What citation drift means—and what it does not
A follow-up can add a new fact, narrow a date range or change the question. The answer's source list may change, stay the same or disappear. That movement is worth recording, but it does not by itself tell you whether the answer is right. A stable citation can be irrelevant to a new claim; a new citation can be the better evidence.
In a 2025 workshop paper, Gokul Srinath Seetha Ram calls changes to references across turns citation drift and measures citation-set stability separately from citation fabrication. The distinction is useful: source overlap and claim support are different questions. The paper's abstract says it analyzed 240 conversations, while its methods describe 240 data points as four LLaMA variants multiplied by 12 seed paragraphs and five turns. Treat that count carefully. The paper did not test Perplexity or ChatGPT Search, and it cannot establish how those products behave today.
The study defines citation stability using overlap between citations on adjacent turns and separately defines fabrication and drift measures. Its methods name four LLaMA variants, 12 seed paragraphs, five turns and 240 total data points; the abstract instead describes 240 conversations.
What it does not establish: This is one workshop study of LLaMA variants using a constrained scientific-writing protocol. It did not evaluate Perplexity or ChatGPT Search. Its abstract and methods describe the sample count differently, and citation overlap is not a measure of whether a source supports a claim.
What the products document about citations
Read product documentation as a description of the interface, not as a guarantee that a citation proves the sentence beside it. The same audit method works across tools, but their search modes and source types are not identical.
| Tool | What its public documentation says | What the documentation does not establish |
|---|---|---|
| Perplexity | Its Help Center describes numbered citations linking to original sources and says follow-up questions retain conversation context. | That description does not show that every follow-up launches a new search or that each carried citation supports every new claim. Open the linked page and check the specific passage. |
| ChatGPT Search | OpenAI says ChatGPT may search automatically when current information would help; responses that use Search may include citations and a Sources list. | OpenAI warns that results and citations can be incomplete, outdated or incorrect. A citation's presence is not a support verdict. |
| Kruqond | Kruqond documents public web and GitHub research plus platform-specific social and video tools, including Reddit, TikTok and other social media. Some workflows retrieve available comments from a public post or video URL. | Inputs and comment coverage vary by platform and operation. A retrieved sample is not a complete social-media export, and this guide is not a matched accuracy test of Kruqond against answer engines. |
Perplexity documents numbered links to original sources, contextual memory and the ability to ask follow-up questions.
Perplexity Help Center: How does Perplexity work?
What it does not establish: This describes Perplexity's documented workflow, not an independent audit of citation accuracy or a promise that every follow-up performs fresh retrieval.
OpenAI says Search may run automatically when a question benefits from current information, responses using Search may have citations, and users can open cited sources. It explicitly warns that results and citations can be incomplete, outdated or incorrect.
OpenAI Help Center: Searching the web with ChatGPT
What it does not establish: This is product documentation, not a head-to-head measurement of current ChatGPT Search and Perplexity answers. The page says Search may be used and citations may appear; neither is universal.
Kruqond lists public web and code, social and video discovery, and comment workflows for specified public posts or videos. It warns that supported platforms do not all offer the same operations and comments are only the retrieved sample.
Kruqond source coverage and access limits
What it does not establish: This is Kruqond's own product documentation. It describes supported inputs and limits; it does not independently verify performance or guarantee full coverage of any platform.
Set up a test another person can repeat
Choose a question whose answer depends on identifiable evidence, such as a version-specific product change or a dated policy. Record the exact first prompt, each follow-up, the product mode, visible model, access tier and date. Keep the original answer and source list before asking the next question.
Use the same prompt sequence when comparing products. If a tool has an explicit Search or Research mode, record it. If you cannot tell whether a follow-up searched again, write ‘not visible’ rather than guessing. Do not switch modes halfway through and then attribute the difference to the model alone.
- Write one narrow research question and the factual claims you expect to check.
- Record the product, mode, visible model, plan or access tier, date and exact prompts.
- Save the first answer, citations and linked sources before continuing.
- Use a follow-up that adds one new factual requirement; keep the conversation context intact.
- Repeat the final follow-up as a standalone question in a fresh conversation, using the same mode where possible.
Audit each new claim against the source passage
Split the follow-up answer into short factual claims. Mark whether each one is new, carried forward, narrowed or corrected. For every citation, open the destination, confirm it is the intended page, and copy the passage that supports the exact claim. Record the page's publication or update date and the date you retrieved it.
