How ChatGPT Decides What to Say About Brands

How does ChatGPT decide what to say about brands? The short answer

How does ChatGPT decide what to say about brands? It does not read a fixed profile of your organisation and recite it back. Instead, it assembles an answer from two layers - what it absorbed during training and what it retrieves live from the web - then weighs which sources to trust before composing a reply.

That distinction matters more than it sounds. The answer a buyer, a regulator or a journalist sees is a synthesis of the public source layer, not a copy of your owned messaging. If the credible sources are thin, dated or contradictory, the answer reflects that.

So the picture an AI assistant paints of you is shaped by material you do not directly control. The good news is that the source layer can be understood, audited and strengthened.

The rest of this piece explains both layers in order - training knowledge first, then live retrieval - and then what it all means for an organisation that wants accurate AI answers.

The two layers: training (parametric) knowledge vs live retrieval

There are two distinct ways a model knows anything about you.

Layer one: parametric knowledge

During training, the model reads an enormous volume of text - reference sources, news, books and the public web - and compresses what it learns into its internal weights. This is called parametric knowledge. It is not a database you can query; it is a statistical impression of the world up to a cutoff date.

If your organisation barely appears in that training text, the model simply has little to draw on. It cannot recall what it never absorbed.

Layer two: live retrieval

When a question is current or specific, the model does something different. It runs a web lookup through a search index, reads back fresh pages, and grounds its answer in what it finds. OpenAI’s own documentation on ChatGPT search describes this live-retrieval behaviour directly.

This is why a fluent-sounding answer can still be wrong. The model predicts likely text from patterns - it is not quoting a verified record. Confidence in tone is not the same as accuracy in fact.

The balance between the two layers shifts by question type. Long-established facts lean on training knowledge. Fast-moving or niche queries tend to trigger a live search.

How retrieval-augmented generation actually works, step by step

The live-retrieval layer runs on a method called retrieval-augmented generation, or RAG. In plain terms: retrieve relevant documents first, then generate an answer grounded in them.

Here is the sequence, without the jargon:

1. Your question is turned into a meaning-based search using vector embeddings - a way of matching ideas, not just keywords.
2. Candidate pages are pulled from an index and ranked for relevance.
3. The top sources are read back into the model.
4. The answer is composed from those sources.

OpenAI’s retrieval documentation and this background on retrieval-augmented generation from Google’s research team both walk through the same retrieve-then-generate mechanic.

This is also why structure matters. Clearly written, well-organised pages are far easier for the system to parse, extract and quote than dense or ambiguous copy. If a fact about you is buried or unclear, it is less likely to surface.

So when people ask where ChatGPT gets its information about their company, the honest answer is: from whichever public pages its retrieval step ranks highly enough to read - plus whatever it already absorbed in training.

Why the source layer decides the answer

If the two layers are the mechanism, the source layer is the lever. A few principles govern which sources win.

Consensus and entity validation

Models lean toward claims that recur consistently across independent, credible sources. They also resolve who you are - your entity - from corroborating signals. When the signals agree, the model is confident and usually correct. When they conflict, it guesses.

Authority over volume

Earned media, analyst and expert coverage, and reference-source accuracy carry more weight than owned content saying the same thing. Repeating a claim on your own pages does not validate it. Independent corroboration does.

Freshness

For current questions, recently updated and well-sourced material is more likely to be retrieved and cited. Stale pages quietly lose ground.

The core thesis follows from all three: you do not edit the model. You shape the public source layer it reads from. That is the whole premise of generative engine optimisation - and it sits at the heart of modern reputation management.

How citations and brand mentions differ across ChatGPT, Gemini, Claude and Perplexity

It helps to separate two things that often get blurred.

A mention is when the model names your organisation in its response. A citation is when it links the specific source it drew on. They are triggered differently, and one can happen without the other.

The engines behave differently here. Retrieval-heavy assistants - Perplexity is the clearest example - surface explicit links alongside their answers. Others, including parts of ChatGPT, Gemini and Claude, often blend sources into prose with fewer visible citations.

This matters for measurement. Tracking what each engine says, and what it sources, is how you find an inaccuracy before it spreads from one model into others. The same wrong claim can echo across engines if it lives in the source layer they share.

Keep expectations measured. No engine “ranks” you the way classic search did. Visibility here is probabilistic, not positional - the same question can produce slightly different answers on different days.

What this means if an AI assistant is wrong about your organisation

The instinct, when a chatbot gets you wrong, is to argue with it. That does nothing. There is no settings panel inside the model, and a single corrected conversation does not change what the next person sees.

The durable fix is to correct and strengthen the underlying source layer the model retrieves from.

The practical levers are consistent:

• Reference-source accuracy - making sure the high-authority sources the model trusts are correct.
• Authoritative earned coverage - independent, credible material that corroborates the truth.
• Clear, consistent owned content - so your own pages reinforce rather than muddy the picture.
• Structured information - organised so machines can parse and extract it cleanly.

This is the same discipline whether the problem is a quiet inaccuracy or a live crisis. When the matter is high-stakes - litigation, a reputational crisis, a public figure under scrutiny - the mechanics do not change, but the response is structured and scaled to the matter. For more on the day-to-day version, see what to do when ChatGPT is wrong about your company and how machine learning shapes your reputation.

This is where Morris McLane operates as the digital execution layer: we audit what the engines actually say and source about you, then act on the source layer to correct it - the work behind our AI search visibility service.

How Morris McLane works the source layer

Knowing the mechanics is one thing; acting on them is the digital execution layer we run. We start by measuring what the engines actually say and source - tracking mentions and citations across ChatGPT, Gemini, Claude and Perplexity so we can see which claims are circulating, which sources they lean on, and where an inaccuracy is creeping in before it echoes from one model into the others.

From there the work is deliberate. We audit the reference sources the models trust and pursue corrections where they are wrong. We help earn the independent, corroborating coverage that carries weight over owned repetition. And we structure your own pages so the retrieval step can parse, extract and quote them cleanly rather than skip past dense copy.

It is patient, source-by-source work, not a settings change - but it is what shifts the retrieved and recalled picture over time. That is the discipline behind our AI search visibility service.

The short version

ChatGPT decides what to say about your brand by combining training knowledge with live retrieval, then weighing which sources are credible and consistent. It is synthesising the public source layer, not reading a profile you control. A fluent answer can still be wrong when the sources are thin or dated. You cannot edit the model - but you can correct and strengthen the sources it reads, and that is what changes the answer over time.

If an AI assistant is getting your organisation wrong, the fix lives in the source layer - start with our AI search visibility service.