A source appearing in an AI answer that mentions your brand is not necessarily saying anything about your brand, and treating the two as the same thing manufactures evidence that does not exist.
The case that made this concrete
In one of our own scans, Stanford’s AI Index appeared in the source panel, marked independent, with a confidence figure of 0.70.
The sentence it came from read: “University research on LLM citation bias (Stanford’s AI Index, Oxford Internet Institute studies).”
Stanford was in the answer. Stanford said nothing whatsoever about the brand. It had been cited to support a general claim about how language models select and cite sources.
The system had counted a source that appeared in the answer as a source that corroborated the brand. Those are different things, and the gap between them is where a lot of AI visibility reporting quietly inflates.
Three relationships a citation can have
When ChatGPT, Claude, Gemini or Perplexity cites a source in an answer that also mentions your brand, that source is doing one of several jobs.
Brand corroboration. The source is cited as evidence about you. “5.0/5 on G2” is G2 being used as evidence about a specific company. This is the only category that should count toward an evidence score.
Context or category evidence. The source supports a claim about the topic, the category, or a method. A research index cited to explain how AI citation behavior works falls here. It is real, it is worth recording, and it says nothing about you.
Competitor evidence. The source substantiates a claim about a different vendor in the same answer. Also real, also not yours.
A report that shows all three under one heading called “independent sources” is not lying about what appeared. It is wrong about what any of it means.
How to classify one yourself
The test is simpler than it sounds and you can run it manually on any report.
Find the sentence the citation came from. Not the source name in a list. The actual claim in the response.
Check whether your brand is named in that sentence, or in the one immediately before it. If the source name and the brand name never appear together in a single claim, that source is not corroborating you.
That single check reclassified our Stanford entry immediately, and it does not require any tooling.
The failure mode underneath the failure mode
There is a second, worse case worth separating out.
Stanford’s 2026 AI Index has nine chapters: Research and Development, Technical Performance, Responsible AI, Economy, Science, Medicine, Education, Policy and Governance, and Public Opinion. None of them is a study of language model citation bias.
Which means this may not be a real source cited for the wrong subject. It may be a claim attached to a source that never made it.
Those need different handling. A correctly attributed source cited about your category should be recorded and excluded from your evidence count. A source credited with a finding it never produced should be flagged, because it tells you something about the reliability of that entire answer.
The tell is usually a specific claim attributed to a named source with no link attached. That is not proof of anything, and it is worth a second look.
Why confidence figures belong on the row
In that scan, the marketplace citations came back at confidence 0.85. The Stanford one came back at 0.70.
The classifier already sensed something was different. That signal was sitting in a detail row nobody looks at, underneath a number in a panel everybody looks at.
If your tool reports per-citation confidence, surface it. The earliest warning that a source has been miscategorized is usually a confidence figure that does not match its neighbors.
What to ask of any report
Three questions, and they work on any vendor:
Does this report distinguish between sources that corroborate the brand and sources that merely appeared in the answer?
Can it return “unclassified” for a citation it cannot categorize, or does every citation get assigned to something?
Does it show me the quoted sentence, or only the source name?
What this costs when it goes unchecked
A miscounted source is not just an inflated number. It changes what you do next.
A brand reading four independent sources concludes it has reasonable corroboration and moves budget to some other problem. The same brand reading two, with one of those two being a source it does not control, would prioritize differently. The report did not just describe the position inaccurately. It redirected the work.
The same failure appears in reverse when a genuinely useful contextual source gets buried. A research organization appearing repeatedly in answers about your category tells you something real about which sources ChatGPT, Claude, Gemini and Perplexity reach for in that space. That is worth knowing. It is just not evidence about you.
A table that cannot say “I do not know” will assign every case to whichever category it superficially resembles. That is not measurement. That is a system inventing corroboration to fill a column.
FAQ
Why would an AI engine cite a source that says nothing about my brand?
Because most AI answers make several claims, and different sources support different claims. A response about your category might cite a research organization for a general point about the market and a review site for a point about you, in the same answer.
Does a contextual citation help my brand at all?
Not as corroboration. It tells you something about the shape of the answer your brand appeared in, which can be useful diagnostically, but it is not a party vouching for you and should not raise an evidence score.
How do I check whether a source really made the claim attributed to it?
Look for a link in the citation. If there is none and the claim is specific, go to the source directly and search for the finding. Named sources with no URL attached to a specific claim are the ones most worth verifying.
Do ChatGPT, Claude, Gemini and Perplexity differ in how they attribute sources?
Yes. Perplexity and Google’s AI Overviews mark citations explicitly, while ChatGPT without browsing enabled often does not, which means attribution there is inferred rather than stated. That difference affects how much you can verify.
Should a miscategorized source be deleted from the record?
No. Record it and reclassify it. A source that appears repeatedly in answers about your category is useful information about the retrieval landscape even when it is not evidence about you.
What confidence figure is low enough to investigate?
There is no universal threshold, and any tool quoting one should explain how it was set. What is more useful is relative: a citation scoring noticeably lower than others in the same scan is worth reading manually.
ABOUT AXIS SUITE
Axis Suite by TrendAxis is the independent intelligence layer that explains what AI believes about your brand, why it believes it, and what decision that belief ultimately drives.
Most tools count citations. Axis Suite tells you which of those citations are actually evidence about you, how many separate parties they represent, and what would remain if your largest source disappeared.
Run a scan at axissuite.ai or read the methodology in the Proof Center.