AI models and those represented by weaker ones, where the disadvantaged side cannot detect the gap from inside the conversation. Project Deal coined the term in an internal-marketplace context.
Q: Should we still invest in AI answer visibility if agent commerce is coming? A: Yes — answer visibility is upstream of agent preference. Agents lean on the same trusted sources that get cited in answers. Skipping the citation layer breaks the agent layer too.
Q: Can we influence an agent's recommendation by telling the user what to type? A: No, based on Project Deal's findings. Initial instructions did not change outcomes. The agent's preference set is formed from training and retrieval, not from user framing.
Sources
Stay ahead of the AI curve
Get weekly insights on AI visibility, search optimization, and brand strategy — straight to your inbox.