如何核查 Perplexity 引用：16 題主張驗證流程

Can You Trust Perplexity’s Citations? A 16-Claim Verification Protocol

A citation badge is a route to evidence, not evidence that the answer used it correctly. A search assistant can point to a real government page and still reverse a trend, carry forward an old year, merge two populations, or attach the page to a sentence it does not support. The useful decision is therefore not “Does this answer have citations?” but “Can a reviewer trace each consequential claim to the exact passage, date, population, and unit that support it?”

If you need a product-level introduction before testing a workflow, Trust Vault’s guide to Perplexity AI explains its search and answer modes. This article adds an operational test: a 16-prompt, claim-by-claim audit that a team can run before an AI-search summary is copied into research, a customer response, or an internal decision. The test below is a reusable protocol, not a claim that we ran Perplexity on these prompts. No live model output was collected for this article.

The decision in one minute

Use a citation-backed AI answer as a research lead when a person can open the sources, check the relevant passages, and correct or reject the summary before it affects someone. Do not treat the citation list as a verification result. For a high-impact claim, require a primary source, a matching time period and population, and a human record of the passage that supports the sentence. If any one of those is missing, the claim remains unverified even when the answer looks polished.

The distinction is visible in current evidence. Perplexity’s own help material says its search answers include direct source links and tells readers to validate information by referencing those linked sources. That is a product description and a user instruction, not an independent accuracy guarantee. The Perplexity help page on Pro Search describes source links and recommends validation. A link makes inspection possible; it does not complete the inspection.

A recent company-run study illustrates why that distinction matters. Searchable reports that it asked four AI systems 106 questions about local US political issues, repeated questions in fresh sessions, and graded 7,449 answers against government data. It classifies an answer as false when two graders agreed it contradicted the official figure. Within that narrow test, the report says 610 answers were false overall and reports a 22.6% false-answer rate for Perplexity. These are the study’s findings, not a universal error rate for Perplexity or for ordinary search tasks. Its questions included deliberately misleading claims, and the test concerned US local political facts before the 2026 midterms. Read the Searchable report and its methodology before applying its numbers elsewhere.

The practical response is neither “citations solve the problem” nor “AI search cannot be used.” It is to choose the verification depth by consequence, preserve the evidence trail, and refuse to average away a fatal mismatch. The protocol in this article helps a reader decide when to accept a sentence as supported, when to rewrite it, and when to stop and consult the original record.

What the word “citation” can mean

People often use “citation accuracy” as though it were one measurement. It is several separate questions. A URL can open but be irrelevant. It can be relevant but fail to support the exact clause. It can support the clause for an earlier year but not the current one. It can support an answer’s direction while contradicting its number. It can be the right source but refer to a different geography or group. A useful review keeps those errors apart.

The SourceBench paper on cited-source quality makes a related distinction: it evaluates the quality of pages referenced in answers, including relevance, factual verifiability, neutrality, freshness, accountability, and readability. Its authors collected 100 queries across several intent types and evaluated 12 systems over 3,996 cited sources. That design focuses on source quality; it does not mean every source-backed answer is factually correct. In its reported argumentative-task results, Perplexity Sonar Pro’s weighted source score was 78.547 on the paper’s scale, while the authors also discuss weaker neutrality and informational-query performance. Those scores belong to that benchmark, system version, query set, and scoring method. They cannot be converted into a probability that a particular answer is true.

A team should distinguish at least five layers:

1. Link validity: Does the citation open the intended page rather than a dead link, login screen, search page, or unrelated destination?
2. Source quality: Is the publisher accountable and suitable for this type of claim? A primary dataset is usually a better source for a figure than a commentary repeating it.
3. Passage relevance: Does the linked page contain the subject of the claim, or merely mention the same keywords?
4. Claim support: Does the cited passage entail the sentence as written, including its number, direction, qualifier, geography, and date?
5. Synthesis accuracy: Did the answer preserve distinctions among sources, or combine incompatible periods and definitions into one confident statement?

