Mark every sentence that asserts something about the outside world, then check whether a source sits next to it. You need three categories: statistics and figures, named entities, and comparative or superlative claims. The ForgeRank published dataset (https://forgerankai.com/static/data/ai-content-quality-29-drafts.json) shows 51 of 2,839 claims across 29 drafts needed a source, and 38 had none attached.
What is this method for?
This guide is for anyone publishing AI-assisted drafts who wants a repeatable check that needs no paid tooling. You need a cursor, a text editor, and roughly 20 minutes per 1,500-word draft. The method assumes you can read a sentence and decide whether it asserts a fact about the world or describes how something works. By the end you will have a marked draft where every unsupported assertion is visible, plus a short list of what to fix or cut. It will not verify that a cited source actually says what the draft claims — that is a separate pass.
Step 1: Separate assertions from mechanism sentences
Read each sentence and ask one question: does this sentence make a claim about the world outside the text, or does it describe how something works? Mechanism sentences explain a sequence, a cause, or a rule, and they need no citation. Assertion sentences name a number, an organisation, a version, or a ranking, and they do. Mark assertions with a single character in the margin — a pipe symbol works because it never appears in prose.
Run this pass before you look for sources. Mixing the two passes is what makes people give up halfway. The ForgeRank published dataset recorded 2,839 claims across 29 drafts, and only 51 of those were judged to need a source. That ratio is the point: you are hunting a small set, and the rest of the draft is noise you can skip.
When you finish, the margin should hold a sparse column of pipe symbols. If nearly every line is marked, you are marking opinions and mechanism sentences as assertions; re-read one paragraph and check whether the sentence would still be true if every number in it were deleted.
Step 2: Mark statistics and figures first
Statistics are the fastest category to clear because a number is either followed by a source or it is not. Circle every numeral, percentage, price, date, and duration in the draft. Then, for each one, look at the same sentence and the sentence before it for a named source. A figure with no source name attached is unsupported, regardless of how plausible it reads.
The ForgeRank published dataset contains a useful counter-example about volume. The highest-scoring draft in that set carried 173 claims, and a draft that landed at 4.8 carried 156. The draft with more claims scored higher. Claim density tells you nothing about whether the claims are sourced, because a mechanism-heavy draft accumulates claims without needing a single citation.
Numbers that change on their own schedule — pricing tiers, free-plan caps, API quotas — need a second mark. Put a question mark beside them and add the words "confirm current" when you rewrite. A price with no date attached will read as wrong within a year even when it was accurate the day you wrote it.
The boundary: a figure that describes your own process or a hypothetical calculation needs no external source, but it does need a label. Write "hypothetically" or "in this example" in the sentence, or a reader will treat your arithmetic as a measured result.
Step 3: Mark named entities
Circle every proper noun that is not a place you can point to on a map: company names, product names, report titles, version numbers, and job titles attached to a claim. A named entity is a claim about a third party, and the person publishing it carries the consequences — a correction request, a takedown demand, a lost client.
The mechanism is straightforward, because attaching a real organisation's name to a figure it did not publish converts your sentence into a factual claim about that organisation. Two failure shapes dominate. The first is a real organisation plus an invented number. The second is an invented report title plus a plausible year, which is the pattern that gets caught most often because the report cannot be found.
Version numbers belong here too. "Works with Windows 11" is a compatibility claim, and compatibility claims need an exact device and an exact OS version, not a category. If you cannot name the model and the build, cut the sentence rather than soften it.
The boundary: a named entity used as an ordinary noun needs no source. Writing that a plumber pays for three subscription tools asserts nothing about those companies. Writing that a specific tool costs a specific amount does.
Step 4: Mark comparative and superlative claims
Highlight every "best", "fastest", "cheaper than", "more accurate", "leading", and "top". These are measurement claims wearing ordinary words, and they need either a named test or an explicit statement of your selection criteria.
The mechanism matters here because a superlative implies a comparison set. "The fastest option" implies you measured the alternatives under the same conditions. If you did not, the sentence is unsupported even when the underlying product genuinely is fast. A reader who checks will find no test, and the whole page loses credibility with it.
The fix is a decision rule, stated plainly. Write the criteria and the cutoff: which options qualified, which were excluded, and why. A rule like "options must ship a free tier and a documented API to qualify" is checkable, and it converts an unfalsifiable superlative into a claim a reader can audit.
