A citation is where the investigation begins.

A cited answer can still give a brand an unfair comparison. The linked page may concern an older product, support only part of the sentence, or repeat an assertion made elsewhere. What a communications team needs is a documented judgment about what the evidence establishes and what response follows. A citation count can point to material worth investigating, though it is too weak to support a correction request or a new positioning claim.

Examine the assertion attached to the link

Start by separating the propositions in the answer. “Company A serves large retailers and offers the fastest deployment” contains a customer claim and a comparative performance claim. A case study about one retail customer could support the first, within limits, and say nothing about the second. The example is hypothetical.

The 2023 paper Evaluating Verifiability in Generative Search Engines assessed whether statements had supporting citations and whether each citation supported its statement. Its human evaluation covered Bing Chat, NeevaAI, perplexity.ai, and YouChat as they worked at the time. Its historical results do not give a current error rate for AI search, but the two verification questions are still useful checks.

The 2023 ALCE research also evaluated fluency, correctness, and citation quality separately. Its benchmark concerned systems that generate answers from retrieved material. It gives a methodological reason to check each dimension on its own, since readable, well-linked prose may not satisfy all three.

A source can be quoted accurately while the surrounding conclusion overreaches. The investigator should keep the relevant passage and mark exactly where the answer adds an interpretation, comparison, or qualification that the source does not establish.

Check the product and the period

Profound’s WHOOP customer account describes AI answers that compared recent competing wearables with older WHOOP generations and used outdated specifications. According to Profound, WHOOP used its FactCheck capability to identify inaccurate claims and route them for review and correction. As a vendor-published account, it does not independently test the products or study outcomes under controlled conditions.

Sentiment alone cannot catch this: an answer can praise both devices while comparing the wrong editions. The useful check covers the named product, generation, specification, and date. A historical review may be entirely accurate and still be the wrong evidence for a current purchasing comparison.

Beyond products, a subsidiary’s operating history should not silently become the parent company’s history. A founder’s past affiliation should not be presented as a current role. Before approaching a publisher, establish whether the error sits in the original material, the model’s attribution, or the comparison the question invited.

As discussed in The question decides who you are compared with, the original question belongs in the record. A query explicitly about a previous generation needs a different assessment from one about today’s alternatives.

Ask what kind of evidence the source supplies

A company’s documentation is a relevant source for its stated specifications, policies, and offer. It is weaker evidence that its implementation is easier than every competitor’s. A customer review describes an experience. Without more evidence, it does not show how common that experience is. An independent comparative test may address performance directly, but only under its stated conditions.

These distinctions set how much weight a passage can bear. The analysis should identify the author, the basis of their knowledge, the product or population examined, and any material limitations. A well-known publisher does not make every inference drawn from its page valid, and an unfamiliar one is not automatically wrong.

When sources conflict, record the disagreement before choosing a conclusion. Two prices may refer to different billing terms. Two product descriptions may concern different markets. If the conflict survives those checks, the recommendation may need to commission verification before anyone publishes a stronger claim.

Separate repeated publication from corroboration

Suppose a manufacturer issues a release claiming improved battery life. Three trade sites reproduce the announcement, and an AI answer cites all three. This hypothetical answer has several cited pages but may still rest on one assertion from one interested party.

The pages show that the claim circulated. Where links, attribution, or matching passages reveal a shared origin, record that relationship in the report. Where the relationship is only suspected, label it as an inference.

In competitive analysis, a rival’s claim, a journalist’s assessment, and an AI-generated description of the rival have different authors. Treating them as interchangeable can lead a team to accuse a competitor of publishing language that appeared only in a model response. Tracking authorship prevents that error and makes any proposed outreach easier to defend.

A visible source trail has limits

A citation shows an observable connection between the answer and a page. It does not reveal every input used to generate the answer or prove that one passage produced the whole account. An answer may include uncited claims, and a linked source may have changed since it was collected.

Keep enough context for someone else to review the finding, including the answer, question, collection date, cited URL, relevant passage, and source date or version where available. Missing support should stay an open finding. Do not fill the gap with a plausible story about hidden model reasoning.

An uncited statement is also not automatically false. Material the answer did not expose may support it. Further investigation can find that evidence, but the report should keep newly found support separate from the citation trail actually observed.

Turn the review into a specific assignment

An outdated owned page may need revision and clearer version labeling. An incorrect third-party specification may justify a documented correction request. A supported limitation may require a narrower message. A comparative claim without adequate evidence may need testing before it can responsibly be promoted.

For each significant finding, the brief should identify the assertion, its subject, the passage examined, the supported conclusion, and the unresolved point. It should then name the proposed work and who can carry it out. That record lets a client assess the reasoning without repeating the investigation.