Passing a lower layer does not pass the next. Link validity is necessary for an audit, but says nothing on its own about source quality. A reputable source may contain a table that the answer has misread. A sentence can also be partly supported: a citation may prove the population size but not the causal explanation that follows it.

That last difference is easy to miss in fluent prose. A sentence such as “The rate fell after the programme, so the programme reduced risk” contains at least two claims: a chronological comparison and a causal inference. A table may establish the first and fail to establish the second. The reviewer should split the sentence before scoring it. One citation beside a compound sentence is not proof of every clause.

Why published results seem to disagree

Public studies do not measure the same thing simply because each mentions citations. Searchable’s October 2026 report grades answers against government facts in a particular set of local political questions. SourceBench scores the quality of cited pages across a benchmark of information, factual, argumentative, social, and shopping intents. One evaluates whether an answer contradicted a reference fact; the other evaluates properties of sources linked by systems. A good source score cannot guarantee a correct synthesis, and a false-answer rate from a high-risk political prompt set cannot be carried over to product research or stable definitions.

The Ofcom research on experiences of generative AI search answers another question. Its qualitative study used 24 in-depth online interviews with UK adult internet users to explore how people experience traditional and generative search. It is useful for understanding user perceptions and checking behaviour, but it is not a representative estimate of factual accuracy. Ofcom explicitly says the participants’ views are theirs, not the regulator’s position, and that the study does not determine the legal status of named products.

An older academic study, “Evaluating Verifiability in Generative Search Engines”, examined four generative search systems in 2023 and reported unsupported sentences and inaccurate citations in its tested sample. It is useful as evidence that answer-level support is a distinct research problem; it is not a current scorecard for today’s product versions. The product, prompts, retrieval pipeline, and evaluation sample have all changed since then.

These results can coexist without contradiction. They use different tasks, samples, dates, systems, and outcome definitions. The right inference is modest: published research gives reasons to verify claim-to-passage support, while no single published percentage tells a buyer what will happen on their own prompts. A team still needs a local acceptance test that records its own task boundaries and versions.

A further operational point comes from McKinsey’s discussion of generative-AI risk controls. The article is older than the newest product changes, so use it for general governance guidance rather than current interface details. It recommends controls at different points in a user interaction, including limiting accessible data, confirming that the request was understood, and using citations to support fact-checking. The useful principle is that citations belong inside a broader process; a citation by itself is not the control.

What people are asking in practice

The query evidence behind this article gives a practical question signal without pretending to be a survey. On 31 October 2026, a targeted Google Search Console query for the candidate Trust Vault page over 23–30 October returned 17 impressions for “perplexity ai search” at an average position of 28, 2 impressions for “perplexity ai what is it official” at 8.5, and 1 impression for “perplexity ai accuracy” at 32. The query was scoped to domain-scoped property for Trust Vault, the exact candidate page, and the stated seven-day date window. These are observed search rows, not market demand estimates and not a promise of ranking. The language suggests two useful reader questions: what makes AI search different, and how can a reader verify an answer’s sources?

That intent aligns with a recurring user concern visible in public discussions: “The link exists, but does the linked page actually say this?” Such discussions are anecdotal; they cannot tell us how often the failure occurs, whether a report is reproducible, or which product settings and model were involved. They are still useful for designing test cases. The audit below turns the concern into checkable fields: exact claim, link destination, supporting passage, source date, scope match, counter-evidence, and observed variation across runs. If a reviewer cannot fill those fields, the item is not verified.

Short video tutorials can be useful for learning where a product displays citations or how to open a source panel. A Perplexity fact-checking walkthrough, for example, demonstrates a user interaction. Treat it as an interface demonstration, not proof that the underlying citations are correct. Interfaces change, videos may show an older build, and a tutorial’s successful example says nothing about failure frequency. The test should be tied to a dated interface version and the actual live answer, not the creator’s explanation of what the interface means.

