Measured

When an AI answers without searching the web

Some AI answers are written from memory and never touch the web. When that happens no amount of optimisation can reach them, and the engines differ enormously in how often they do it.

Last updated 11 October 2026

The short answer

An AI assistant does not always look anything up. Sometimes it answers from what the model already knows and never issues a search at all. That answer can still name brands, recommend companies and sound exactly as confident as one built from live sources — but there is no retrieval step in it, so nothing you publish can be found by it.

We record whether each run came back grounded in sources. Across the 24 scans carrying per-engine telemetry, between 2 September and 9 October 2026:

Engine	Runs measured	Answered without searching
ChatGPT	344	81 — 23.5%
Claude	936	115 — 12.3%
Gemini	1,062	3 — 0.3%
Perplexity	488	0 — 0.0%

The spread is the finding. Roughly one ChatGPT answer in four is written without consulting the web. Perplexity, in 488 runs, never did it once. These are not small differences in degree; they are different products answering the same question in different ways.

Why it decides whether optimisation can work at all

Most advice about getting cited by AI concerns what is on your page: whether a crawler can reach it, whether the content is in the HTML, whether the claim a buyer searches for appears in a form a model can lift. All of that assumes a search happens.

When it does not, the answer is drawn from training data that was frozen months earlier. A page you published last week cannot appear in it. Neither can a correction, a new product, or a price change. The only thing that moves an ungrounded answer is what the model absorbed before it shipped, which is not a channel anyone can optimise on a quarterly cycle.

So the practical reading is per engine, not general. Work aimed at Perplexity and Gemini reaches essentially every answer they give. The same work reaches around three ChatGPT answers in four. That is still most of them — but a brand told it is “invisible on ChatGPT” deserves to know that a quarter of the sample was never a contest.

How we count it

A run is grounded if the engine returned sources with its answer, and ungroundedif it answered with none. A run that failed — a timeout, a rate limit, an error from the provider — is neither, and is excluded from the denominator rather than counted as a refusal to search. An outage is not a behaviour.

That distinction is the reason this page exists as a measurement rather than an impression. The obvious way to compute it is to count the results we stored with no verdict, and it is the wrong way: a missing verdict covers “the engine did not search” and “the call failed” at the same time, and would publish our outages as ChatGPT’s temperament. The per-engine counters we record separate the two at the point of measurement.

Why this does not contradict our methodology page

How we measure says that an answer without a search and a call that fails are both recorded as no grounded answer, and neither is counted as a miss. That is correct, and it is about scoring: for the purpose of a citation rate, both mean the same thing, which is that we have no evidence either way and will not invent some. This page is not about scoring. It asks how often each engine behaves a particular way, and for that question the two are not interchangeable, so they are counted apart.

What to do with this

Read a ChatGPT result as a smaller sample than it looks. Across the runs above, roughly a quarter of its answers had no retrieval step for your content to enter.
Do not average the engines into one number and act on it.A blended “AI visibility” figure hides the fact that the engines are not running the same process.
Treat Perplexity and Gemini as the honest test of your content. They searched on effectively every run, so if your page is not showing up there, that is about the page and not about whether anyone looked.
Be wary of anyone reporting a no-search rate of zero across the board. It is not what the engines do, and measuring it requires recording groundedness per run, which most tooling does not.

Limits, stated because they bound what the table can carry. Figures come from the 24 scans that carry per-engine telemetry, not from every scan we have ever run. ChatGPT’s 344 runs are far fewer than Gemini’s 1,062, because it was added to the panel later and runs on Growth and above only; its interval is roughly plus or minus four and a half points, so “about a quarter” is the honest phrasing rather than 23.5% precisely. Perplexity additionally errored on 54 of 542 calls in this period, which is excluded from the rate above and is a separate problem we are still measuring. The brands sampled are mostly Malaysian and several are media publishers, so this is the behaviour of four engines on one market’s questions, not a universal constant.

How to check this yourself

The underlying record is one JSON column, engine_stats on geo_scans, holding grounded, ungrounded and errored run counts per engine per scan. The rate is the second divided by the sum of the first two. We describe the method rather than only the result so the figure can be checked, and rerun as the scans grow. If you measure something different, we would like to see it: hello@mediadesk.asia.