The answer shows no web search: how to report what evidence remains

When an AI answer shows no web search, there are three different situations behind that one blank, and you should never report them the same way. The engine may really have answered without issuing any searches this time. The engine may have searched but not told the monitoring tool what it searched for. Or your own records may simply not cover that answer. Only the first is a claim about the engine, and even it does not prove where the answer's content came from. This guide gives you a way to sort a blank into the right case and wording you can paste into a report.

In short

• "No search shown" is not one state. Separate an observed no-search, an engine that does not report searches, and a gap in your own records.
• Sources are stronger evidence than queries. A returned source shows a search happened even when the query text was never captured.
• A no-search reading never proves the answer came from training data, memory or any specific origin. Report the observation, not the cause.
• A "did not search" cell in a rate can be an artefact of what an engine reports, so check the engine before comparing rates.
• Use the reporting phrases and the decision tree below so the same blank is described the same way every week.

Why does an AI answer show no web search?

An answer shows no web search when nothing on record shows a search, and that has several possible causes: the engine did not search, the engine searched but did not report it, or the capture never ran for that answer.

Some AI engines decide for themselves whether to search. DiscoveredBy never instructs ChatGPT (app) or Gemini (app) to search, so some of their answers cite no sources at all, and that is normal behaviour rather than a fault (see Engines and measurement). Other engines never expose their searches, so a blank there is about reporting, not behaviour. A third group is about your data: capture may not have run, or the answer was collected before a capture feature existed.

This matters because the words you choose in a report become the reader's belief. "The engine did not search" and "we have no record of a search" lead to very different decisions about what to change.

What is fan-out, and what does "not recorded" mean?

Fan-out is an engine expanding one prompt into several searches of its own before it writes an answer. It is detail about a single answer, not a separate answer.

A monitoring tool can capture that expansion only where the engine exposes it. For each answer, DiscoveredBy works out a fan-out state fresh from the queries and sources on hand, rather than storing a status that could drift out of sync (Fan-out). There are seven states. Three are easy: exact, grouped and queries only all mean queries were recorded, at decreasing levels of precision about which query returned which source. The other four are where blanks live: no search, not applicable, searches not recorded and not recorded. The next sections take them in turn.

Enlarge image

Case 1: did the engine really answer without searching?

If capture ran and found no sub-queries and no sources, the state is no search, and it is the one blank that is a claim about the engine: it issued no sub-queries this time.

The docs are careful about this. no_search is a real statement about the engine's behaviour on that run, which is why it is kept apart from the states where your records are thin. Two limits still apply.

First, "this time" is the scope. Engines that choose whether to search may search on the next run of the same prompt. One observation is not a property of the prompt.

Second, no search does not tell you why the answer said what it said. It does not show the content came from training data, and it does not show which sources shaped it. Say "on this run the engine reported no searches and returned no sources" and stop there.

Case 2: did the engine search but not report it?

Some surfaces never report their searches, so an empty fan-out panel is a statement about the surface, not about the answer.

DiscoveredBy marks fan-out as not applicable for Google AI Overviews, Google AI Mode and Gemini (app), which report only the sources their answers cite, and for Perplexity answers collected through its earlier Search API, which return a ranked result list rather than a generated answer (Fan-out). This state is checked before every other, so such an answer is never misread as an engine that searched and told us nothing.

Perplexity's current Sonar collection is a different case. It returns the sources it found but not the text of its searches, so an answer with any source reads searches not recorded. The docs say this is a gap in what the engine reports, not a sign that it answered without searching. It sits under case 3 for that reason: sources exist, queries do not.

If you need to inspect what a Perplexity answer drew on, our sibling guide on inspecting a Perplexity answer's sources covers that workflow.

Case 3: did the tool simply not write it down?

If sources exist but no queries were recorded, or capture never ran, the gap is in the record. Two states cover this, and the docs deliberately keep them apart.

