The first check is the cheapest: does the sentence assert something about the outside world? If yes, it needs a source attached inline. ForgeRank's published dataset (https://forgerankai.com/static/data/ai-content-quality-29-drafts.json) scored 29 drafts and logged 2,839 claims, of which 51 were flagged as needing a source and 38 arrived with none attached. That gap is the whole problem in one number.
This is an evidence checklist, scoped to attribution and verifiability. It is not the general quality checklist published earlier. Every item here is a decision about a single sentence.
How were these picked?
- Each check had to address one of two questions: does this claim need a source, and is the attached source acceptable.
- Checks that only improve prose, structure, or readability were excluded.
- Every check maps to a figure or rule already documented in the ForgeRank published dataset, the ForgeRank methodology page, or Google Search Central.
- Any check that could not be tied to a concrete instance in the 29-draft dataset was cut.
Quick comparison
| Check | Best for | Standout strength | Main limitation |
|---|---|---|---|
| External-world test | Sorting claims from opinions | Catches unsourced assertions fast | Slow on dense technical prose |
| Inline attribution | Published drafts | Reader verifies without leaving page | Clutters short-form copy |
| Source cap awareness | Long drafts | Explains missing citations | Cap hides real gaps |
| Dimension tier read | Comparing drafts | Isolates the weak dimension | Needs full scoring run |
| Delete-the-claim rule | Unsourceable assertions | Removes liability entirely | Costs you a good sentence |
| Scaled-abuse screen | Bulk publishing | Matches Google's stated target | Threshold is judgment-based |
| Core-update alignment | Site-level planning | Ties to a stated 40 percent goal | Site-wide, not per-sentence |
| Verbatim quote discipline | Regulatory figures | Prevents paraphrase drift | Slows drafting |
| Distribution check | Pre-publish gating | Shows where drafts cluster | No single-draft score pair exists |
1. The external-world test — best for sorting claims from opinions
Run this first on every draft. A sentence that asserts something about the world outside your own head needs a source attached inline. A sentence that states your position, your process, or your recommendation does not.
The mechanism matters because the scorer's flag rate depends on how many claims survive this filter. ForgeRank's published dataset recorded 2,839 claims across 29 drafts, with 51 judged to need a source. That ratio tells you the filter is doing real work: most sentences in a typical draft are claims about the world, and each one is a citation slot.
The consequence for the writer is uncomfortable. You cannot fix an unsourced claim by rewording it into a softer assertion. Softening produces the vague qualifiers that read as filler. Either attach the source or delete the sentence.
The boundary: opinion, method description, and direct reader instruction sit outside this test. "Open Search Console today" needs no citation because it asserts nothing about the world.
2. Inline attribution — best for published drafts
Attach the source name to the figure in the same sentence. Not a footnote, not a reference list at the bottom. The reader and the extractor both need the attribution adjacent to the number.
The mechanism is extraction. A paragraph lifted out of your article and placed in an answer surface keeps only what travels with it. A number that depends on a footnote three screens down arrives stripped of its source, which makes it an unsourced claim in that new context.
ForgeRank's published dataset puts a number on the failure mode: of 51 claims judged to need a source, 38 had none attached. That is roughly three in four. If your draft matches that rate, you are shipping unsourced assertions at scale.
The consequence is that attribution becomes a drafting habit, not an editing pass. Write the source name as you write the number.
The boundary: a figure that changes often, such as pricing or plan limits, needs the source name plus a note to confirm before acting. Stale attribution is its own failure.
3. Source cap awareness — best for long drafts
The source analysis behind the ForgeRank published dataset runs on a capped number of claims per draft, and 13 of the 29 drafts hit that cap, according to the ForgeRank methodology page (https://forgerankai.com/blog/ai-content-quality-data). If your draft is long, the cap explains why some claims show no citation flag.
The mechanism: a capped scan stops counting at a fixed point, so the claims past that point go unexamined. A draft that looks clean may simply have run out of scan budget before reaching its weakest paragraph.
The consequence for the writer is that a low flagged-claim count is not proof of clean sourcing. It is proof that the scan reached the end of its budget. Check the tail of long drafts by hand.
The boundary: this applies to drafts long enough to hit the cap. Short drafts below the cap get full coverage, so the flag count is meaningful there.
4. The dimension tier read — best for comparing drafts
Every score in the ForgeRank published dataset resolves into four dimension tiers, and reading them tells you which dimension dragged the draft down. The pattern is mechanical: all ten 4.0 drafts scored 4.0 across all four dimensions, and all ten 4.8 drafts scored 4.0/4.0/7.0/4.0.
The mechanism is that each step in the distribution is a single dimension moving. A draft at 4.8 versus 4.0 differs on one tier, not four. That means the fix is narrow: find the dimension that moved and repair that one.
The consequence is that you stop rewriting whole drafts. You target the tier.
The boundary: this reading works when you have a full four-dimension score. A single overall number tells you nothing about which dimension to fix.
5. The delete-the-claim rule — best for unsourceable assertions
Some claims cannot be sourced, and the correct fix is deletion. Not rewording, not hedging. Cut the sentence.
The mechanism: an unsourceable claim is one where no acceptable source exists, because the claim was never true in the first place. Rewording it preserves the assertion and hides the absence of a source. Deletion removes the liability along with the sentence.
The consequence for the writer is real cost. You lose a sentence you liked. In exchange, you stop publishing assertions you cannot defend.
The boundary: a claim you know to be true but cannot cite belongs in the delete pile too, unless you can point to a checkable public fact. "Everyone knows this" is not a source.
