Method
Verifying a source: support and relevance
A source is only usable once you can point to the exact passage that backs your sentence and answers the same question your reader has. Here is a four-step test, with a worked example that fails.
You are writing an article and the source in front of you looks authoritative. How do you verify a source before you use it? The test is not whether the source exists. It is whether the passage you retrieved actually supports the sentence you want to write, and whether it is relevant to the reader's question. Take the sentence "AI visibility is measured only through APIs." Find the passage in the source and ask yourself: does that passage talk about measurement or about access? If it talks about access, the sentence is not supported. Rewrite it or drop the source.
Four steps with checkable criteria
RankAI offers this method as a working proposal, not as a universal standard. Each step has a criterion you can check without relying on memory.
- Isolate the passage, not the source. Criterion: you have a verbatim quote or a line range, not a link to the homepage. If you cannot point to the exact spot, the source is not verified yet.
- Support test. Criterion: replace your sentence with the passage and check whether the claim still holds without additions of your own. If it needs an "in practice" or a "that is" to make sense, the support is missing.
- Relevance test. Criterion: the passage answers the same question your reader has, not a nearby one. A source about API access is not relevant to a sentence about how AI visibility is measured.
- Check the chain of observation. Criterion: ask whether the passage describes what a person sees or what an API returns. What comes back from an engine through an API is the API response, not the screen a person sees; Gemini and Google AI Overviews are not the same observation. This matters when you verify an outside source, because a quote that blurs the two levels can look supported while describing something else.
Worked example: a test that fails
Sentence in the article: "AI visibility is measured by querying engine APIs." Passage found in the source: "APIs return the engine's response, not the screen a person sees." Support test: the passage says what the API returns, not that visibility is measured that way; the sentence adds a conclusion the passage does not contain. Result: insufficient support. Relevance test: the passage is about access, the sentence is about measurement. Result: not relevant. What to do when it fails: rewrite the sentence as "what comes back from an engine through an API is the API response, not the screen a person sees," or look for a passage that explicitly discusses measurement. If neither route is available, cut the sentence.
What separates this from a generic source check is the gap between a retrieved source and a cited source. A retrieved source is a document you found and read. A cited source is a passage you have pinpointed and put through both tests. Classic fact-checking tends to stop at the reputation of the source or at whether the claim appears somewhere in the document. The bar here is higher: a retrieved source is not enough, you need the cited passage. One useful edge case is a source that contains the right sentence in a section about something else. It passes the support test but fails the relevance test, and for that sentence it is unusable. A second edge case is a source that describes an API and gets used to talk about what a person sees: the passage is true, but it observes a different level and does not support the sentence. In neither case does the method say the source is false. It says the source is wrong for that sentence. Apply this before you write, not after.
A checkable limit
The method does not establish whether a source is true in absolute terms. It establishes only whether a passage supports and concerns the sentence you are writing. If the passage cannot be pinpointed, the test cannot be completed and the source should be treated as unverified. When a visibility figure is unavailable, the product shows it as N/A, never as zero. For the full picture of how RankAI handles missing measurements, see the in-depth page.