A practical acceptance test, not a vendor ranking

The test has two jobs. First, it checks whether a citation supports the claim it is attached to. Second, it checks whether the answer’s wording stays within what its sources establish. It deliberately does not rank Perplexity against another vendor and does not claim to represent the whole web. Teams can use the same worksheet for a research assistant, browser search summary, or internal retrieval system.

Method and limits

This protocol was prepared on 31 October 2026 from current Perplexity help documentation, the current Perplexity AI-search landing material, the Searchable report updated October 2026, SourceBench on arXiv, Ofcom’s 2025 qualitative study, a management article, a relevant video walkthrough, and public user discussions. The original reusable fixture contains 16 prompt scenarios, selected deliberately to cover four failure-prone classes: changing facts, source-to-claim mismatch, misleading premises, and consequences that require a primary record. The fixture is not a random sample of user queries. No Perplexity outputs were generated, no answer rates were measured, and no claim is made about the performance of any product on the fixture. Its purpose is to make a future test reproducible and to force reviewers to record failures precisely.

For a live run, record the date and time, region, signed-in status, product plan, search mode, selected model if shown, prompt text, session identifier or screenshot, and every cited URL. Repeat each prompt in a fresh session if you want to test run-to-run variation. Do not mix runs from different product modes and call them a single result. If settings are not visible, record “unknown.” Preserve this information with the worksheet so a later reviewer can tell what was actually tested.

The source set behind the article is not a substitute for legal, election, medical, investment, or safety advice. High-consequence questions require the official source and, where appropriate, a qualified professional. The fixture is intentionally small enough for a human to audit, not large enough to estimate a platform error rate. It avoids a false precision trap: a team can report the number of failures it observed in its own run, but should not pretend that number generalizes beyond that exact task set and run condition.

The 16-prompt fixture

Use neutral wording first. Then use a deliberately loaded variant where specified. This lets a reviewer see whether the answer challenges a false premise instead of silently inheriting it.

Replace bracketed values with a real location, time range, or named programme before running. Never use actual personal data in the prompt fixture. For law, benefits, health, safety, or money questions, choose the exact jurisdiction and write the policy date into the query. For a test of ordinary research, use the first 12 prompts and label the last four “not run”; do not invent outputs to fill a table.

The claim ledger: the original reusable asset

The useful unit of review is one atomic claim, not one answer. Copy each factual sentence into a separate row. Split a sentence if it contains a number and a causal interpretation, or if different clauses rely on different sources. Record the sentence as returned before rewriting it. That makes it possible to distinguish an answer error from a later editor’s correction.

A claim is supported only if a reviewer can point to a passage that establishes the statement as written, including its scope and time. It is partly supported if the page confirms only a subset, such as the direction but not the percentage. It is unsupported when the source does not establish the sentence, even if the sentence might happen to be true. It is unresolved when evidence is inaccessible, inconsistent, or too weak to decide. “Unresolved” is not a euphemism for supported.

For a recurring workflow, save the ledger as CSV or a database table with one row per claim and a separate table for sources. Do not put a dozen URLs into a single cell; doing so makes it hard to know which source supports which statement. If one source supports multiple claims, link the claims to the same source record but capture a distinct passage for each. If the answer provides a citation marker for the whole paragraph, the human still needs to decide whether the destination supports each sentence in that paragraph.

A practical evidence note should be short enough to reproduce. Record a table name, row label, quoted phrase, and page date; for a web article record the heading and the sentence; for a PDF record page and table. Keep the quote brief and within copyright rules. A screenshot can document what the interface showed, but it does not prove the source itself was accurate. A copy of the underlying dataset or immutable official release is stronger when the claim depends on a number that may later be revised.