Use plain labels: supported, partly supported, contradicted, not supported by this source, or unavailable. A source can support one clause but not another. Keep the claim narrow enough for the evidence: a page announcing a feature does not necessarily show who can use it, when it became available or whether it works in every region.
Treat an answer's own explanation of why its citations fit as a lead, not as verification. The decisive check is still the cited material itself. If a source is a search snippet, abstract, repost or metadata-only result, label what you actually saw.
- Assign a short ID to each factual claim in the follow-up answer.
- List every citation marker attached to the claim; leave it blank if there is none.
- Open the source and record the exact supporting or conflicting passage.
- Check dates, versions, populations, geography and other scope limits.
- Label the result and preserve unresolved disagreement instead of averaging it away.
Keep citation stability separate from citation quality
A source list that changes is not automatically a failure. A follow-up may call for a newer or narrower source. A source list that stays fixed is not automatically a success. The old pages may not address what the user just asked. Report the dimensions separately so readers can see what was measured.
If you calculate a stability score, define exactly what counts as a citation and keep it separate from your claim-support labels. A changed bibliography can reflect the new question; only reading the sources tells you whether the answer remains grounded.
| Dimension | Question to ask | What not to infer |
|---|---|---|
| Source-set stability | Which URLs were retained, added or dropped between turns? | URL overlap does not show whether sources are accurate or relevant. |
| Claim support | Does the cited passage support this exact statement? | A real, authoritative page can still be a poor match for the claim. |
| Coverage | Does each material factual claim have an adequate source? | A long source list does not ensure that every claim is covered. |
| Context and freshness | Does the page match the version, date, place and population in the claim? | A source that was once current may no longer answer the question. |
Use a fresh conversation as a control, not a cure
Run the final follow-up as a standalone prompt in a fresh conversation. Compare its claims and sources with the answer inside the longer thread. This can reveal that context, a changed prompt or retrieval selection matters, but it does not identify the cause by itself. If results vary, repeat the same protocol and report the number of runs instead of selecting the most favorable answer.
Reddit discussions offer useful leads, not controlled product evidence. In one thread about follow-up citations, a commenter suggested checking whether the linked source list changed and asking the final question in a new thread. Another discussion about research workflows includes advice to open sources rather than trust the summary. Treat those as user suggestions to test, not proof that a specific product has a defect or that a fresh thread fixes it.
One Reddit poster reports that citations sometimes seemed less well matched after two or three follow-ups; a commenter recommends opening the follow-up sources and comparing the standalone question in a fresh thread.
Read the Reddit post and the fresh-thread comparison comment
What it does not establish: This is an anecdotal report and a commenter’s proposed test, not an independently reproduced result or evidence of the product's internal retrieval behavior. The post listed 11 comments; the fetcher returned five and marked comment coverage incomplete.
In a separate discussion, commenters advise opening the linked sources and checking them before relying on a research summary.
Read a comment in a Reddit discussion about verifying Perplexity research
What it does not establish: This is one participant's advice, not a survey, tested workflow or representative measure of user opinion. It is included as an example of community practice, not proof of product performance.
Compare products on the same task—and disclose the limits
Give each product the same research question, follow-up sequence and verification standard. Compare what each one actually returned: claim support, source relevance, coverage, freshness, access limits and the effort required to verify. Record product mode and date because interfaces and models change. Do not turn a documentation summary or one example into a winner ranking.
Perplexity and ChatGPT Search are answer-search workflows. Kruqond also supports research across public web pages and platform-specific social sources, including Reddit and TikTok; some tools retrieve available comments from selected posts or videos. Those jobs overlap but are not identical. If a study needs comments or video discussion, compare those source workflows directly and disclose how many posts, comments or transcripts were actually available.
A good report shows the prompts, sample, source links, support labels and missing evidence. If an answer cites an inaccessible page or a snippet rather than a passage you opened, say so. Readers can then decide whether the result fits their own research rather than relying on an unexplained score.
Scope and disclosure
Kruqond publishes this guide and is one of the products named in it. The comparison summarizes the linked public product documentation and selected public Reddit comments. No matched test of current Perplexity, ChatGPT Search or Kruqond citation accuracy was run. The Reddit retrieval was incomplete and anecdotal. The guide was prepared with AI assistance; sources were checked on November 17, 2026. Product features and source access can change, so verify the current documentation and inspect the original pages before relying on a result.