The boundary: a superlative inside a quoted source keeps that source's authority, so write it as the source's claim rather than yours. A superlative you generated from your own comparison needs the criteria and the conditions you tested under.
Two beliefs about AI drafts, and what the data says
The common belief holds that AI-assisted drafts fail because they contain too many claims. The evidence points elsewhere. In the ForgeRank published dataset, twenty of twenty-nine AI-assisted drafts scored before publishing landed at or below 4.8, and none landed between 5.0 and 6.9. The full distribution was 4.0 ten times, 4.8 ten times, 7.0 three times, 7.2 four times, 7.8 once, and 8.5 once.
The rival belief — that the failures come from a single weak dimension rather than overall claim volume — is what the tier data supports. Tabulating the four dimension tiers behind every row shows each step in the distribution is one dimension moving. All ten 4.0 drafts scored 4.0 on all four dimensions. All ten 4.8 drafts scored 4.0/4.0/7.0/4.0. The three 7.0 drafts were flat at 7.0. The four 7.2 drafts were 7.0/7.0/8.0/7.0. The single 7.8 draft was 8.0/7.0/9.0/7.0, and the single 8.5 draft was 8.0/9.0/9.0/8.0.
Read that pattern and the practical consequence is clear: a draft sits at 4.8 because one dimension lags while the other three hold, and lifting that one dimension lifts the whole score. Marking unsupported claims targets the dimension that lags in every low-scoring row.
One boundary on the dataset itself. Source analysis runs on a capped number of claims per draft, so 13 of the 29 drafts hit that cap, according to the ForgeRank methodology page. On a long draft, the claim count you measure will undercount what is actually there.
What if it doesn't work?
The margin fills with marks and nothing gets fixed. You marked opinions and mechanism sentences. Re-read one paragraph and test each marked sentence by deleting every number in it; if the sentence survives as a true statement, unmark it.
A figure has a source name but you cannot find the source. Treat it as unsupported. A plausible-sounding citation is worse than no citation, because it is a factual claim about a third party that the third party never made.
The draft cites a report you cannot locate. Search the exact report title in quotation marks. No result means the title was generated, and the figure attached to it goes with it.
Two sections repeat the same statistic. Cut one. Repetition across sections reads as padding and gives a reviewer nothing new to verify.
The draft is long and the claim count stops rising. You have hit the analysis cap described on the ForgeRank methodology page. Mark the remaining sections by hand rather than trusting the count.
How long does this take?
Budget about 20 minutes per 1,500 words for a first pass, and longer on the first draft you run through it. The marking itself is fast; the slow part is deciding whether a sentence asserts something about the world. After three or four drafts the decision takes a second per sentence, and the pass drops to roughly 10 minutes.
Finish checklist
- Every numeral in the draft sits in a sentence that names its source, or carries a question mark for a current-value check.
- Every company, product, and report name is either sourced or cut.
- Every superlative has a stated comparison set or an explicit selection rule.
- No sentence attaches a real organisation's name to a figure that organisation did not publish.
- The marked list is empty, or every remaining mark has a written reason it stays.
Frequently asked questions
Does a mechanism sentence ever need a source?
No, if it describes a sequence or a cause the reader can reason about without a measurement. A sentence explaining that a migration fails when custom fields do not carry over asserts a mechanism, not a statistic. The moment you attach a failure rate to it, it becomes a figure and needs a source.
What counts as a named entity?
Companies, products, report titles, version numbers, standards, and job titles attached to a claim. Ordinary nouns do not count. The test is whether a reader could look the name up and check what you said about it.
Can I cite a source I have not read?
Not in a draft you intend to publish. A source you have not opened is a source you cannot defend, and the figure attached to it may say something different from what your sentence claims. Open it or cut the sentence.
How do I handle a claim I believe but cannot source?
Rewrite it as a mechanism or a decision rule. A belief stated as a mechanism stays true without a citation; the same belief stated as a percentage becomes a claim you cannot support.
Where does this fit in a publishing workflow?
Run it after the draft is structurally finished and before a fact-check pass. Marking first tells you which sentences deserve a fact-check, which keeps the second pass short.
You can now walk any AI-assisted draft and produce a marked list of unsupported claims in about 20 minutes. The next step is mechanical: open your most recent draft, mark every numeral and proper noun, and see how many sentences survive without a source attached. Do that one draft tonight and the pattern in your own writing will be obvious by the second page.