Score evidence fit, not the visual polish

A weighted score can help compare routine answers, but it must not let a fatal defect disappear into an average. Use a 0–2 scale for each claim: 0 means absent or wrong; 1 means partial or ambiguous; 2 means direct and scope-matched. Score six dimensions: link opens, source authority fits, passage is relevant, exact claim is supported, time and scope match, and material counter-evidence is acknowledged. The maximum is 12. The scale is an internal decision aid, not a scientifically validated measurement and not a quality score from Perplexity.

Set an acceptance threshold before reviewing output. One workable internal rule for a low-impact desk-research summary is a score of at least 10/12, with no zero in passage relevance, exact support, or scope. For a customer-facing statement or a consequential decision, do not rely on a threshold alone: require direct support from an authoritative primary source, a second reviewer for the critical claim, and a recorded decision owner. No numeric score overrides a hard fail.

Hard fails include a dead or unrelated URL; a source that discusses the subject but not the stated claim; a number whose year, unit, denominator, or geography differs; a trend described in the opposite direction; a causal statement supported only by a correlation; or an important limitation omitted from the paraphrase. Any hard fail means rewrite, remove, or independently verify that claim before use. A failed critical claim blocks the answer from being used as a decision input even if every other citation is excellent.

For an answer containing a recommendation, separate factual premises from value judgments. “The rule applies to these accounts” is a factual premise. “Therefore we should migrate” depends on cost, risk appetite, and alternatives. The cited source may support the first proposition but cannot decide the second. The decision owner should list the assumptions and choose the action explicitly, rather than attributing the recommendation to a citation.

How to run the audit without turning it into busywork

Start with the decision, not the citation count. Ask what the answer will be used for and what would happen if one critical sentence is wrong. A summary used to find a source can tolerate more uncertainty than an answer pasted into a customer email or runbook. This classification determines which prompts to run, how many claims to inspect, and who must sign off.

Next, capture the live conditions. Write down the exact prompt, mode, model label, locale, date, and whether the answer came from a fresh session. If the product hides a setting, record that as unknown. When comparing two answers, hold the prompt and session conditions as constant as practical, and run them separately. Do not compare one answer from a long research mode with another from a short search mode and infer that the model alone caused the difference.

Then split the answer. Highlight every factual claim and assign a stable identifier such as R1-C01. Divide multi-clause sentences until each row has one proposition. Keep interpretation and advice in a different column. This step is often where a reviewer discovers that a paragraph contains more claims than citations, or that a sentence quietly moves from “may” to “will.”

Open each cited page. Do not stop at the search snippet or the AI’s source preview. Check the date, author or responsible organization, source type, and whether the relevant text appears in the page. If a page is paywalled, blocked, or absent, mark it inaccessible and find the underlying public record. A citation that is impossible for the intended user to verify does not meet the same usability bar as an open source.

Compare claim and evidence line by line. Preserve the units. If the answer says “up 12%,” establish whether the source says a 12% relative increase or a 12 percentage-point change. Check whether “users” means all users, a survey sample, registered accounts, or customers in a particular region. For current questions, compare the publication date and effective date; a page updated today may still describe a rule that took effect last year.

Search for the strongest counter-source, not just another page repeating the first one. For a government statistic, read the agency’s definition and revision note. For a disputed claim, look for the primary record and a credible source offering a different interpretation. A news story can help locate the record but should not replace it when the record is available. If sources genuinely conflict, explain why the evidence cannot be collapsed into a single confident statement.

Finally, choose an action for every claim: keep with citation, narrow the wording, update the figure, add a missing qualifier, seek a second source, or delete. The reviewer should not spend time polishing a sentence that lacks evidence. A concise answer with four verified claims is more useful than a long answer whose citation markers are decorative.

Repeatability and variation across runs

Answers can differ because a live index changes, because a product selects another model, because search mode changes, or because the wording changes. A one-run screenshot is a record of one event, not a durable measure of product behavior. If the operational question is whether a workflow can be trusted over time, repeat the exact prompt in fresh sessions and compare the claims, not only the list of URLs.

