How to check whether ChatGPT and Claude recommend your client

The short answer

Write the questions your client’s buyers would ask, without the client’s name in them. Ask each one in ChatGPT, Claude and Google search, in a fresh session, more than once. For every answer, record three things: whether the client is named, whether one of its pages is linked as a source, and who is named instead. Then ask the same questions again later, the same way, and compare.

That is the whole check. The rest of this guide is how to do each step so the result holds up when you show it to the client.

Step 1: Write the questions buyers actually ask

The most common mistake is asking about the client by name. “What does Ashcombe Fabrication do?” tests whether the assistant can describe the company. It says nothing about whether the assistant would recommend it to a buyer who has never heard of it.

Write unbranded questions in the buyer’s words, not the client’s:

• Category questions. “Who are the best contract metal fabricators for medical devices in the Midwest?”
• Problem questions. “Our supplier can’t hit tolerances under 0.05 mm. Who else should we talk to?”
• Comparison questions. “What should I look for when choosing a precision fabrication partner?”

Ten to twenty questions is enough to start, spread across the segments the client sells into. Keep a separate short list of branded questions for the accuracy check in step 5.

Step 2: Ask in a clean session

Assistants personalise. Chat history, memory and custom instructions all shape the answer, and your history is full of the client. Use a temporary or logged-out chat where the assistant offers one, and turn memory off. For a client that sells locally, ask from the region its buyers are in, or name the region in the question.

Ask in more than one place. ChatGPT, Claude and Google’s AI Overviews draw on different sources, and a client can be visible in one and absent from the others.

Step 3: Record mentions and citations separately

For every answer, note:

1. Mentioned. Is the client named anywhere in the answer?
2. Cited. Is one of the client’s pages among the links or sources? Search-backed answers show them; answers from the model’s own knowledge often don’t.
3. Named instead. Which companies are recommended, and which of their pages are cited.

The two numbers mean different things. A mention shows the assistant already knows the client well enough to bring it up. A citation shows the assistant’s search found a page of the client’s that answers the question. A client can have one without the other, and the fix is different for each.

Step 4: Ask more than once

The same question can produce a different answer a minute later. One try is an anecdote. Ask each question at least twice and count: “named in 4 of 30 answers, cited in 1” is a result a client can act on, and a baseline you can beat.

Step 5: Check accuracy, not just presence

Now ask the branded questions. Does the assistant describe what the client does correctly? Does it get the customers, the locations and the main products right? Does it repeat an old positioning, a discontinued product or a competitor’s claim? Being described wrongly can cost more than being absent, because the buyer believes it.

Step 6: Repeat, and compare

Keep the questions, the assistants and the method fixed, and run the check again whenever something changes: after you fix the client’s site, publish a new page, or the market moves. The change between runs is what the client cares about, and it only means something if the method didn’t change too. Add new questions as a separate set rather than editing the old ones.

What the results usually mean

• Neither mentioned nor cited. The assistant doesn’t have a confident picture of the client. Start with the site: one plain sentence on what the client does and for whom, structured data, and pages that state the differences a competitor can’t claim. Why AI visibility matters covers the reasoning.
• Mentioned but never cited. The assistant knows the name, but nothing on the client’s site answers the question being asked. Write the page that answers it directly, in the buyer’s words.
• Cited but described wrongly. The page the assistant reads says the wrong thing, or an outdated page still ranks. Fix it at the source.
• A competitor named every time. Look at which of their pages get cited. That is the content to match or beat.

How a Haycion workspace runs the check

A client’s AI agent is the client’s public, structured profile that Claude, ChatGPT and other AI assistants read, plus the research that works for the client: their market, competitors and vendors. You build it in the client’s workspace and run its visibility check whenever you want, to see whether AI assistants recommend the client and who wins instead.

The check follows the method above. The workspace writes the buyer questions from the topics the client should own, asks them in ChatGPT, Claude and Google AI Overviews, and records for each one whether the client was mentioned, which of its pages were cited, and who was cited instead. Each run is compared with the last, so the change is a number rather than an impression. Alongside it, a 17-check site readiness audit scores the client’s site from 0 to 100, and every failed check comes with a fix, an effort estimate and a snippet to paste.

You send the result as a report under your own name. How AI visibility is measured goes through each part in detail. Public data only. Nothing from inside the client.