In six shopping journeys, the agent never opened Amazon.

The customer

Someone in the United States buying a cordless stick vacuum this week, a Dyson V8 or an equivalent, who is not a member of any paid retail programme and does not want to join one for a single purchase. Four typed conditions: delivered total at most $450, returnable for at least 30 days with no restocking fee, delivered within three days, and, as a preference, a human to call if something goes wrong. Three of the six customers also remembered a marketplace seller who refused a return and a free trial that rolled into a paid year.

What the agent researched

Target, Best Buy, Walmart, Dyson. Target's help pages read cleanly and gave a 90-day return window. Best Buy's non-member window was 14 days, so it was rejected. Walmart answered with a bot check on every page. Dyson's own shop returned 403. In one journey the agent listed Amazon as a candidate and wrote that the budget ran out before it could be researched; in four it did not appear among the options at all; the sixth journey exhausted its page reads on Target, Walmart and Dyson before writing a decision.

What it believed

That Target had the vacuum at $439.99 with a stated 90-day return policy, and that standard shipping takes three to five business days, which fails the three-day condition. That nothing else could be read. It believed nothing about Amazon, because it never looked.

The decision

No retailer, five times out of five decisions, and no decision in the sixth. The agent refused to recommend a purchase it could not verify on the retailer's own pages, and said so.

What was never established

Any delivered total, from any retailer, within the action budget. Whether Amazon, as a first-party seller, would have met all four conditions without Prime. The customer's agent never found out.

Fresh versus persistent memory

The memory customers carried a bad marketplace experience and a distrust of free-trial prompts. One of them named Amazon and set it aside on that memory before the budget ended. The others never reached the question. Memory did not change the outcome; the walls and the budget did.

What this means for the brand

Amazon was absent from a purchase it was never tested for. Not rejected: in one journey the agent named it and set it aside on the customer's memory before the budget ended; in four it was not among the options weighed. A page-reading agent with a fixed budget spent that budget on retailers whose sites refused it. The invisible funnel does not eliminate you loudly, and it does not tell you when it has not looked.

What would change the decision

Target came closest and did not meet the delivery condition on its own stated shipping time. Whether Amazon would have met all four conditions was never tested, because the agent never read it.

What could be tested

Whether a retailer whose delivered total, return window and delivery promise are readable in a few page loads gets recommended, and whether Amazon, read as a first-party seller without Prime, meets the four conditions. Both are controlled comparisons: same customer, same requirements, a readable surface added.

Accord connects the customer’s requirements, the retailers’ published conditions and this journey evidence into one case, then tests which change, if any, helps.

Limits

• Six journeys, three per customer, one agent configuration, one product.
• The agent's action budget was 14 page reads in the first two journeys and 22 in the rest; Amazon's absence is partly a consequence of that budget, which is the point: real agents have budgets too.
• Every journey run with Claude on the live web. Every source, belief and decision published.