An independent research project on AI answer accuracy. Every answer measured against zero. Makefield measures how AI answer engines describe companies: per engine, per language, per location. It never does the fix.
Ask several AI engines the same question about your product, three times each. You will rarely get the same answer twice, and the most fluent answer can be the wrong one. Makefield writes the right answer down first, keeps every transcript, and measures every answer against it.
AI visibility · in the open
What AI says about your company, measured in the open.
When a buyer asks an AI assistant about your category, an answer gets written, with or without you in it. Makefield writes here, in the open, about how to find that answer and measure it, with evidence.
How to read this page
AI often answers questions about products before anyone opens the maker's website.
I write down the right answers before I ask, keep every transcript, and score what the engines actually said. The red line is zero: the locked truth. Everything an engine says is set against it. Above the line is an overclaim, below it an omission or a false claim, and the further from the line, the larger the error.
- Invisible from the inside. If an engine tells a buyer in Munich that a product has a property it does not have, no complaint reaches the company that makes it. No internal dashboard flags it.
- Different on every engine, and it moves. A mention on one engine is not a mention on another. The answer shifts week to week and changes with the language of the question. Asking once yourself gives you an anecdote, not a measurement.
- The stakes are rising. These answers increasingly touch regulated claims, about safety, specifications and health. A false machine-made claim in a company's name is one nobody there wrote.
How a reading works · an invented example
The red line is the truth. Everything else is measured against it.
Tap an engine to read its answer. Switch the language, play the three draws, or ask once, the way you would yourself.
- Question
- Is the device waterproof?
- Locked truth
- No. It is rated IP54: protected against splashing water, not against immersion.Source of record: manufacturer specification sheet, locked before the first question
Illustrative. The product, the engines and the answers are invented to show how a reading works; they are not results. Engines are anonymised on purpose. A real reading ships with its frozen results table and its SHA-256 fingerprint.
The method · ACE, with an invented example
Audit, Correct, Engineer. Then Measure, and keep watching.
ACE is a loop: measure what the engines say, let the company change things, measure again against the noise, and keep watching. Makefield measures; it never does the fix. Step through a reading as it moves around the loop.
Illustrative. The engines and values are invented to show how the loop works; they are not results. ACE, explained step by step →
Studies · published with their limits
Ask the same question in another sitting, and the answer moves. By how much?
Before anyone can say an AI answer changed, they need to know how much it moves on its own. The floor test measured that: eighteen questions to eight answer engines, in seven capture waves between July and August 2026, each answer compared with a repeat from the same wave and with answers from other waves. In 85 of 103 question-by-engine cells, the answers moved more between waves than inside one.
The study is exploratory, and says so. It compares wording, not facts, and it cannot separate change over time from a difference between sittings.
Read the floor test → All studies → The next study needs two research partners →
What a reading scores
Seven things, scored the same way every time.
The first four are what the engines say about you. The last three are the machinery that shapes it. Each one has its own lesson in the open course.
- RecognitionDoes AI know you exist?
- AccuracyIs what AI says about you true?
- Competitor framingWhich set AI puts you in, and how often you are in it at all.
- RecommendationThe outcome you can't chase directly.
- DocumentationThe words a model actually reads.
- Structured dataMachine-true, and why schema isn't enough.
- Citation and authorityWho AI trusts to describe you.
Point at or tap a row to see what it is meant to move. Each of these links is a hypothesis, not a finding.
The obvious question
“Why not just ask ChatGPT yourself?”
You can. It will answer, confidently, once. That is the trap. One question to one engine is a single draw from a system that answers differently on the next draw. It comes with no error bar, and the more fluent it sounds, the harder it is to tell that it is wrong.
Makefield asks the engines repeatedly, with the correct answer written down and locked first, then measures each answer against the source of record. The result is not an answer. It is a scoreboard you can reproduce and defend, and it shows the thing a single prompt hides: the confident, fluent, false claim.
From the field notes
Short findings, one question at a time.
A field note is one observation, not a study: what an engine said, what that shows, and where it stops.
- AI engines don't even agree on what the question isI asked one question. Some engines heard 'recommend a tool,' others heard 'hire a compliance auditor.'
- The entity-recognition mountain every new brand starts onYou don't begin at zero. You begin sharing your name with townships, a golf brand, and a biotech start-up.
- Recommended ≠ cited: two different gamesFrom one day of questions I built two lists: who the assistants recommend, and who they cite. The name recommended most often was barely cited.
The method, in full
What a complete reading contains
Readings of named companies are not published here. The example reading shows every part a real one has: the scope, the locked ground truth, every engine's answers, the scoreboard and the fingerprint. The product and the answers are invented.