A citation is a source the answer points to, usually as a link. A mention is your brand name appearing in the answer text, with or without a link. A paraphrase is your content restated with no name and possibly no link. These are different events, and a count that merges them tells you little.
The four ways a brand shows up
| Pattern | Name in text | Link to your domain | What it tells you |
|---|---|---|---|
| Linked citation | Often | Yes | The engine surfaced your page as a source |
| Unlinked mention | Yes | No | The engine knows or repeats your name; the source is unclear |
| Paraphrase | No | Maybe | Your content may have shaped the answer, but the answer doesn't name you |
| Link without name | No | Yes | A source card or footnote points at you while the sentence stays generic |
The last row is the one people overlook. Here is a hypothetical, not an observed answer: the text says "one provider offers a free tier," and a source card beside it shows your domain. A search for your brand name would miss it. A check for your domain would catch it.
Linked citation
A linked citation gives the reader a path to your page. It is the strongest and easiest signal to verify, because you can click it and see which URL was used.
Google says it is updating AI Mode and AI Overviews with clearer inline links and source context [1]. It also says the updates are meant to help people find websites, brands and sources more easily [2]. That describes the direction of Google's interface. It does not say how often any given site gets linked, or by what rules.
Even a linked citation needs a second look. Check which URL is linked. It may be a page other than the one you expected, such as a different page on your site or a syndicated copy.
Unlinked mention
An unlinked mention is your name in the answer text with no link to you. It is ambiguous. In our reading, the engine may have pulled the name from a page it retrieved, from a third-party page that discusses you, or from what a model absorbed in training. Without a link you can't tell which.
An unlinked mention can still matter to a reader. It just isn't evidence that your site was used as a source.
Paraphrase
A paraphrase uses your facts or wording without your name. It is the hardest pattern to detect and the easiest to over-claim. Matching phrases suggest influence, but common facts are stated the same way across many sites. Treat a paraphrase as a lead to investigate, not a result to count.
Which engines expose structured citations
What you can verify depends on the interface, not the brand name of the engine. Google's announcements confirm inline links and source context in AI Overviews and AI Mode [1][2]. For any other assistant, open the interface and check three things:
- Does the answer show source links at all?
- Are links attached to specific sentences, or collected in a list at the end?
- Do links appear in every mode, or only when the assistant searches the web?
Record these answers per engine. An engine that shows no source links can only produce mentions and paraphrases for you to observe. An engine that attaches links to sentences gives you much stronger evidence per answer.
Why counting mentions without links is noisy
Each of these is a separate source of error:
- No provenance. A name in text doesn't show whether your page was the source. Two identical mentions can have completely different causes.
- Name misses. Searching answers for your brand name misses the link-without-name pattern entirely, so a name-only count can understate citations.
- Ambiguous names. If your brand is a common word or shares a name with another company, some matches will be about someone else.
- Variant spellings. Abbreviations, product names and misspellings need a list of aliases, or they are silently skipped.
- Context ignored. A name in a "which tools to avoid" list counts the same as a recommendation unless someone reads the answer.
- Run-to-run variation. In our reading, the same prompt can produce different answers on different runs. One snapshot is a sample of one. The article on why citation percentages are misleading covers what honest measurement needs.
A practical way to record what you see
Log each observation as a row, not a tally:
| Field | Example value |
|---|---|
| Engine and mode | Name of the product and whether web search was on |
| Prompt | Exact text, saved verbatim |
| Date | Day of the run |
| Name in text | Yes / No / Variant |
| Link to your domain | Yes / No, plus the URL |
| Context | Recommended, listed, compared, criticized |
| Notes | Possible paraphrase, wrong entity |
Keep linked citations and unlinked mentions in separate columns when you report them. Don't sum them into one "visibility" number.
What to do next
- Pick the engines your audience uses and note whether each one shows source links.
- Run a fixed set of prompts and log results with the table above. The manual method for checking AI chatbots gives a starting prompt set.
- Report linked citations, unlinked mentions and suspected paraphrases as three separate lines.
- For background on how this fits the wider discipline, see what GEO is.