Does Claude search the web live when it answers?

Claude can search the web, and it does so when a question needs current information, but it will also answer from memory when it judges that enough. Because both are possible, the only safe approach is to ask which one it used. It answers that question directly and specifically.

Why this is the most important question in measurement

We have this on record, with dates. On 27 July 2026 we put 78 buyer questions to ChatGPT for clawlaw.in, our sister company, with no brand named in any of them. It confirmed afterwards that it had run no live web search for any of the 78, so the whole reading described what the model remembered. In that reading the order by breadth of mention was ProVakil on 28 questions, CLAW on 21 and Legistify on 18, which ranks what the model had absorbed rather than what was true that week.

The live half worked on a completely different timescale. On 6 August 2026 ChatGPT cited the page clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india as a source for a vendor due diligence question. That page was fourteen days old, and at baseline the same question set had named the company nowhere in that part of the market.

Neither reading was wrong. They measured different things, and what they could see was months apart. What would have been wrong is treating them as the same number, or comparing one month's memory based reading against the next month's live one and calling the difference progress.

How to force the useful version

Say plainly in your prompt that you want a live web search for each question, and that it should tell you next to each answer if it did not search. Then check the reply against that instruction rather than assuming it complied.

Ask afterwards how many searches ran in total. A specific count, and a list of retrieved pages, is what turns an impression into a measurement.

One more detail worth knowing: Anthropic publishes the crawler names it uses, including ClaudeBot for content collection, Claude-User for fetches prompted by a user's request, and Claude-SearchBot for search indexing. If any of these is blocked at your robots.txt or your firewall, the live half of this engine cannot reach you no matter how you phrase the prompt. Check your own file rather than trusting your memory of it.

What each mode tells you

Memory tells you what the wider web said about your category a while ago, which is a reasonable proxy for how established you are. It cannot show recent work. It moves on the timescale of model releases, which nobody outside the company controls.

Live search tells you whether the pages you published can be found and are worth quoting today. This is the one that responds to what you do, and it is the only half where a change this month can show up this month.

The five signs you can read without asking

Asking is better, and when you cannot, the following five signs are reliable enough.

• Citations or links in the answer. The clearest evidence that pages were read.
• Named recent specifics: a date, a version, a figure that matches your current page.
• A stated uncertainty about anything recent, which often means it knows it has not looked.
• A fact that is exactly one version out of date. A stale but once true fact is a strong sign of memory.
• A direct test: ask what a named page says today. Memory cannot answer that honestly.

A worked example of why this matters

Put the two readings above side by side on purpose. The reading of 27 July 2026 was 78 questions with no live search behind any of them, and it produced a breadth order led by ProVakil. The reading of 6 August 2026 used a page that was fourteen days old, which no memory could have contained.

If we had set those two next to each other and called the difference progress, we would have credited our own work for what was largely a change of instrument. Worse, the wrong conclusion would have been actionable. It would have told us that publishing works in a way the evidence did not support, and the next quarter's plan would have been built on it.

There is a second trap in the same place, and we have that on record too. In September 2026, on our own site aiknowsus.com, we asked Perplexity how many of the questions it had actually searched for. It withdrew its earlier statement, saying it could not honestly substantiate the claim that it had run a live search for each question, and in another batch that its claim to have searched all five was not adequately supported. That happened in three separate batches. So ask for the search count, and then check it against the pages the engine says it retrieved.

The habit that prevents this costs one line in a prompt and one column in a sheet. Ask for the search count, and write it next to every reading. Then compare like with like: searched questions against searched questions, memory answers against memory answers, and never the average of the two.

How to record it

Keep two separate numbers and never average them. Write the search count next to every reading. If you only ever keep one number, keep the live one, because it is the one you can change.

Also record the surface. A reading through an API and a reading in the app are not interchangeable, particularly around when a search happens, and a team that mixes the two will eventually report a change that is only a change of tool.

What is not knowable

Anthropic does not publish the rule that decides when a search happens, how many pages are read, or how the retrieved pages are weighed. Asking the model what it did is good evidence about that one conversation and it is not documentation of a policy.

Two further limits. Even a live search reflects what a search index can find, so a page published this morning may not be reachable yet. And no reading, in either mode, can be turned into a promise. Nobody can guarantee a position in an AI answer.

Common questions

Can I make it search every time?

You can ask, and it usually helps, and you should still verify rather than assume. Tell it to say so when it has not searched, then read the reply for that statement. If you need guaranteed live retrieval for a measurement, use an engine built entirely around retrieval as your instrument.

How often should I check?

Monthly for the full set, with the search count on every row. Memory based answers will not move in a week, so a weekly check here mostly produces the same result twice and tempts you to conclude something from noise.

Why do two runs of the same question disagree?

First suspect the search count: one searched, one did not. Then the retrieved page set, which varies. Then ordinary variation in generated text. Only after all three would you suspect that anything about your own position changed.

An answer cited a competitor's page for a fact about us. Why did it not use ours?

Either your page was not retrieved, or it does not state that fact plainly, or the crawler cannot read it. Ask which pages it retrieved to find out which. Then publish the fact where a buyer asking that question would land, in text, dated, on a page that returns its content in raw HTML.

Is being cited the same as being recommended?

No. A citation says your page was useful. A recommendation says your claims held up and you fitted the question. On this engine a vendor page is often used for facts and passed over for the recommendation, so keep the two columns apart.

If memory cannot be changed, is long term reputation work pointless?

The opposite. The memory half is built from how widely and accurately you were written about, so the slow work of appearing on pages you do not own is exactly what feeds it. It pays nothing this quarter and it is the only thing that pays in the one after next.

What to do first

Write the search count into your sheet for the readings you already have, if you can recover it, and add the question to your prompt from today. Then run your three most valuable questions and compare only the searched ones against each other. That single change makes every future reading honest, and it costs nothing.