Inline citation: definition in AI answers
99Visibility Research · Updated October 21, 2026 · 4 min read
Short answer
An inline citation is a reference placed inside an AI answer, next to the claim it supports, that points to the source behind that claim. Engines attach it in one of two ways. Span-level citation ties a short quoted passage of the source to the sentence. Source-level citation ties a numbered link to a whole page. Because the units differ, citation counts from different engines cannot be added together.
Key takeaways
- An inline citation sits next to the claim it supports, not in a list at the end of the answer.
- Span-level citations quote a specific passage, while source-level citations point to a whole page by number.
- Anthropic's API cites text spans and Perplexity cites whole sources, so one citation on each is not the same unit.
- A citation shows a source was credited, not that it shaped the answer, so counts are a noisy proxy for influence.
- Compare citation numbers within one engine over time, never as a single sum across engines.
What an inline citation is
An inline citation is a pointer embedded in the answer text itself. It connects one claim to the source the engine credits for it. It differs from a reading list at the bottom of an answer, which says nothing about which source backs which sentence.
In AI search, inline references are generated by the engine, not written by the publisher. You control whether your page can be fetched and what it says. You don't control whether the model attaches your URL to a claim. For the broader concept, see what an AI citation is.
How it works: span-level vs source-level
Engines don't agree on what a citation points to. As of October 2026, the documented approaches fall into two groups.
| Style | What the citation points to | Example |
|---|---|---|
| Span-level | A short quoted passage from the source | Anthropic's web search tool returns a cited_text field of up to 150 characters |
| Source-level | A whole page, usually by a numbered marker | Perplexity cites whole sources by integer index |
A cited text span tells you exactly which sentence the engine relied on. A numbered link tells you only which page it credited. Anthropic's tool docs also expose the URL, title and a page_age field for each result, which the Claude web search documentation covers in detail.
OpenAI's API adds a third distinction. It returns the URLs the model consulted separately from the URLs it cited, and OpenAI states the consulted list is often larger. So a page can be read and still never appear as a citation. The differences between Claude and Perplexity citation styles are worth knowing before you compare any reports.
Why it matters for AI visibility
The unit problem breaks naive reporting. One span citation on Claude and one numbered link on Perplexity measure different things. Adding them produces a number with no shared denominator. Google's own documentation adds a further split: AI Overviews and AI Mode "may use different models and techniques" and can return different links.
Citation is also not the same as use. Research on how LLMs cite (arXiv:2606.28358) repaired over 90% of missed citations by amplifying specific attention heads, without changing the sources. Citation emission is partly separate from which sources influenced the answer.
What to do with this:
- Track citation rate per engine, and label the unit each engine uses.
- Where an engine gives a span, read it. It shows the exact sentence the engine took from you, so you can check whether AI quotes your site accurately.
- State key facts (a price, a number, a one-line definition) plainly in HTML that crawlers can fetch without running JavaScript. That gives an engine a sentence it can quote. It does not guarantee a citation.
Common misconceptions
"More citations means more influence"
Not reliably. Retrieved sources outnumber cited ones in OpenAI's API, and citation output can diverge from source use. Treat counts as a proxy.
"A citation total across engines is my visibility"
It isn't. Spans and whole-page links are different units, and retrieval fires at each engine's discretion per query. Any blended score averages answers given with and without live retrieval.
"Markup makes engines cite a passage"
No evidence supports that. Google says structured data isn't required for its generative AI features. The comparison of structured data and plain-text facts explains what the evidence actually shows.
Frequently asked questions
What is a cited text span?
A cited text span is the exact passage of a source that an engine attaches to a claim in its answer. Anthropic's web search tool returns it as a `cited_text` field of up to 150 characters. It is more precise than a numbered link, because it shows which sentence the engine relied on, not just which page it credited.
Why can't I add citation counts from Claude and Perplexity together?
They count different units. Claude's API cites short text spans, while Perplexity cites whole sources by integer index. One page could earn several span citations in a single Claude answer but only one numbered link in Perplexity. Summing them mixes passages with pages. Report each engine separately and compare each one against its own history.
Does being retrieved mean I will be cited?
No. OpenAI's API separates the sources a model consulted from the sources it cited, and OpenAI states the consulted list is often larger. A page can be read, used as background, and never shown as an inline reference. Access and retrieval are necessary for a citation, but neither guarantees one.
Can I make an engine quote a specific sentence from my page?
Not directly. The model decides what to cite. What you control is whether the page is reachable by AI crawlers and whether the fact is stated plainly in the HTML before JavaScript runs. A specific, quotable sentence gives the engine something to cite. Whether it does is outside your control.
Sources
This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.