Your brand is mentioned without a source: what can you conclude?

When an AI answer names your brand but shows no source, you can conclude one thing for certain: on that day, for that prompt, on that engine, the answer named you. You cannot conclude where the engine learned it, that a particular page caused it, or that the mention will repeat. A missing citation is a fact about the answer's links. It is not a record of the model's provenance, and a visible link cannot be reliably reverse-engineered into a training source either. What you can do is grade each claim you want to make by the evidence behind it, and this post gives you a checklist for that.

In short

• A mention is your brand named in an answer; a citation is a link the engine put in that answer as a source. They are recorded separately, and either can occur without the other.
• No visible citation does not identify a training source. It also does not prove the engine did not search.
• What you may say gets stronger with evidence: one observation, then repeated runs, then several engines. Claims about what appeared beside the mention (third-party links, a page a person has read, your own page) are separate, and all describe co-occurrence.
• Whether a mention can carry a source depends on the engine, so a bare mention means different things on different engines.
• Write findings as observations of an answer, never as explanations of why the model produced it.

What is a brand mention without a source?

It is an answer where your brand is named but the answer contains no link the engine offers as support for that statement. A mention is your brand named in an AI answer. A citation is your domain linked as a source for that answer, and retrieval is an engine pulling a page into its context without linking it. The key terms page treats these as three different things, and they feed different numbers: mentions become brand visibility, citations become citation rate.

The first thing to check is which of four situations you are in, because they look identical in a screenshot and mean different things.

Why does no visible citation not identify a training source?

Because a citation records what an engine linked in one answer, not what shaped the model. DiscoveredBy stores a citation as a link an engine puts in its own answer, pointing at a URL as the source for what it just said (Citations). Nothing in that record says how the underlying model was trained, and the answer data described here contains no field for a training source.

That has two consequences you can state plainly to a stakeholder.

First, an answer with no links might come from the model's stored knowledge, from a search the engine ran and chose not to credit, or from something else the engine does not expose. The evidence cannot separate these. Second, an answer with links may name you for reasons that have nothing to do with those links. Co-occurrence is not causation.

Does the engine decide whether to search?

For at least some engines, yes, and that changes how you read a bare mention. The engines reference says DiscoveredBy never asks ChatGPT (app) or Gemini (app) to search, so some of their answers cite no sources at all. A source-less mention on those engines can simply be an answer where the engine chose not to look anything up.

Engines also differ in what they report, which is why the same absence of a link is not comparable across them:

• Cited sources only. Google AI Overviews, Google AI Mode and Gemini (app) report only the sources their answer cites, with no list of pages searched. ChatGPT (app) is listed the same way on Retrieved vs cited, though since 29 September 2026 it also reports the pages it found.
• Full or partial search results. Claude, Perplexity and Grok expose more of what their search brought back. For Claude, every page its search returned is stored as a citation, so "cited" there is broader than "credited in the text". Grok is similar: its saved citations include every source its search came across (Metrics defined).
• Perplexity marks the sources its answer used, so a Perplexity citation is one the answer text points at.
• Google may show no AI answer at all. That run has no answer text and no citations, so it is not an answer that named you without links. It is a different state, and answer-based numbers such as brand visibility leave it out.

Enlarge image

So "mention without a source" is a cleaner observation on some engines than others. On Claude, an answer that names you alongside stored citations is still not one where the text credited a source for you. For Claude, Perplexity and Grok, the Retrieved vs cited screen separates pages an engine returned from pages the answer attributed; for the other engines it does not compare the two.

How do you check what the answer actually linked?

Open the answer and read its sources before you write anything down. The Citations screen has an owner control (All, Your Citations, or External) and an Every citation tab that lists one row per citation with a link to the source. The Sentiment screen's Evidence card holds the mention list and the citation list on two tabs (Sentiment).

Enlarge image

Three checks matter before you grade a claim:

1. Which links are in this specific answer? A domain that is cited often elsewhere tells you nothing about this answer.
2. Is the link yours by domain, and only by domain? Attribution to you is one domain compared to another, done once when the citation is stored. A syndicated copy of your article on another domain is attributed to that other domain, never to you.
3. What did the engine's own text say? Brand reasons stores, for each reason an answer gives about a brand, a quote copied from the answer, so you can read the stated reason in the answer's own words.

One engine quirk can confuse this check. ChatGPT (app) and Gemini (app) can write a small link after a sentence they support, with the site's address as the link text. Mention extraction blanks those links out, so a site named only in such source links is not counted as a brand mention (Engines). If your domain appears as a source link but the answer is not counted as naming you, that is the rule working as designed, not a missed detection.

What can you conclude? The claim-strength checklist

Grade every claim by the weakest evidence it rests on, and use the wording in the last column. This is the deliverable: copy the table into your reporting template.