For each run, preserve the answer text, citation destinations, and visible mode. Compare each claim across runs using four outcomes: same and supported; same but support differs; changed but still supported; or changed and unsupported. A changed citation is not automatically a failure: an alternative primary record may be equally strong. Conversely, a stable citation is not automatically success if it repeatedly fails to support the claim.

Use a simple disagreement register. Record which statement changed, what source changed, whether the primary record changed, and whether the revision changes the decision. If a value varies across runs but the official table is stable, the model has an inconsistency. If the underlying agency later revises the table, the difference may be legitimate freshness. The audit needs to distinguish system variation from source revision.

Do not present repeatability as a percentage unless the denominator, prompt set, and repetition design are explicit. If a team runs 16 scenarios twice, that produces 32 answer instances, not 32 independent user questions. If the same scenario is repeated five times, report five runs of that one prompt and say so. A fixture designed to expose failure types is useful for acceptance, but it is not representative sampling.

Failure modes and recovery actions

A real page, wrong claim. The link works and the publisher is reputable, but the cited paragraph does not say what the answer claims. Split the sentence, mark the citation unsupported, and search for the original record. Do not let the domain’s reputation substitute for passage-level support.

The correct number, wrong period. The number appears on the page but belongs to another year, month, base period, or publication vintage. Write the unit and period into the claim, then use the dataset’s date and revision notes. If the value changes frequently, include an “as of” date in the answer.

The statistic moved, the meaning did not. An answer says the value rose by 12%, while the source reports a change of 12 percentage points. The magnitudes may be very different. Recalculate from the original counts when possible; otherwise quote the source’s exact measure and wording.

A source is secondary but presented as original. A blog links to a report, and the answer cites the blog as though it generated the data. Follow the link chain to the original issuer. Keep the commentary only when it adds an interpretation, and label that interpretation as such.

The prompt smuggles in a conclusion. “Why did X cause Y?” assumes causation. Rewrite it as two questions: did Y change, and what evidence links the change to X? If causation is not identified by the source design, say that the evidence is descriptive or correlational.

The answer cites both sides without weighing them. A list of pro and con sources can create false balance. Record source type, directness, data, and limitations. A primary record and an anonymous post do not carry equal evidentiary weight just because each has a link.

The source preview replaces reading. A snippet can omit a table footnote or a sentence that reverses the interpretation. Open the page or mark it inaccessible. When the page cannot be read, do not promote the claim from unresolved to supported.

The citation count is mistaken for coverage. Five URLs may all repeat one syndicated source. Group by underlying origin. For independent confirmation, look for distinct data collection or a separate primary record, not five domains copying the same release.

A citation checker agrees with the answer. Automated claim matching can help triage, but a second model is not an independent authority if it relies on the same retrieved pages or scoring assumptions. Use automation to identify passages, then have a human decide whether they entail a consequential claim.

The reviewer changes the prompt mid-test. Adding “be accurate and cite sources” after seeing a poor answer changes the condition. Save the original failure, run the new prompt as another test, and do not erase the first outcome. Prompt tuning is a mitigation experiment, not proof that the original workflow was safe.

When a failure occurs, keep the response out of the downstream action path. Preserve the raw output and evidence row, assign an owner, and record the corrected version as a new revision. If an answer has already been shared, correct the specific claim and notify the people who relied on it. Deleting the failed row or quietly editing the prompt destroys the evidence needed to improve the process.

The smallest useful implementation

A spreadsheet is enough for a one-time evaluation. Use one row per atomic claim and columns for run ID, prompt ID, claim text, cited URL, passage, source date, scope, support score, hard fail, reviewer, and action. Keep the test fixture in a separate tab. Freeze the original answers so the reviewer is not scoring a rewritten version.

