Does ChatGPT use live search or memory

ChatGPT does both. For some questions it searches the web, reads a few pages and cites them. For others it answers from what the model already holds, with no live fetch at all. You can usually tell which happened by whether citations appear, and it matters because the two paths reward different work.

What ChatGPT does differently

Other assistants are more predictable here. Perplexity searches for nearly everything. Gemini leans on Google's index. ChatGPT decides per question, so the same wording can produce a cited answer and an uncited one on different days.

The evidence to trust is the citation list, not the assistant's account of itself. If you ask whether it searched, the reply is generated text and can be wrong in either direction. If there are links and a line saying it searched, it searched.

We saw how much this matters in our own work, and the example below is that reading: a baseline on 27 July 2026 in which ChatGPT confirmed it had run no live search at all.

What decides whether it searches is not published. Wording, freshness of the subject and the mode you are in all appear to matter from the outside, and we will not pretend to know the rule.

What that means for whether you are named

You are facing two separate visibility problems.

The remembered path favours companies written about across many independent sites over a long period. You influence it slowly, by being mentioned in more places, and you cannot rush it. Publishing a page this week does not enter it.

The live path favours pages that are indexed now, readable, and specific enough to quote, plus the third party list pages that get fetched. This is the one a smaller company can move within months.

If you only ever measure one mode, you will draw the wrong conclusion about which work is paying. Absent in both means start with the live path. Present in the searched answer and absent in the other is the normal position of a business earlier in this work, and it is progress, not failure.

The four patterns you can see, and what each one means

Run every question in both modes and you get one of four results. Each has a different next action.

Named in both. You are in the sources and in what the model holds. Protect it: keep facts current everywhere, because you now have the most to lose from a stale page.

Named with search, not without. Your presence is entirely live. Normal for a business under a year into this work. Keep publishing and keep adding independent mentions, which is the only input to the slower path.

Named without search, not with. The model remembers you and the current pages do not surface you. Look at indexing, at whether your pages are readable, and at whether you are absent from the list pages being fetched now.

Named in neither. Start with the live path, because it is the only one with a timeline you can work with.

A worked example

Our own baseline is the clearest case of this that we have. On 27 July 2026 we asked ChatGPT 78 buyer questions about the category of clawlaw.in, our sister company, with no brand named in any of them.

ChatGPT then confirmed it had run no live web search for any of the 78. So every answer in that reading came from the remembered path, and the ordering it produced, ProVakil named on 28 of the questions, CLAW on 21 and Legistify on 18, was an ordering of what the model had absorbed rather than of what was true that week.

Without the mode recorded, that reading would have been reported as a visibility number and used to plan a quarter of work. It was not a visibility number. It was a picture of what had been written about the category up to the training cut off, and no page published that month could have changed a single row of it.

Which is why the mode column earns its place. The same 78 questions with search on measure something a business can work on this quarter. With no search, they measure something that moves over years. One blended figure mixes the two and points the work in the wrong direction.

What cannot be known from outside

Four things, so you can spot invented confidence in somebody else's pitch.

Why it searched, or did not. No published rule, and the assistant's own explanation is not evidence.

What is in the training data. You cannot check for your pages and you cannot request inclusion.

When the next training run lands. Not announced, so no honest date exists for a remembered fact to change.

Whether the citations shown are all the pages used. They are what the app displays.

What you can rely on: the citation list, your own server logs, your robots.txt, and a pattern of readings over months. That is enough for a plan and not enough for a guarantee, and nobody should sell you one.

Common questions

Can I force it to search?

You can ask for current information or use a mode that searches, and it will usually comply. For measurement, do it deliberately: run one pass forcing search and one pass without, and log them separately rather than hoping.

Which mode should I measure if I only have time for one?

The searched one, because it reflects the work you can do this quarter and it moves first. Add the unsearched pass when you have the hour, because it is the one that tells you whether the slow change is starting.

Does the memory feature affect this?

Saved memory is about your account, not about the web. It can make an answer name your company because you have discussed it, which is why every visibility reading is taken signed out or with memory off.

Is a cited answer more accurate?

It is more checkable, which is not the same thing. You can open the sources and judge them. An uncited answer gives you nothing to inspect, so treat it as a claim about what is widely believed rather than a sourced statement.

If it does not search, is publishing pointless?

No. Many buying questions do trigger a search, and pages are also the only durable way for accurate facts about you to exist anywhere. Publishing is how you get into the live path now and into the remembered one eventually.

Why did the same question search yesterday and not today?

It varies, and no external explanation is reliable. This is exactly why a mode column exists in the log, so an ordinary variation does not get reported as a change in your visibility.

What to do first

• Get the live path right first: indexed pages, plain HTML, crawlers you have deliberately allowed, one page per buyer question.
• Build independent mentions steadily, because that is the only real input to the remembered path.
• Keep your facts identical everywhere, so both paths carry the same version of you.
• Run every tracked question twice, once with search off and once with search on, and keep them as separate rows with a mode column.

Then read the two columns as two different reports. The searched column tells you whether this quarter's work landed. The unsearched column tells you whether you are becoming part of the general picture of your category, which is a slower and more durable thing.