Two rules make the table work. A claim is no stronger than its weakest input, so a level 5 claim built on one answer is still one observation. And levels 4 to 6 describe co-occurrence, however tidy the match looks; they are different questions about what appeared beside the mention, not a strict ladder above level 3.

Here is a log you can copy for each mention you plan to report:

Mention claim log
Prompt:
Engine (and collection: app or API):
Dates checked (list each):
Named on N of M runs:
Other engines checked:
Links in the answer (domains):
Any link to our domain? (yes/no):
Did a person open the cited pages? (yes/no, who, when):
Do the pages state what the answer states? (yes/no/partly):
Claim level reached (1-6):
Wording used in the report:
What we cannot say:

Worked example: Quillstone

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. Over a 28-day window, the team's monitoring shows Quillstone named in 18 analysed answers across four engines. They split those by what the answers linked.

Checking the arithmetic: 6 + 7 + 5 = 18, and each row adds to its own total.

Five answers name Quillstone with no links. The temptation is to report "AI knows us without needing our site". The claim log says otherwise. Two of the five are ChatGPT (app) answers, an engine that decides for itself whether to search, so a bare mention may just be a no-search answer. Two are Claude answers, and Claude stores the results its search returned as citations, so no links means no stored results; the record does not say why. One is a Google AI Overview.

The seven answers with other-domain links are more interesting. The analyst opens the cited pages. In three of them a third-party comparison page lists Quillstone next to Brieflane and Clausewise, matching the sentence in the answer. That is level 5 for those three answers: the cited page states what the answer states. In the other four, the linked pages do not mention Quillstone at all. The answer cites them for other sentences, so those four stay at level 4. The six answers that linked Quillstone's own domain reach level 6 at most: the page was in the answer's sources, and nothing more is claimed.

The report says: "Quillstone was named in 18 answers across four engines in the window. In 3 answers the cited third-party page listed Quillstone in the same terms. In 5 answers no source was shown, and we cannot say why." It does not say where the model learned about Quillstone, and it does not propose that the comparison page "drives" mentions. That page becomes a hypothesis worth watching, for example by checking whether mentions on the same prompts change if it is updated, which is a separate test.

Enlarge image

What can this analysis not tell you?

It cannot tell you why an engine produced an answer, and it cannot tell you the answer will repeat. The limits, stated plainly:

• No provenance. Neither a link nor its absence identifies training data or the pages an engine internally relied on. For engines that report only cited sources, the pages it searched are simply not exposed.
• No causation from co-occurrence. A cited page that mentions you sits in the same answer; the record does not say the answer used it for that sentence.
• Answers vary. One answer is one observation. Levels 2 and 3 exist because a single run may not repeat.
• Collection channels differ. Do not average bare-mention rates across app and API engines as if a missing link meant the same thing.
• The reverse limit is real too. A source-less mention is still a real mention. It counts toward brand visibility and is a legitimate observation of what the engine said.

Common mistakes to avoid: treating "no citation" as "no search"; reading a cited page's title instead of its text; crediting a source because it ranks first in the list; and pooling mentions from engines that expose very different amounts of source detail.

Frequently asked questions

Does a mention without a citation mean the AI learned about us from training data?

No, it does not. A missing citation shows that no link was recorded for that answer, and that is all. The engine may have searched and chosen not to credit anything, may not have searched, or may be using knowledge you cannot observe.

Is a mention with no source worth less than a cited mention?

They measure different things, and you should not merge them. A mention counts toward brand visibility whether or not the answer linked anything. A citation counts toward citation rate only if the link points at your domain.

Can I find which page an AI used to mention us?

Not reliably. You can see which pages an answer linked, and on some engines which pages it returned, using Retrieved vs cited. Neither tells you which page shaped the specific sentence about you unless a person reads the page and it matches.

Why does the same prompt mention us with links one day and without the next?

ChatGPT (app) and Gemini (app) decide for themselves whether to search, so a linked answer and an unlinked one can both occur for the same prompt on those engines. Track the mention across several days before you read a pattern into a single answer.

Should we publish content to turn bare mentions into cited ones?

Treat that as a hypothesis, not a plan with a promised result. Updating a page that already appears in related answers is a reasonable test, but there is no guarantee of a change in citations or mentions. Compare like for like over a comparable window before deciding it worked.

How does this differ from a wrong or invented claim about us?

A bare mention says only that you were named. An incorrect statement in the answer is a separate problem; see what to do when AI gets your pricing or product facts wrong.

Next step

If you want to separate mentions from citations in your own data, the terms are explained in AI visibility, brand mentions, and citations: what each number tells you, and the glossary entries for brand mention and citation give the short definitions. The sources feature shows what DiscoveredBy records about cited pages.

To grade your own mentions, sign in to DiscoveredBy, open a prompt where your brand appears without links, and fill in the claim log above for the answer before you report it.