---
title: Measured honestly, or not at all.
canonical: https://chosenbyagent.com/method
updated: 2026-10-22
audience: both
offers: []
description: How ChosenByAgent runs assistants and agents, stamps every reading, calibrates by hand and what it never claims.
---

# Measured honestly, or not at all.

A reading is only useful if you can repeat it. So every one of ours can be.

## Run

We turn real buyer intents into scenarios in your markets and languages and run them through ChatGPT, Gemini and Claude via their official APIs with web search, or through agent harnesses with official access. Single answers vary a lot, so every cell is repeated — 15 times for routing audits, around 325 for choice tests.

## Stamp

Every result carries the model, the harness version and the date. When a model changes, comparisons across the change are flagged instead of being mixed into one trend.

`RUN · gpt-6.1-sol · harness 0.4 · 2026-10-21`

## Map

We record where agents and assistants fail: sold listings shown as live, wrong prices, lost attribution, pages that only render with JavaScript, bot management that blocks agents. API answers and consumer apps do not always agree, so a sample is checked by hand in the consumer apps and, where you share them, against your own referral logs.

## Fix

Each finding comes with the change that would fix it, who owns it, and a re-run after the next model release to show whether it held.

## What we don’t claim

- We do not forecast what any particular consumer agent will buy.
- We do not report “agent conversion”: simulated agents almost never refuse to buy.
- Our shares are a versioned index of model behaviour, not market share.
- We do not automate consumer assistant apps whose terms forbid it, and we never place real orders.