> ## Documentation Index
> Fetch the complete documentation index at: https://docs.scite.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Citation analysis tools

> Traverse the citation graph, page through citing statements, and run field-level bibliometrics through MCP.

Four tools analyze how papers cite each other. Use them when the connections between papers matter more than keywords.

| Tool | Answers |
| - | - |
| `citation_graph` | Which papers does this build on, which build on it, and do they agree? |
| `get_citation_statements` | What exactly do citing papers say about this one paper? |
| `bibliometrics` | How big is this field, who leads it, and how contested is it? |
| `citation_dynamics` | When were these papers cited, by whom, and did citation continue after a retraction? |

Requires a Scite premium subscription. Counts reflect Scite's index coverage, so say so when you report them.

## citation\_graph

`citation_graph` traverses citations from up to 10 seed DOIs. Each edge is `{s, t, d}`: the citing paper `s`, the cited paper `t`, and `d`, the hop distance from a seed.

```json theme={null}
{
  "name": "citation_graph",
  "arguments": {
    "seeds": ["10.1126/science.1225829"],
    "direction": "in",
    "depth": 1,
    "include_intent": true,
    "max_edges": 200
  }
}
```

| Parameter | Type | Description |
| - | - | - |
| `seeds` | string\[] | Up to 10 seed DOIs. Required unless `claim` is given |
| `claim` | string | Start from a claim instead of seeds. See [the claim graph](#the-claim-graph) |
| `direction` | string | `out` (papers the seeds cite), `in` (papers citing the seeds), or `both`. Default `both` |
| `depth` | integer | `1` for direct citations, `2` for citations of citations. Default `1` |
| `include_intent` | boolean | Add each edge's Smart Citation `type`, `section`, and per-stance `confidence` |
| `include_snippets` | boolean | Also attach up to 3 citing-text `snippets` per edge. Implies `include_intent`. Keep `max_edges` modest |
| `max_edges` | integer | Cap on edges returned. Default `500`, maximum `2000` |

The response carries `edges`, a `papers` map of every DOI to its `title`, `year`, and any `editorialNotices`, and `truncated` when `max_edges` was hit.

Edge `type` is `supporting`, `contradicting`, or `mentioning`. `contradicting` here is the stance `search_literature` reports as `contrasting`. Snippet text returns only for open-access and unrestricted sources; closed-publisher edges still carry type and section.

Seeds expand automatically to their linked preprint and published versions, so a paper split across an arXiv DOI and a journal DOI traverses as one work.

### Check coverage first

`seed_coverage` maps each seed to how many of its edges resolved at hop 1, and `low_coverage_seeds` lists seeds with too few to be reliable. arXiv-heavy fields, such as machine learning, often under-resolve. A thin graph means missing edges, not a poorly connected paper; fall back to `search_literature` for low-coverage seeds.

### Analyses you can derive from edges

The edge list answers the classic citation-analysis questions without extra calls:

* **Common references**: traverse `out` from several seeds and keep DOIs that appear as `t` for every seed.
* **Common citers**: traverse `in` and keep DOIs that appear as `s` for every seed. These are often surveys.
* **Bibliographic coupling**: traverse `both` at depth 2, so the second hop reaches other papers that cite the seed's references. Rank those papers by how many reference targets they share with the seed.
* **Co-citation**: rank papers frequently cited by the same citers.
* **Citation classics**: rank DOIs by in-degree.

Add `include_intent` first to filter any of these to supporting or contradicting edges.

### The claim graph

Pass `claim` instead of `seeds` and the tool runs `check_claim` for you. The response adds `claim_check`, with the verbatim supporting, contradicting, and stating passages, and builds the graph from the papers that state the claim, always with intent on the edges.

From there, call `read_fulltext` on `claim_check.readNext` (the stating papers) and `claim_check.readNextCiting` (the papers holding the counterexamples and evidence) to read the experiments, assumptions, and proofs behind each node.

## get\_citation\_statements

`get_citation_statements` lists the Smart Citation statements that cite one paper, with paging. Unlike `citation_graph`, it pages cleanly through papers with tens of thousands of citers.

```json theme={null}
{
  "name": "get_citation_statements",
  "arguments": {
    "doi": "10.1038/s41586-020-2012-7",
    "type": ["contrasting"],
    "limit": 20
  }
}
```

| Parameter | Type | Description |
| - | - | - |
| `doi` | string | Required. DOI of the cited paper |
| `type` | string\[] | Any of `supporting`, `contrasting`, `mentioning` |
| `term` | string | Keyword filter over the citing text, journal, and author names |
| `limit` | integer | Statements per page, `1`–`100`. Default `20` |
| `offset` | integer | Statements to skip. Default `0` |

The response carries `total` and `hasMore`. Stance uses Scite's expert label when one exists, else the classifier's; `confidence` is the classifier's 0–1 score. `snippet` is verbatim for open-access and licensed sources; otherwise `contentDenied` is `true` and only type, section, and confidence return.

For very heavily cited papers `approximate` is `true`: `total` counts a bounded slice, `offset` is capped (`capped: true` when yours was past it), and `hasMore` turns false at the last reachable page. Narrow with `type` or `term` rather than paging further.

