What Actually Happens When You Cross-Check a Thesis

You are three weeks from submission. The chapters are done, the argument holds, and somewhere in the reference list there are entries you added in year two and have not looked at since. You know, in a vague and uncomfortable way, that the citations and the bibliography have drifted apart. You do not know by how much, and checking 250 references by hand is a job you keep postponing.

This is a walk through what that check actually looks like when a machine does it — what each stage asks of you, what it costs, and where the decision points are. No marketing abstractions; just the sequence.

Step one: the document, not the reference manager

The first design decision is the one that shapes everything after it. CitationLab starts at the document your examiners will read — not at Zotero, not at Mendeley, not at your .bib file.

That matters because your reference manager holds what you meant to cite. Your thesis holds what you actually cited. Those two sets diverge over four years, and only one of them is going to be examined. Drop the whole thesis in and the engine reads all of it — body, appendices, and any chapter-end reference lists that appear after the main bibliography.

Upload the .docx if you have one. PDF works, but text pulled from a PDF — especially tables, multi-column layouts or line-wrapped text — is occasionally less precise than a native Word file. The upload screen says this before you choose rather than after you get odd results, which is the right time to say it.

Worth knowing up front: upload, extraction, review and cross-check are free. The AI steps are priced in credits, and the cost sits next to the button before you run anything. You can get all the way to a complete deterministic audit of your thesis without spending a credit.

Step two: review, in three separate buckets

Extraction produces three piles, each with its own count, and they are kept apart deliberately: the citations pulled from your text, the entries parsed from your reference list, and the dates the extractor could not place.

That third pile is the interesting one. A date the extractor cannot confidently attach to anything does not get thrown away and does not get guessed at — it is screened into an unresolved bucket with the reason attached. You can look at why. Nothing is discarded silently, which means the thing you are most afraid of — a citation quietly vanishing from the analysis — is the thing the design specifically prevents.

Here you meet the first optional AI step. Ref[In] AI cleanup tidies messy in-text forms — spacing, ampersands, inconsistent et al. — and flags entries that do not read as citations at all. It is advisory. Nothing is rewritten in your thesis. Every suggestion is yours to approve or dismiss, and the price, your balance and the number of runs left are shown before you press, not after. It is charged once per session; re-runs are included.

Step three: de-duplicate, and the 2014a rule

De-duplicate is the way out of Review, and the AI pass has to have run first. The same source cited two different ways collapses into one entry with the citation count preserved — you do not lose the fact that you cited it eleven times.

One rule here is worth dwelling on, because it is where most tools quietly corrupt a thesis:

Suffixes are never invented — a 2014a is only a 2014a if your thesis says so.

If you cited Ahmed twice in 2014 and never disambiguated them, a tool that helpfully invents 2014a and 2014b has just introduced a discrepancy between your text and your bibliography that did not exist before — and you will not notice until an examiner does. Refusing to invent a suffix leaves you with a visible problem instead of a hidden one. That is the correct trade.

Step four: the cross-check itself

Now every citation is paired against the reference list parsed from your own document. Matching runs on what the guide calls the Date-ID-Style backbone, in a fixed order of confidence:

1. Shared identifiers first — a DOI, PMID, legal instrument or standard number. If both sides carry the same identifier, the pairing is not in doubt.
2. Then exact author-year — the conventional match.
3. Then the reference line as written — falling back to the literal text when nothing stronger is available.

This step is deterministic and free. No model decides whether your citation matches your reference; it is a rule applied the same way every time, which is why you can re-run it as often as you like without cost and get the same answer.

Four outcomes: matched, questionable (a pair resting on weak evidence), missing (cited but absent from your list), and orphan (listed but never cited). Match review sits on top with verified, questionable and not-judgeable counts, and an AI re-check available on demand — the first ten pairs are free.

Above that, the alignment audit re-reads the pairs resting on weak evidence — a bare year marker, two different styles, a contested reference — and proposes corrections rather than quietly accepting them.

The door that closes behind you

Read this bit before you click. Proceeding to Verify & Edit closes Organising permanently. The interface says so before you press, not after — but it is still a one-way door, and it is the single point in the workflow where you should stop and make sure you are ready.

Before you go through it, download your Level 1 report. It is the deterministic audit of your thesis as it stands, available as a PDF or a shareable link — and the link stays live for a supervisor even after Organising has closed. If your supervisor wants to see where you started, this is the artefact to send.

Step five: finding what is genuinely missing

Whatever is still missing goes to the finder, and the ordering here is deliberate. Six databases — Crossref, OpenAlex, Semantic Scholar, PubMed, Google Scholar and Google Books — first establish that a citation is genuinely missing rather than merely misspelled, hyphenated or accented differently. That verification happens before any AI judgement enters the picture. PubMed is only consulted where a citation reads as biomedical, so a sociology thesis is not sent to a medical index.

Your personal library searches first. Upload PDFs, DOCX or BibTeX and your own sources outrank the internet — which is what you want, because the PDF sitting in your folder is the edition you actually read and cited, not a similar-looking record someone else catalogued.

Step six: annotate, format, export

Once references are finalised, annotation builds in order and you can produce an annotated bibliography and an annotated paper alongside the citation run summary. Formatting supports fifteen citation styles, with bibliography styles drawn from the Citation Style Language project and rendered through Citation.js.

The Level 3 report is issued at export: in-text citations and reference list updated and validated. Because the score bands are on a fixed scale, you can put your Level 1 and Level 3 reports side by side and read the difference directly. That comparison is the honest record of what the work actually changed.

What to expect the first time

Expect a worse number than you were hoping for, and do not read anything into it beyond what it says. The Ref[In] Score — short for REFerence [IN]tegrity — is a referencing-quality measure: not a similarity score, and not a finding of misconduct. A long, honest, carefully-argued thesis routinely comes back with substantial gaps on the first pass, because that is what four years of drafting does to a reference list. The number is a to-do list, not a verdict.

• Walk every screen in the product tour, or read the full Product Guide (PDF).
• Try it on your own document with the live demo.
• See what a real run turns up in one orphan, one missing and finding missing citations.
• Check credit costs on the pricing page, or browse help and the FAQ.

The reason to run it three weeks out rather than three days out is simple: finding a missing reference is quick, but tracking down a source you cited in 2023 and can no longer identify is not. Give yourself the room.