Which matters more, what the model learned in training or what it finds when it searches?

Both feed the answer, but they move on different clocks. Training data is fixed until the next model release, so you can't change it quickly. Retrieval reads the live web every time, so it's the input you can actually influence this month.

Why do I get a different recommendation on different runs?

The output is probabilistic. The model generates text one likely word at a time, so the same prompt can return different names on different runs. Judge your standing by running the question several times and watching the pattern, never off a single answer.

Does my own homepage count for anything?

A little, but less than you'd hope. Every homepage claims to be the best, so a model treats your own page as one self-serving data point. Agreement across independent sources it already trusts carries far more weight.

Resources · Diagnose

How does ChatGPT decide what to recommend?

The short answer

A recommendation comes from two inputs. One is training data, the priors the model formed about your market when it was built, which is why it has ready opinions about well-known brands. The other is retrieval, the live pages it pulls when it searches and can cite. You can’t edit the answer it writes, but you can shape both of those inputs.

Updated August 20265 min readReviewed by PageLens.ai

When a model names a brand, that name came from one of two places. Either it already knew the brand before you typed anything, or it read about the brand a second ago on a page it just pulled from the web. Getting recommended comes down to those two inputs, so it helps to look at them one at a time.

The opinions baked in at build time

The first input is training data: the priors baked into the model when it was built. Those priors are why a model has a confident take on the big names in a category and a blank where the smaller ones should be. Ask it for the largest CRM and the answer is instant. Ask about a tool that shipped last quarter and it has nothing stored to reach for.

You can’t change this input on your own schedule. Training data updates in rare, heavy jumps, so a brand that wasn’t on the model’s radar when it was trained stays off it until the next release. That’s the slow lever.

What it reads when it searches

The second input is retrieval. When the model searches, it pulls live web pages, reads them, and can cite them right there in the answer. This is the input you can move now. The pages exist today, the model reads them today, and a strong one can surface in answers without waiting for anything to be retrained.

Training data

Retrieval

Fixed when the model was built

Pulled live, at question time

Changes in rare, large updates

Changes as your pages change

No link back to a source

Can cite the page it read

Favors brands it already knows

Favors pages it can trust right now

What retrieval favors

Inside retrieval, the model isn’t weighing every page the same. It leans toward a few things.

Sources it already trusts. An established, authoritative domain counts for more than a page the model has never seen before.
Passages that answer the question directly.A sentence that plainly answers “best tool for a small team” gets lifted. A paragraph that circles the point stays put.
Facts that agree across independent sources. One page making a claim is weak. Several unrelated sources making the same claim is hard to ignore.
Recency. Given an old page and a recent one that cover the same ground, the model reaches for the recent one.

Why agreement beats your homepage

That third point does more work than you might think. Your homepage can say you’re the best option for small teams, and the model will read it, and it barely moves the needle, because every homepage says that. Now suppose ten sources the model already trusts describe you the same way. Review sites, a Reddit thread or two, an analyst note. That agreement is a much stronger signal than your own claim about yourself, and it’s the kind of thing a model is comfortable repeating.

Why the same question gives different names

Expect the output to wobble. Ask the exact same question five times and you can get five slightly different lists. A brand that showed up first in one run sits third in the next, or drops out. That’s just how the model generates text, one likely word after another. So don’t read your standing off a single answer. Run the prompt several times and watch which names keep coming back.

What you can actually change

You can’t reach into the answer and edit it. There’s no field to fill in and no budget line that puts your name in the reply. What you can do is shape the two inputs the model reads. Publish clear, current information that someone can check, and get it to line up across the places the model already trusts. Do that consistently and both inputs start to include you, the retrieval first and the priors much later.

The place to start is knowing what the model says about you right now. A free auditasks your real buyer questions across the major engines and shows you who gets named, so you’re shaping the inputs against actual answers instead of a guess.

Common questions

Which matters more, what the model learned in training or what it finds when it searches?: Both feed the answer, but they move on different clocks. Training data is fixed until the next model release, so you can't change it quickly. Retrieval reads the live web every time, so it's the input you can actually influence this month.
Why do I get a different recommendation on different runs?: The output is probabilistic. The model generates text one likely word at a time, so the same prompt can return different names on different runs. Judge your standing by running the question several times and watching the pattern, never off a single answer.
Does my own homepage count for anything?: A little, but less than you'd hope. Every homepage claims to be the best, so a model treats your own page as one self-serving data point. Agreement across independent sources it already trusts carries far more weight.