Formatting errors in a bibliography belong to that category of mistake nobody praises you for avoiding and somebody always finds.
I once edited a draft where the prose was good and nineteen of the thirty-one reference entries were wrong. The author said the list was auto-generated. It was. It was also half-finished, because the metadata fed into it was half-finished.
Tools can do more with formatting than you would expect, and there is a floor they cannot get under.
The two systems
GB/T 7714-2015 and APA 7th are the pair most Chinese academic writing runs into. The logic differs.
| GB/T 7714-2015 | APA 7th | |
|---|---|---|
| In-text | Numeric [1], or author-year | Author-year (Chen, 2020) |
| Authors | All if 3 or fewer; if more, first 3 plus 等 | All if 20 or fewer; if more, first 19, ellipsis, last |
| Type marker | Required, e.g. [J] [M] [EB/OL] | None |
| Punctuation | Full-width, Chinese punctuation | Half-width, italic journal titles |
| Web sources | (publish date)[access date]. URL | Usually no access date unless the content changes |
Small differences, and mixing them is visible at a glance. The most common: an author-year citation inside a Chinese paper that still carries a GB type marker, producing something like [J] Zhang, 2020. No layout tool fixes that for you.
What auto-generation needs first
A journal article needs author, title, journal, year, volume, issue, page range, and a DOI if it has one. A web source needs author (which may not exist), title, site name, publish date, URL, access date.
Miss one field and the output is half an entry. On the web, missing fields are normal. A WeChat article has no byline. A news site has no publish date. A technical blog has two dates that disagree with each other.
So the real work isn't formatting. It is assembly.
I do it in three passes.
Capture what is on the page. While you highlight, the extension records the title, site name and publish date. All of it is already rendered on screen.
Fill gaps from Crossref. If there is a DOI, Crossref has the author, volume, issue and pages, complete. Mechanical, worth doing.
Hand edits win. This one matters. Scraped metadata is frequently wrong. The site name becomes a section name ("Culture" instead of the publication). The date becomes "3 days ago". The author becomes the editor. So wherever you have corrected something, the corrected value has to survive export. It cannot be overwritten by a re-fetch.
Inkory exports in that order. If it grabbed something wrong you can fix it before export, and your value is what gets formatted. That beats "fully automatic", because a reference list is a document an editor reads, and you are the one signing it.
A checklist you can run
Don't paste the export straight in. Five checks, two minutes.
- Author count and the truncation word. 等 in GB, "et al." in APA. Is the count right?
- Year. Publication year, not update year, not the year you happened to see it. Preprints and final versions are often a year apart.
- Volume and issue. Chinese journals frequently have an issue and no volume, in which case write the issue only. Page numbers come from the printed pages, not the PDF index.
- DOI. Paste it after doi.org and click. If it doesn't resolve, the entry is fake or stale.
- Access date. Under GB/T 7714, a web page with no fixed publish date requires one. With a clear publish date it is optional, but it has to be consistent across the list.
The fifth is where people get caught. Twenty of thirty entries carry an access date, ten don't, and the editor reads that as carelessness.
One counterintuitive piece of advice
If you are writing a thesis, don't expect a one-click export to be submittable.
Not because the tools are bad, but because behind every format rule sits a judgement. Is this a journal article or a conference paper. Does this issue have a volume number. Is this a preprint or the final version. Is this web page stable enough to cite at all. Machines cannot make those calls, and the calls decide whether the output is right.
The legitimate use of a tool is this: it removes about 80% of the mechanical labour and leaves the remaining judgement to you. What you save is typing time, not the obligation to check.
One thing to do when you are finished. Read the reference list as if somebody else wrote it, entry by entry. Whatever you can follow, a reader can follow. Whatever you cannot is probably broken.
Related: Zotero and browser highlights don't do the same job · AI drafts: how to find the citations that aren't real