Reviews, Reputation and External Evidence for AI Search
How AI search uses outside sources When buyers ask an AI search product to compare businesses or recommend a product, information beyond…
When an AI answer includes a page in its sources panel, the page may discuss the topic without supporting the specific statement you are checking. A source panel is a list of links associated with an answer; a citation is a link attached to some part of the answer. Neither guarantees that the page supports every detail.
That distinction matters when you are checking a number, date, recommendation, or claim about what caused an event. As of November 11, 2026, OpenAI says ChatGPT’s Sources panel can include both cited sources and other relevant links. To check whether a page supports a claim, open it and compare its evidence with the answer’s exact wording.
A source-panel entry tells you that ChatGPT surfaced a page in connection with its answer. An inline citation, when present, points to a source associated with a particular passage. OpenAI’s ChatGPT web search help describes both cited sources and other relevant links in the panel.
So a page can appear in the panel without clearly supporting the sentence you are checking. Even an inline citation gives you a starting point for checking, and you still need to verify whether the linked page supports each nearby claim. OpenAI’s web search API documentation also distinguishes inline citations from a broader set of returned sources. The documentation describes the API, leaving open whether ChatGPT’s consumer interface works identically. It shows that finding or returning a page and attaching it to a particular passage are separate actions.
Other AI search products present sources differently. Google says AI Overviews and AI Mode can run multiple related searches and show links associated with a response. That helps explain why a group of links need not function as a sentence-by-sentence bibliography, but Google’s description does not reveal how ChatGPT handled a particular answer.
A page may cover the same product, person, study, or event without establishing the answer’s exact claim. For example, a page about a study could identify its authors and topic but say nothing about the result an AI answer attributes to it. The page is relevant; the particular claim remains unverified.
The page might describe one location, time period, or group, while the answer makes a broader statement. Or the page may support one part of a sentence but not the rest. If a sentence combines a date, a statistic, and a conclusion, check each part separately.
An answer can also combine information from several pages. One cited page might support a detail while another supports a different detail. The final summary could still go beyond what any of the pages establishes. Public documentation does not show exactly which passages shaped each sentence in a particular answer, so the panel alone cannot show how the answer was put together.
Finally, a page may be out of date, unclear, or weakly sourced. It might repeat a claim without linking to the underlying data, or use a definition that differs from the answer’s. If a page loads, you can reach it. That does not prove that an AI service visited it, stored it in its search index, ranked it for the answer, cited it, or relied on it.
Start with the exact statement you doubt. If it contains several factual parts, separate them. Then open the closest inline citation, if the answer has one. If you are checking ChatGPT, treat other panel links as additional leads because the panel can include relevant links that are not citations for that statement.
Read the page itself rather than relying on its title, search snippet, or the AI’s summary of it. Search within the page for a key phrase, name, or number, then read the surrounding text. Ask whether the page supports the same people or subject, place, time period, definitions, and level of certainty as the answer.
Check the page’s publication or update date and the evidence behind its claim. For an important decision, look for the original document or data, or other strong evidence. If the page supports only part of the answer, or another credible source disagrees, treat the claim as unresolved until you check further.
You can ask the AI to identify the passage behind each claim to help locate relevant text. But the AI’s explanation of its own sources is not independent verification. Confirm the passage on the page.
Research has measured gaps between citations and the claims they accompany, though its results apply to the systems and tasks tested. A 2023 audit of four generative search products found that 74.5% of citations supported the sentence associated with them, while 51.5% of generated sentences were fully supported by their citations. Those figures measure different things: whether a citation supported its sentence, and whether a sentence was fully supported. These 2023 results describe the audited products at the time; they do not give current accuracy rates for ChatGPT or every AI search product. See Evaluating Verifiability in Generative Search Engines.
A separate evaluation of medical references found that the share of responses not fully supported by their listed sources ranged from about 50% to 90%, depending on the model and test conditions. The study examined medical questions and specific systems, so those results should not be generalized to all AI answers. They do show why readers should check source support carefully when claims carry risk. The authors describe their methods in How well do LLMs cite relevant medical references?.
Citation quality also varies by task and system. A 2026 study of OpenScholar, a system designed to answer scientific-literature questions, reported citation accuracy comparable to human experts in its evaluation. That result does not establish the reliability of consumer chat products, but it cautions against treating every AI citation as equally weak. The relevant question is whether this page supports this claim under these conditions. See the study on synthesizing scientific literature with retrieval-augmented language models.
A source panel helps you find pages associated with an answer. It does not tell you, by itself, whether each page supports the answer, which passage supports which claim, or whether the answer represents the page accurately. The interface also varies by product, and product behavior can change.
A claim may be correct even when the displayed citation is weak, because the answer could be supported by other evidence. Even a strong source does not guarantee that the AI summarized it correctly. Check the claim and the page independently.
For medical, legal, financial, and safety decisions, rely on evidence appropriate to the subject and seek qualified professional advice when needed. An AI citation is a starting point for verification, not a substitute for it.
Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.
Take a look at vroni.com