• Searches not recorded: at least one retrieved source exists for the run, but zero queries. We cannot say the engine did not search, only that we did not write down a search.
• Not recorded: fan-out capture did not run for this execution at all. There are no queries and no sources to show, and the page has nothing to tell you about what the engine searched for. For example, ChatGPT (app) answers collected before 2026-09-29 read this way, because its searches were not captured before then.

Searches not recorded also reads almost the same as no search on the page, yet the two mean different things on purpose. No search is a fact about what the engine did; searches not recorded is a fact about what was written down. Collapsing them would make a claim about the engine that your records cannot support.

Does a source prove a search happened?

For counting web search, yes: a returned source counts as evidence of a search even when the query text is missing.

In the answer-features table, exact, grouped, queries only and searches not recorded all count as searched, "because a returned source proves a search even without its query" (Fan-out). No search counts as did not search. Not recorded reads Not recorded and is left out of the rates (with an exception for older ChatGPT (app) answers, covered below). Not applicable is blank, with Gemini (app) as the exception, since its web search is read from whether the answer showed a source.

That last rule has a consequence you should hold onto. A rate of "did not search" is only as good as what that engine reports, which is the subject of the next section.

When is "did not search" really "nothing reported"?

Some engines can produce a "did not search" reading for an answer that may have searched, because the reading is derived from sources and queries rather than from the engine's own statement.

The answer-features docs describe how a few engines are read:

• Gemini (app), and ChatGPT (app) answers collected before 2026-09-29: an answer that searched but showed no source reads as did not search. This holds even though the fan-out panel reads not recorded for those older ChatGPT (app) answers, because the web search column is read from sources there.
• Perplexity (Sonar): an answer with a search count but no returned source reads as did not search.
• Grok: a search of X counts as a search.
• Google AI Overviews and AI Mode: the web search cell is blank, because Google's AI answers are part of Google Search (Answer features).

So before you compare "web search" rates across engines, read the table of which engines report what. A blank is never shown as 0% for a feature the engine does not report. A 0% means answers were checked and none held the feature, which is a different thing, and only as strong as the engine's reporting behind it.

The rate labels matter too. Blank means the engine does not report the feature. Not recorded means the answer was checked but searches were not captured (applies to the API engines for web search). Not checked means the answer has not been checked yet. 0% means checked and none held it.

What evidence remains when no search is shown?

Even with no visible search, you still have the answer text, its cited sources if any, the mentions in it, and the collection channel that produced it. Report those.

• The answer text and mentions. Whether your brand is named, in what role and with which stated reasons, is observable regardless of search state. These are observations of an answer, not proof of why the model produced it.
• Cited sources. A cited page is always counted as found for ChatGPT (app). On an engine that reports web search, an answer that cites sources does not read as "did not search" in the web search column, even if the query text is missing.
• Retrieved versus cited. Where an engine reports the pages its search returned, Retrieved vs cited compares those with what the answer credits. Channels differ: Google AI Overviews, Google AI Mode, Gemini (app) and ChatGPT (app) read as Cited sources only and are not compared. Rates are computed for Claude, Perplexity and Grok. For a related workflow, see Your page was retrieved but not cited: what to check next.
• Repetition. Because an engine may search on one run and not the next, the same prompt across days is the honest way to see a pattern. See Prompts.

Enlarge image

The deliverable: reporting language for each case

Use one sentence per case. These phrases stay inside what each state supports.

Add one line to every report that uses these words: "No recorded search does not show where an answer's content came from."

The decision tree

Follow this from the top for any answer with no visible search. The first "yes" decides the wording.

Enlarge image

1. Is the answer collected on a surface that reports no searches?
   (Google AI Overviews, Google AI Mode, Gemini (app), or Perplexity's
   earlier Search API)
   YES -> "not applicable". Use the engine-does-not-report phrase. STOP.

2. Did fan-out capture run for this answer?
   (Not for ChatGPT (app) answers before 2026-09-29; check the panel note.)
   NO  -> "not recorded". Report nothing about search. STOP.

