Why ChatGPT and Perplexity disagree about your brand (and what moves each one)

Two different questions, two different answers

Ask “best light roast coffee subscription?” and ChatGPT names three brands that are not you while Perplexity puts you first. Or the other way round. People read that as one of them being wrong. Neither is; they are answering two different questions.

ChatGPT's model, asked without web search, answers from memory: what it learned in training, which stopped at a cutoff. The gpt-4o-mini model that powers many integrations knows the web as it was in October 2023; Claude Haiku 4.5 reliably knows it to February 2025. Perplexity answers from a search it just ran: it fetches pages that rank for your question today, writes an answer from them, and cites them.

So if your brand is younger than the model's cutoff, or was rarely written about before it, “ChatGPT: not mentioned, Perplexity: mentioned” is the normal state of the world, not a bug in either engine. The useful question is what each one is made of, because the two are moved by different things.

What model knowledge is made of

A model recalls a brand the way a well-read person does: by having seen it often, described consistently, in many places. Your own site counts, if you let the training crawlers read it. So does everything anyone else wrote: reviews, directories, forum threads, comparison posts, news. A brand mentioned in a thousand places under one clear name gets recalled; a brand mentioned in ten, or under three different names, gets forgotten or confused with a neighbour.

Two consequences follow. First, model knowledge changes only when a new model version is trained. Nothing you publish this week reaches the model next week; it reaches the next generation, if it is read widely enough by then. Second, whether to let the training crawlers in is a real decision. You can block AI training without leaving AI answers, but the price is that future models learn about you only from what other sites say.

There is also plain randomness. The same model gives a different answer to the same question on different calls, and a brand can be named in two of three tries. That is why the Monitor asks ChatGPT and Claude three times per prompt on Pro and shows the range, not one answer.

What live search is made of

A search-grounded answer is built in three steps: the engine searches for your question, fetches the pages that rank, and writes an answer from them with citations. Your visibility in that answer comes down to whether pages that mention you rank for the question and can be fetched by the engine's crawler: PerplexityBot and Perplexity-User for Perplexity, OAI-SearchBot for ChatGPT search, Googlebot for AI Overviews.

This side moves in days, and it moves for the reasons ordinary search does. Publish a page that actually answers the question. Keep it retrievable: server-rendered text, structured data, the search crawlers allowed. Earn mentions on the pages that already rank, because an engine that finds you on three of its five sources will name you. Live search still varies from run to run as the index changes, but the variation is smaller and, when it moves, you can usually see why.

The ChatGPT people actually use is both

One more source of confusion. The ChatGPT app searches the web when it decides a question needs it, so a user there can see your brand through OAI-SearchBot's index even when the model alone has never heard of you. Our “ChatGPT” column deliberately asks the model through the API with no search, because that measures what the model knows. Perplexity and Google AI Overviews measure the search side. Read the two kinds of column together and they tell you which lever to pull; read either alone and you will pull the wrong one.

Reading a disagreement

• Model no, search yes. You are new, or not yet written about much, and the search side is fine. The work is off your own site: get described, consistently, in the places that rank, and decide whether the next model generation should read your pages. Then wait for that generation; nothing else brings the model round.
• Model yes, search no. An established brand whose current pages do not rank or cannot be fetched for this question. Check robots.txt for the search crawlers, check whether the page needs JavaScript to show its text, then publish something that answers the question directly. This is the one you can fix this month.
• Both no. Either nobody asks the question that way, or the category is owned by names that appear in every answer. Look at which competitors the answers name and whether your prompts are the ones customers actually type.
• Both yes, but third. You are in the answer after two others. The engines are describing what “best” means to them; the pages they read decide the order. Make the reason to choose you legible on the pages that get fetched.

How to measure it without fooling yourself

1. Same prompts, same engines, every week. A one-off comparison tells you about one day.
2. Label every result by how it was measured: live search or model knowledge. Averaging the two hides the lever.
3. Ask the model-knowledge engines several times and look at the range. A move inside the range is noise; a move that persists across runs is drift.
4. Count the competitors named in the same answers. Share of voice is relative: a flat rate while a rival climbs is a loss.

This is what the Monitor does, with every result labelled live search or model knowledge and the method written up on the methodology page. The free tier asks ChatGPT and Claude your three prompts, four runs a month; Pro adds Perplexity and Google AI Overviews, repeat asks with the range, and runs it for you weekly or daily. Either way, start with the two columns side by side. The disagreement between them is the most useful number on the page.