What ChatGPT tells your customers
when you are not in the room.

Your own ChatGPT knows who you are. It has your history, your account, your context. So when you ask it about your business, it flatters you. A stranger's does not.

Model Mirror asks the questions your customers actually ask, from a clean position, across four assistants with live search on. Then it reports what came back.

Three things, measured separately

Most tools report one number: were you mentioned. That overstates how well you are doing, because being listed among competitors is not the same as being sent a customer.

The gap between the first and the second is usually the finding.

And the part you can act on

Assistants ground their answers in pages they can retrieve. They recommend whoever appears in those pages, which is normally a directory, a review site, or a listicle rather than anybody's own homepage. So the report names the specific domains cited for your category. That is where a fix has to land. Rewriting your own site usually changes nothing.

If your name is shared, the numbers lie

We learned this on ourselves. An early run scored AUTX at 33% visibility. The assistants were describing a defunct stock ticker, an autotransformer product line, and a fibre material that happen to share the name. Real visibility was zero.

Model Mirror requires corroborating evidence before counting a mention, and reports name collisions separately instead of folding them into a flattering total. Any tool that counts raw name matches will tell you that you are doing better than you are.

Run it inside your AI assistant

Model Mirror ships as an MCP server, so Claude and other assistants that support MCP can run the check during a conversation rather than sending you to a website.

"autx-model-mirror": {
  "command": "python3",
  "args": ["-m", "apps.mirror.mcp_server"],
  "env": { "ANTHROPIC_API_KEY": "sk-ant-..." }
}

Bring your own key. It spends against your environment, never ours.

What an audit includes

• 30 to 50 buying-intent questions across four assistants with live search.
• The three-tier breakdown, per assistant, so a single weak surface is visible.
• Who gets recommended instead, ranked by how often.
• The cited source domains, which is the prioritised fix list.
• The questions where you never came up, each one a customer with intent.
• A logged-out screenshot as corroborating evidence.

What it does not include

Any promise that a fix produces a ranking change by a given date. Grounded answers vary between runs. The report states what was observed and what the citation evidence supports, and a repeat scan is what measures movement. Anyone promising otherwise is guessing.

Request an audit

Email sales@autx.ai with your business name, what you sell, and your city if you serve one. We reply to every message.