## bibliometrics

`bibliometrics` computes field-level citation analytics in one aggregation pass. It counts citations received by the papers in scope and returns no paper text.

```json theme={null}
{
  "name": "bibliometrics",
  "arguments": {
    "term": "\"base editing\"",
    "date_from": "2016",
    "group_by": "journal",
    "order_by": "citations",
    "baseline": { "term": "CRISPR", "date_from": "2016" }
  }
}
```

### Scope

Pass at least one filter: `dois` (up to 2,000), `term`, `title`, `abstract`, `journal` (word match) or `journals` (exact keys), `author`, `affiliation`, `publisher`, `topic`, `paper_type`, `year`, `date_from`, `date_to`, `citation_types`, `sections`, `has_tally`, and the editorial filters `has_retraction`, `has_concern`, `has_correction`, `has_erratum`, and `has_withdrawn`.

`term` matches title, abstract, and full text, and every word must match. Check `total_papers` and tighten with quotes, `AND`/`OR`, dates, `topic`, or `journals`.

### Grouping

`group_by` sets the bucket dimension: `year` (default), `journal`, `author`, `affiliation`, `topic`, `mesh`, `paper_type`, `oa_status`, `editorial_notice`, `citation_type`, or `citation_section`. `order_by` ranks buckets by `papers`, `citations`, `supporting`, `contrasting`, or `mentioning`; `year` buckets are always chronological. `limit` caps buckets at up to 100.

### What it returns

| Field | Contents |
| - | - |
| `total_papers`, `date_range` | The scope you measured. Always report these |
| `overview` | Annual growth rate, mean document age, co-authors per paper, single-author and multi-institution shares |
| `citations` | Totals by stance, plus per-paper mean, median, p90, p99, and max |
| `indices` | h, g, fractional h, a, e, and m-quotient, with an `exact` flag |
| `distribution` | Uncited count and rate, and counts above 10, 20, 50, 100, and 1,000 citations |
| `top_papers` | The 5 most-cited papers, with DOIs |
| `buckets` | Per bucket: papers, share, citations, citations per paper, stance counts, and `contrast_rate` |

`contrast_rate` is contrasting ÷ (supporting + contrasting). Most fields sit around 0.05–0.08; a bucket well above that is worth reading.

### Normalized and custom indices

Pass `baseline`, an object with the same filters describing the field, to get `normalized`: `mncs` (1.0 = field average), `tncs`, and Leiden `pp_top1`, `pp_top10`, and `pp_top50`.

Pass `custom_indices` to compute your own h, g, or at-least indices over any tally field. For example, `{"type": "h", "field": "contrasting", "name": "negative_h"}` returns the number of papers each disputed by at least that many contradicting citations.

Set `include_raw: true` (with `raw_limit` up to 1,000) to get per-paper rows for computing other indices yourself.

> [!WARNING]
>
> Indicators support judgment; they don't replace it. Normalize across fields with `baseline`, and don't rank individual researchers on h-index or journal impact factor alone. See the Leiden Manifesto and DORA.

## citation\_dynamics

`citation_dynamics` analyzes when and how a set of papers was cited. Pass up to 100 DOIs, or up to 500 with `summary_only: true`. The DOIs usually come from `bibliometrics` (`top_papers` or `include_raw`), `search_literature`, or `citation_graph`.

```json theme={null}
{
  "name": "citation_dynamics",
  "arguments": {
    "dois": ["10.1126/science.1225829", "10.1016/j.cell.2013.02.022"],
    "include_trajectories": true
  }
}
```

| Parameter | Type | Description |
| - | - | - |
| `dois` | string\[] | Required. Papers to analyze |
| `summary_only` | boolean | Return the aggregate only, allowing 500 DOIs |
| `include_trajectories` | boolean | Add each paper's `citations_by_year` |
| `include_network` | boolean | Add `co_citation`, `coupling`, and `co_authorship` pairs and `top_authors` within the set |

Per paper, it returns citation timing (`first_citation_year`, `peak_year`, `cited_half_life`, `within_1y`/`within_3y`/`within_5y`, `delayed_recognition`), `self_citation_rate`, stance counts, editorial notices with `citations_after_notice` and `stance_after_notice`, and reference-list measures (`mean_reference_age`, `price_index`, `journal_insularity`).

The aggregate adds `citations_by_citing_year`, `cohorts` (cited year × citing year), `stance_by_citing_year`, `references_by_year`, and `top_cited_references`.

Common uses:

* **Journal citation-rate estimates.** Pass one journal's papers from years Y-1 and Y-2 (from `bibliometrics` with `journals`, `year`, and `include_raw`). Then (`cohorts[Y-1][Y]` + `cohorts[Y-2][Y]`) ÷ papers approximates a two-year impact factor, and `cohorts[Y][Y]` ÷ papers approximates immediacy for year-Y papers. These are estimates from Scite's index, not official Journal Impact Factors: the denominator counts every paper supplied rather than citable items only. Above 500 papers, `summary_only` covers only the 500 most-cited, so the value overstates the rate.
* **Post-retraction citation.** `stance_after_notice` still showing supporting citations means papers kept citing the work as valid.
* **Shifting stance.** A rising contradicting share in `stance_by_citing_year` is a signal to read the contradicting passages, not a verdict that the finding is weakening.
* **Fair comparisons.** Use `within_3y` instead of all-time counts, and quote `self_citation_rate` beside any individual index.

## Related

* [Verification tools](https://docs.scite.ai/mcp/tools/verification.md): `check_claim`, which feeds the claim graph
* [Literature tools](https://docs.scite.ai/mcp/tools/literature.md): `read_fulltext` for the papers you find in the graph
* [Citation model](https://docs.scite.ai/concepts/citation-model.md): how statements, tallies, and citing publications relate


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.