TL;DR
- An AI citation gives you a source to inspect. The link alone does not prove that the source supports the whole claim.
- Check two things: whether the linked text supports the sentence, and whether that text is reliable evidence.
- Research has measured both weak claim support and citations to fabricated content. Those studies test different risks.
- Keep the exact claim, source passage, scope and review decision in one record before counting a citation as useful.
Look beyond the source link
This matters when a cited answer names your product for a use case you do not support. It also matters when an answer about a rival looks persuasive because it has several links. A count of links cannot settle either question.
Two Measures That a Link Count Misses
Liu, Zhang and Liang's 2023 study found average citation recall of 51.5%: the share of checked statements fully supported by their citations. Average citation precision was 74.5%: the share of citations supporting the attached statement. The team tested four search systems on 1,450 queries each, with responses collected in February–March 2023.
| Measure in the 2023 study | Average result | What was checked |
|---|---|---|
| Citation recall | 51.5% | Whether statements needing verification were fully supported by their citations |
| Citation precision | 74.5% | Whether each citation supported its attached statement |
Source: Evaluating Verifiability in Generative Search Engines, Tables 5–6. Averages across tested systems and query groups; different denominators. These are historical results, not current product error rates or stages in a funnel.
A reply could have several sound links yet leave a key claim without support. It could also cite a page that supports only half of a sentence. The two measures help expose those different gaps. They should not be added, subtracted or turned into a single “trust score” for your brand.
The study used Bing Chat, NeevaAI, perplexity.ai and YouChat as they worked in early 2023. It did not test today's versions. Its lasting value for a content team is the distinction between covering the claims and checking the links, rather than a current ranking of tools.
Historical February–March 2023 responses from four systems. Averages across tested systems and query groups; these are not current product rates or stages in a funnel. Liu, Zhang and Liang, 2023.
A Source Can Support a Sentence and Still Be Wrong
A newer October 2026 paper by Qi Liu and colleagues reports that 8 of 10 tested AI-search platforms cited a fabricated concept within seven days of its publication. The researchers used controlled web publications with traceable markers. The result shows why a citation cannot serve as an independent fact check of the source.
| Research and direct source | Method and scope | Actual result | Implication for a B2B review |
|---|---|---|---|
| Liu, Zhang and Liang, 2023 | Human review; four systems; 1,450 queries per system | 51.5% average claim coverage; 74.5% average citation precision | Compare each material claim with the linked text |
| Qi Liu and colleagues, October 2026 | Controlled publication of a fabricated concept; ten platforms; seven-day window | Eight platforms cited the fabricated concept | Inspect the source's own evidence and origin as well |
The newer paper also reports a separate mapping of 17,211 citation instances across 6,356 source domains. That March 2026 mapping is not the denominator for the fabricated-concept result. The authors state that their designed query mix does not represent real traffic. Controlled topics also cannot tell a B2B team how often ordinary buyer questions receive false answers.
These studies complement each other. The first asks whether an answer stays faithful to linked material. The second shows that published material can enter citations despite being fabricated. A faithful summary of a weak source can still mislead a buyer.
For a brand, the practical response is to preserve the chain from answer to source to underlying proof. More pages that repeat a claim may share one weak origin. A familiar domain name also does not identify who wrote a specific page or how its claim was checked.
A Review Record That Leads to a Decision
Start with claims that could change a buyer's decision: product limits, supported systems, measured results, service scope or the cost of an option. Save the full answer and its date first. Then split compound claims so each part can be checked against the cited text.
The following record is an illustrative example for a fictional software supplier. It is an editorial tool, not a client result.
| Field | Completed example | Review consequence |
|---|---|---|
| Exact answer claim | “The connector works with every ERP and needs no custom code.” | Split coverage from setup effort |
| Cited source passage | “Standard mappings cover the listed ERP versions. Custom fields may need code.” | The page supports a narrower claim |
| Underlying record | Current supported-version list and setup guide; product owner confirms both | Check version scope and the custom-field exception |
| Support judgment | Partial support; “every” and “no custom code” exceed the source | Do not log the original sentence as a supported product claim |
| Action and owner | Content lead checks whether the source page could be clearer; product owner approves precise limits | Update ambiguous source wording if needed; retain the answer and review date |
The team has learned more than “we were cited.” It can now see how the answer stretched the source and which facts need to be clearer. If the source already states the limit plainly, record that the error is in the generated answer. A website edit cannot force an external system to correct it.
Use simple review outcomes: full support, partial support, no support, or unresolved. Keep source quality as a separate note. A page behind a login, a missing document or a changed passage belongs in unresolved until you can inspect it. Do not quietly count it as either correct or false.
For repeated reports, retain the checked sample and its selection rule. “Full support in the reviewed answers” is more precise than “our AI accuracy score.” Our AI brand accuracy article addresses the wider task of comparing brand claims with approved business facts. This record focuses on what a specific citation can establish.
What Should Change in Your GEO Report?
Keep citation presence and citation quality visible side by side. For each useful buyer question, record whether the brand appeared, which URL was linked, what claim the link was meant to support and what the source review found. That lets a team tell broader exposure from stronger evidence.
Also keep the unresolved cases in view. Excluding hard-to-check claims without noting them can make a report look better while making it less useful. Choose the sample before reviewing outcomes, and use the same rule in later checks. For the broader measurement plan, see how to measure AI visibility.
AI Citation Support: Common Questions
Is a broken link proof that the answer is false?
No. A broken link blocks that route to verification. The claim may be true, but the citation no longer lets you check it. Look for the original record and mark the case unresolved until the evidence is available.
Can several citations jointly support one sentence?
Yes. Different sources can support different parts of a claim. Check each part, then check whether the combined evidence justifies the full wording. A set of relevant links still may not support a stronger claim such as “always” or “best.”
Should a team check every citation manually?
Prioritize claims that could change a buying decision. For a large report, define a sample before you see the results and disclose its scope. Automated tools can help flag missing or mismatched text; a final judgment still needs to account for meaning and context.
Does correcting our page guarantee a better AI answer?
No. A clearer page gives readers a better source and removes ambiguity you control. Save the change and recheck the relevant question later. Report the observed answer separately from the website work, without claiming that one forced the other.
Review the claims behind the links
If your visibility report contains citations that your team cannot tie to clear product evidence, bring the saved answers and source pages to SemShift's AI visibility audit. We can discuss a review of material claims, source gaps and content priorities before those citations become sales claims.