6. The scaled-abuse screen — best for bulk publishing
Google's spam policies name scaled content abuse — mass-producing many pages to manipulate rankings — as the target, regardless of whether AI or humans wrote them, according to Google Search Central. Run this screen at the site level, not the sentence level.
The mechanism: the policy targets volume produced for ranking manipulation, so the trigger is the pattern across pages, not the authorship of any one page. A hundred thin pages on adjacent queries reads as scaled abuse even if every page is clean.
The consequence: your evidence checklist cannot save a site-level volume problem. Per-sentence attribution and per-site publishing rate are two separate controls, and you need both.
The boundary: AI authorship alone is not a violation under the stated policy. The screen targets mass production aimed at rankings.
7. Core-update alignment — best for site-level planning
Google Search Central states that the March 2024 core update aimed to reduce low-quality, unoriginal content in search results by 40 percent. Treat that as the site-level goal your evidence practice serves.
The mechanism: an update targeting unoriginal content rewards pages that add something the existing results do not have. A sourced figure from your own published dataset is original content. A restatement of a figure everyone already cites is not.
The consequence for the writer is that attribution alone does not clear the bar. A draft can be fully sourced and still be unoriginal if every source is someone else's.
The boundary: the 40 percent figure describes the update's stated aim, not a measured outcome for your site. Do not read it as a guaranteed reduction.
8. Verbatim quote discipline — best for regulatory figures
Regulatory and policy figures get quoted exactly, with the source named. Google Search Central's stated aim for the March 2024 core update is to reduce low-quality, unoriginal content in search results by 40 percent — quote that, do not round it.
The mechanism: paraphrase drift changes figures while keeping the source name attached, which produces an attributed number the source never published. That is worse than no citation, because it misrepresents a named institution.
The consequence is a small drafting tax. Copy the figure, paste it, keep the source adjacent.
The boundary: this applies to figures and policy language. Ordinary factual statements can be rewritten in your own words as long as the figure and its source stay intact.
9. The distribution check — best for pre-publish gating
Before publishing, check where your draft sits against the known distribution. 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, per the ForgeRank published dataset.
The mechanism: the gap between 4.8 and 7.0 is empty in this dataset, which means drafts cluster at the low end until a dimension changes. A draft sitting at 4.8 needs a dimension fixed, not polish.
The consequence: you get a gate. If your draft scores at or below 4.8, do not publish it and hope. Find the dimension.
The boundary: the dataset contains no before-and-after score pair for any single draft. You cannot infer how much a given fix raises a score.
Two beliefs about AI content evidence, and what the dataset says
The common belief is that AI-assisted drafts fail because of how they are written — tone, rhythm, structure. The belief the evidence supports is narrower: they fail on attribution. The ForgeRank published dataset shows 38 of 51 source-flagged claims arrived with no source attached, and 20 of 29 drafts scored at or below 4.8 before publishing.
Resolve it against the data. The distribution is 4.0 x10, 4.8 x10, 7.0 x3, 7.2 x4, 7.8 x1, and 8.5 x1. The jump from 4.8 to 7.0 happens when a dimension moves, and the dimension tiers show each step is one dimension changing. Tone is not the dimension that moves.
Which AI content evidence check should you run first?
- If your draft is under the source-analysis cap, run the external-world test first — full coverage makes the flag count meaningful.
- If your draft is long and hit the cap, hand-check the tail before trusting a clean flag count.
- If a claim has no acceptable source, delete it rather than reword it.
- If your site publishes at volume, run the scaled-abuse screen before any per-sentence work.
- If your draft scores at or below 4.8, read the dimension tiers and fix the one that moved.
Is one sourced claim enough to pass?
No. One sourced claim does not clear a draft. ForgeRank's published dataset logged 2,839 claims across 29 drafts, and 51 of those needed a source. A draft with one citation and forty unsourced assertions fails the same way a draft with zero citations does. The bar is proportional: every claim that asserts something about the outside world carries its own source, and the count of sourced claims should track the count of external-world assertions. A single citation is a start, not a pass.
Frequently asked questions
What counts as a claim that needs a source?
Any sentence asserting something about the world outside your draft: a statistic, a date, a price, a policy, a study result. Opinions, method descriptions, and direct instructions to the reader do not need one. The test is whether a reader could ask "says who?" and expect an answer.
Can I cite my own published dataset?
Yes, and it is the strongest source type available. ForgeRank's published dataset at https://forgerankai.com/static/data/ai-content-quality-29-drafts.json is cited throughout this piece for exactly that reason. First-party data you published is verifiable and original, which serves both the attribution check and the core-update goal.
What if I cannot find a source for a true claim?
Delete the sentence. An unsourceable claim you believe is true still fails the check, because the reader cannot verify it and the extractor cannot carry it. Rewording preserves the assertion and hides the missing source, which is worse than cutting it.
Does the source cap mean my long draft is clean?
No. The ForgeRank methodology page notes that 13 of 29 drafts hit the cap, which means the scan stopped before reaching every claim. A clean flag count on a long draft may only mean the scan ran out of budget. Check the final third by hand.
Why does the March 2024 core update matter to an evidence checklist?
Google Search Central states the update aimed to reduce low-quality, unoriginal content in search results by 40 percent. Sourced, first-party figures are original content. A fully sourced draft that only restates other people's numbers still reads as unoriginal against that stated aim.
Start with one sentence
The evidence checklist comes down to one habit you can start tonight: open your current draft, find the first sentence that asserts something about the world, and check whether a source name sits in that same sentence. If it does not, you have three options — attach the source, delete the claim, or move it to the delete pile. Do that for the next ten claims and you will know your draft's real sourcing rate before you publish it.