How can I tell if an AI searched the web?

Look for visible sources, links or a note that it searched. If an answer has none, it may have come from the model's memory, though not every interface shows its work.

Does an answer from memory get brands wrong?

It can be out of date or miss new brands, because it reflects what the model learned before its training ended. It can also be perfectly accurate for stable facts.

Do assistants choose when to search?

Yes, in most products the assistant decides per question, unless the user turns search on or off or the interface forces it.

Does this affect which brands get recommended?

Yes. Memory favours brands that were well established in the training material. Live search can bring in newer brands and recent coverage.

When an AI answers without searching: memory vs. live web answers

An AI assistant can answer in two ways. It can search the live web, read pages, and base its answer on what it finds. Or it can answer from memory, using only what the model learned in training. Memory answers are faster and can be fine for stable facts. For a buying question, they tend to favour brands that were already well known when the model was trained.

What is the difference, concretely?

A live web answer comes with evidence you can follow. The assistant ran searches, read results and cites some of them. It can mention a product launched last month, a price change or a new review.

A memory answer comes with nothing to follow. The assistant writes from patterns in its training. It can be fluent and confident, but it cannot know about anything that happened after its training material was gathered, and it cannot show where a claim came from.

The longer comparison is in training data or live search: where AI answers come from.

Who decides whether to search?

In most products, the assistant does, question by question. It judges whether the question needs fresh or specific information. A question like "what is the capital of France?" does not. A question like "which invoicing tool is best for a two-person consultancy?" usually does, but not always.

That decision is not constant, even for a single assistant. It can differ between two askings of the same question, and it differs between assistants. A research lab that wants to compare assistants must therefore record whether each answer used search. Otherwise it would be comparing a live answer from one assistant with a memory answer from another and calling it a fair test.

In Citewake's Edition 0, the share of answers that ran a search differed by assistant, and one assistant did not search on a substantial share of questions even though the request asked it to. The figures are in the findings guide and the editions.

Why does it change who is recommended?

Because the two routes favour different brands.

Memory favours the established. A brand that was widely written about before training gets recalled easily. A brand that launched recently may be absent.
Search favours the well covered right now. A brand appearing on current roundups, reviews and community threads can be pulled in even if the model barely knew it.
Memory is steadier, and search is more variable. A memory answer depends on fixed learning, so it changes mainly when the model changes. A search answer depends on what the searches return, which can vary.

For a young brand, this is encouraging and sobering at once. Search gives a route in. But if an assistant skips search, that route is closed for that answer.

How can you tell which one you got?

There are signals, none perfect.

Visible sources. Links, footnotes or a "sources" panel suggest search.
A mention of searching. Many interfaces show a status such as "searching the web" while working.
Recent specifics. Details like current pricing or a recent launch suggest fresh information.
Silence about sources. An answer with none may be from memory, or the interface may simply not display them.

For a researcher, the reliable method is to use an official interface that reports whether searches were run and what they returned. Where a provider reports nothing, you record "not reported" instead of guessing.

What should a researcher do about it?

Treat "searched or not" as a variable, not a footnote.

Record it for every answer. Alongside the model, version and settings.
Report it per assistant. Averages hide the difference.
Do not drop memory answers. They are real answers that real people read. Excluding them would measure a cleaner world than the one buyers live in.
Say what you asked for. If the request tells the assistant to search, say so, and report how often it complied.

Citewake's method sends one neutral instruction to search the web before answering, so that each answer is meant to be a live one, and then reports what actually happened. The methodology explains the setup.

Why is this hard to see from the outside?

Interfaces differ in how much they reveal. One may show a clear "searching" step and a list of links. Another may fold sources into small footnotes, and another may show nothing at all, even when it searched. An answer with no visible sources is therefore not proof of an answer from memory. This is why consumer testing by hand is a weak guide to which route was used. Reading the provider's own record, where it reports whether a search ran, is far more reliable. When no such record exists, the honest label is "unknown", and a careful study keeps those answers in a separate group rather than assuming either route.

What does this mean for a brand?

Both routes matter, and they need different work.

For memory, the work is long-term: being clearly and consistently described across the web, so that the next model learns who you are.
For search, the work is current: being present on the pages that searches return, and being easy to read. See how to get recommended by ChatGPT for practical steps.

You can check how an assistant treats you with the Citewake checker, then repeat the test over time, because both routes shift.