ChatGPT, Claude, Gemini and Perplexity name different companies for the same question. Here is why.

What differs between the assistants?

Where they search, what they read, and what they remember.

Every assistant can answer in two ways. From memory, using what the model learned in training. Or with search, reading pages on the live web before it writes. Both parts differ from one assistant to the next.

Some also have a fetcher that visits a page when a user asks it to, such as ChatGPT-User and Perplexity-User. OpenAI and Perplexity both say these fetches, because a user started them, may not follow robots.txt.

Memory differs too. Each model was trained on its own data, up to its own date. A company that was written about widely before one model’s training ended may be well known to that model and barely known to another.

Being named by one assistant tells you little about the others.

Why does even one assistant give different names?

Because the answer is written fresh each time.

An assistant does not look up a stored answer. It writes a new one each time, and the search results it reads can change between runs too. Anthropic’s documentation says results are not fully deterministic even at a temperature of zero, the setting meant to make answers most repeatable.

You can see the spread within one assistant in our own data. On 17 October 2026 we asked Claude, with web search on, fifteen buyer questions in each of eight B2B categories. Each category’s answers named between 112 and 248 different companies, and no company was named in more than ten of its category’s fifteen. Related questions in one category, one assistant, one day, and still a scattered list.

So one answer from one assistant is a sample of one. To say where you stand, you need several questions, more than one assistant, and a count.

How do you check all four in one sitting?

Same questions, same day, same settings, one sheet.

About an hour for ten questions.

• Write ten buyer questions with no company names in them, in the words your buyers use.
• Open ChatGPT, Claude, Gemini and Perplexity in fresh chats with personalisation off where you can. Turn web search on where it is optional.
• Ask each question in each assistant on the same day. Paste every answer into one sheet, one row per answer.
• For each row, record the companies named, in order, and the cited pages if shown.
• Count, for each company, how many of the forty answers name it, and in how many of the four assistants.

Then sort the companies into three groups. Named by all four. Named by some. Named by one. Your competitors in the first group are the ones every buyer meets. The second group is where the work pays off fastest, because some assistant already reads about them.

If you are missing from one assistant only, look at what makes it different. A company named by ChatGPT and Perplexity but not by Gemini may be well covered on pages those two read, and weaker in Google’s index. A company named by all of them with search on but none from memory has a footprint problem, not a page problem.

What do you do with the differences?

Fix the shared foundations first, then the gaps per assistant.

Most of the work helps every assistant at once. Plain pages that answer buyer questions in their words. Review profiles that are complete and current. A consistent description of what you do across the web. Search crawlers allowed in.

Then look at the gaps. If one assistant never names you, check its search crawler in your robots.txt and your bot protection, check whether you appear in the search index it draws on, and look at the pages it cites for your questions. Each assistant’s citations are a list of the pages it trusts for your market.

Re-run the same forty answers monthly. The change in each assistant’s count is the result worth reporting. A single run, in a single assistant, is not.