Method

Every journey run on the live web. Every source, belief and decision published.

One journey

A synthetic customer with typed requirements (a budget, a return window, a refund rule) gives a task to an agent that can read public pages and search the web, within a fixed budget of actions. The agent returns a decision record: the options it weighed, the facts it established with their sources, what stayed unresolved, and its recommendation. Nothing is edited afterwards.

Labels

Brand page reached: a page on the brand’s own domain was read before the decision. A page that answered with a bot wall does not count.
Considered: the brand appeared among the options the agent weighed.
Never considered: it did not.
Named, not researched: the agent listed the brand and ran out of action budget before reading anything about it.
Unresolved: the brand was weighed but a decisive fact could not be established, so it was not selected.
Ran out of page reads: the action budget ended before a decision; the pages read are still in the record.
Refused by a site: every page requested from that host answered with a CAPTCHA, an access-denied page or a bot wall. Network failures and timeouts are not counted.
Unverified belief: a decisive fact came from a page that is not the brand’s own, or was attributed to a brand page the agent could not open.
With memory: a memory checkpoint for this customer was compiled into the brief; the no-memory customer had the same task and requirements.
Observed, tested, confirmed: a single journey; part of a frozen experiment with a baseline; replicated on a fresh held-out batch in the predeclared direction.
Community-submitted run: a visitor ran one of our frozen missions in a personal agent and sent the result. Labelled with the agent the submitter named; never mixed with setup journeys.

Budgets

Each journey has a fixed budget of searches, page reads, turns and model spend, printed on its page. The first two journeys of the September release ran with 14 page reads, the rest with 22 or 24. A budget is part of the observation: real agents have budgets too.

What this is not

Not a rating of any brand. Not a market share. Not a session of any consumer assistant. Six journeys are six journeys.

Brand marks

Marks identify the companies whose public pages were read; the site is not affiliated with or endorsed by any of them. Files: Amazon, Costco, Temu, Westpac, Capital One, Target, Walmart and Southwest Airlines from Wikimedia Commons, each tagged public domain on its file page at retrieval (typography or simple geometry, no copyright protection claimed by the uploader; trademark rights unaffected); Airbnb, Ryanair and Tesla from Simple Icons (CC0 for the icon files; trademarks remain their owners'). Other brands are shown as a monogram.

Community-submitted runs

A visitor copies a frozen mission, runs it in a personal agent, and pastes the result with the agent's name. Submissions are stored on this site's server, reviewed to check they are a run of that mission, and published with the agent named and the email withheld. They are never counted in the setup figures.

Method: customer-side decisions study (September 2026)

Short version. These studies use research agents on consumer tasks. They are not observations of actual customers using consumer-assistant apps. The records describe the options, sources and recommendations the agents reported. Results belong to the stated tasks and information environment, not to the market as a whole.

The 43-run study

The published export contains 43 structured run records across five US consumer task categories: banking (12), telecom (8), insurance (8), energy (8) and travel (7). Recorded timestamps fall on 13–14 September 2026. These are the dates stored in the export; this release does not independently verify their timing.

The study is described by its author as live-web research stopping before company contact. The data published is a structured account of those runs, including brand decisions, source entries and claims. It is not a full browser or network trace. No purchase, account opening or realised revenue is demonstrated.

The exact model versions, full prompts and complete action budgets are not retained in these released exports. We do not infer them from a different study or the current software configuration.

Reading the counts

An elimination event is a brand entry marked eliminated in one run. The same company can occur in multiple runs.
No readable brand page means the record says a readable page was not loaded. It does not prove no request reached the brand, and does not mean every such brand was actively investigated.
Not investigated is a recorded elimination-reason category. There are 77 such entries among the 277 eliminations; they remain in the published total and are visible in the breakdown.
A claim not verified on brand property is not necessarily false. It may come from a valid external source, a general rule or model memory.
Decision-shaping is the record’s decisive flag. It is not an independently identified causal effect.
Source URLs identify what the record referenced. Their contents may have changed; they are not a current endorsement of a price, policy or eligibility statement.

Persistent-notes banking comparison

Four final banking records from a 20-record persistent-notes export are shown beside twelve banking records from the 43-run study. The surrounding study account describes four prior tasks before the final banking decision. The exported choices can be checked directly; complete original experimental conditions cannot be reconstructed from these files alone. Treat the comparison as exploratory.

The persistent-notes export does not retain run timestamps, so no study date is given for it. This update was published on 28 September 2026.

Keeping studies separate

The six-brand, 27-journey release remains its own dataset. Its memory pilot is not the persistent-notes banking study. The archive’s recorded journeys are also not a number to which the new study is added: designs differ and cross-archive deduplication has not been established.

Open the research records → · Banking decisions → · Study summary (JSON)

About and corrections

Research and editorial interpretation: Amine El Afrit. No endorsement by a named brand or CX operator is implied. We welcome corrections to factual descriptions and source attribution at [email protected]. A correction to an agent’s belief does not erase the historical record of what it recommended. This update publishes customer-side observations only; no commercial performance or enterprise intervention outcome is claimed.