3. Are any sources recorded for the answer, but no queries?
   YES -> "searches not recorded". Say sources were returned, query text
          was not recorded. Count it as searched. STOP.

4. Are there no queries and no sources, with capture having run?
   YES -> "no search". Use the observed no-search phrase, scoped to
          this run. Then ask: did the same prompt search on other days?

5. Otherwise queries are recorded -> "queries only", "grouped" or "exact".
   Report the pairing level the state supports.

Worked example

Illustrative example: Quillstone and its competitors are fictional, and the numbers are made up to show the method.

Quillstone sells document-review software to legal and compliance teams. Its content lead tracks the prompt "best document review software for a compliance team" and opens the run for one day. Five engines appear in the fan-out panel, four of them without a list of queries, and the weekly report draft says "ChatGPT (app) and Gemini (app) did not search". The lead sorts each one.

The draft was wrong on Gemini (app): the panel makes no claim about its behaviour. Its web search cell would read as did not search because it showed no source, which is a reading derived from sources, not a statement by the engine.

For ChatGPT (app), the lead does not stop at one run. Across ten daily runs of the same prompt, two show a recorded search. The report says "ChatGPT (app) recorded a search on 2 of 10 runs (20%)". It does not say why, and it does not say the other eight answers came from training data. Brieflane and Clausewise appear in both search and no-search answers, so the lead notes that competitor mentions were observed with and without a recorded search, and leaves the cause open.

Common mistakes and what this cannot tell you

• Treating "no search shown" as "no search happened". Only the no search state is a claim about the engine, and it is scoped to one run.
• Treating "no search" as a cause. It does not identify the origin of the answer's content, and no state in the product tells you that.
• Comparing web search rates across engines as if all report equally. Read which engines report what first.
• Overstating the other way. A blank does not mean the tool is broken. Some engines expose sources but not queries, and that is a documented limit.
• Ignoring dates. ChatGPT (app) answers before 2026-09-29 read as not recorded in the fan-out panel, while the web search column reads them from sources alone. A trend line across that date mixes two kinds of record.
• Assuming every captured query has a category. A classification pass labels a limited number of queries, and labels need a plan that includes them, so a query without a chip is not necessarily unimportant (Fan-out).

Engine behaviour changes. Check Engines and measurement for the current collection method before you write a claim about any engine, and keep a dated note when a state's meaning changes.

Frequently asked questions

Does "no web search" mean the AI used only its training data?

No. A no-search reading means the engine reported no searches and returned no sources on that run. It does not show what the answer's content came from, and a monitoring tool cannot establish that. Report the observation and leave the origin open.

If I see sources but no search queries, did the engine search?

For counting purposes, yes. A returned source counts as evidence of a search even without its query text, so the answer reads as searched. The state is called searches not recorded, and it describes a gap in what was captured, not a claim that the engine did not search.

Why does Gemini (app) show no fan-out at all?

Gemini (app) reports only the sources its answer cites, never the searches behind it, so its fan-out reads not applicable. Its web search is inferred from whether the answer showed a source. An answer that searched but showed no source will read as did not search.

Why does the same prompt search on some days and not others?

Engines such as ChatGPT (app) and Gemini (app) decide for themselves whether to search, and DiscoveredBy never asks them to. Runs of the same prompt on different days can therefore differ. Look at the pattern across several days before drawing a conclusion.

Should I count a "not recorded" answer as an answer with no search?

No. Not recorded means fan-out capture did not run, so the record says nothing either way, and the docs leave it out of rates. The one wrinkle is ChatGPT (app) answers from before 2026-09-29, which the web search column reads from sources alone, so check the Answer features page before quoting a rate that spans that date.

Next step

Open a prompt's page in DiscoveredBy and read the fan-out panel for each engine before you write the next weekly summary. It shows the fan-out state for each engine, which maps onto the cases above, and the Fan-out and Answer features docs define every state. For related reading, see Query fan-out: turn captured searches into content questions and How often are returned sources also credited in the answer?. To check your own answers, sign in to DiscoveredBy.