For repeat runs, use a versioned folder layout such as audit-date/product-mode/prompt-id/. Store a text export or screenshot, a source snapshot where reuse is lawful, and a manifest.csv. Do not store credentials, personal data, private search history, or client material in a public repository. If a source page is copyrighted, capture only the short passage needed to substantiate the specific claim and link to the original page.

A simple local script can enforce the hard-fail rule before averaging. The script should not decide whether evidence is true; it should refuse a pass when a critical claim has a zero for support or scope. Here is a small CSV validator using Python’s standard library. Save it as audit_gate.py, create a CSV with the named columns, and run python audit_gate.py claims.csv.

import csv
import sys

REQUIRED = {"claim_id", "impact", "support", "scope", "action"}
CRITICAL = {"high", "critical"}

with open(sys.argv[1], newline="", encoding="utf-8") as file:
    rows = list(csv.DictReader(file))
if not rows or not REQUIRED.issubset(rows[0]):
    raise SystemExit(f"CSV must include: {', '.join(sorted(REQUIRED))}")

blocked = []
for row in rows:
    try:
        support = int(row["support"])
        scope = int(row["scope"])
    except ValueError:
        blocked.append((row["claim_id"], "missing score"))
        continue
    if row["action"].strip().lower() == "hard-fail":
        blocked.append((row["claim_id"], "hard fail"))
    elif row["impact"].strip().lower() in CRITICAL and (support == 0 or scope == 0):
        blocked.append((row["claim_id"], "critical claim lacks direct support or scope match"))

if blocked:
    for claim_id, reason in blocked:
        print(f"BLOCK {claim_id}: {reason}")
    raise SystemExit(1)
print(f"PASS {len(rows)} claims: human review still required")

The script checks input completeness and a fail-closed rule; it does not fetch URLs, parse evidence, or certify truth. It assumes the human reviewer assigned scores consistently. A passing output means only that the recorded rows contain no listed hard fail and that critical rows have non-zero support and scope. A release owner must still review source class, date, and the primary passage.

A decision matrix for teams

This is a risk rule, not a vendor comparison. A well-performing tool on one set of queries can still fail a particular high-consequence claim. Conversely, a weak answer can still help a researcher find the document needed to answer correctly. The relevant unit is the task and its decision boundary.

What to do when the citations conflict

When two cited pages disagree, do not ask the model to “pick the more reliable one” and accept its choice without inspection. Identify the underlying reason: different reporting periods, definitions, jurisdictions, sample populations, revisions, or genuinely contested interpretation. The oldest document may be the legal authority; the newest may be a summary. The most recent number is not necessarily the correct number for the question.

Create a conflict note with four fields: claim at issue, source A and its date, source B and its date, and what would resolve the disagreement. Then look for the issuing agency’s definition, methodology, or original document. If the record remains ambiguous, state the ambiguity and the decision needed. For a buyer or analyst, an explicit unresolved issue is more actionable than an invented consensus.

The 2026 Searchable election study provides an example of why scope needs to remain attached to a result. Its reported Perplexity figure concerns a fixed set of local political questions, repeated under a specified procedure and judged against official data. It includes misleading-premise prompts and a particular pre-election period. A researcher should not rewrite that as “Perplexity gets 22.6% of answers wrong” in general. The permissible statement is narrower: “In Searchable’s described test of US local political questions, its graders classified 22.6% of Perplexity’s tested answers as false by their stated definition.” The attribution, sample, and definition travel with the number.

SourceBench offers a different useful example. Its eight dimensions force a reviewer to look beyond whether a link exists: accountability, freshness, neutrality, relevance, and layout all affect whether a person can audit a page. Yet it deliberately evaluates source quality, not the correctness of the answer text. A source-quality score should not be renamed an “accuracy rate.” If your worksheet collapses these dimensions into one number, report the components as well as the aggregate.

A compact workflow for a small team

A practical team can assign three roles without creating a committee: the requester defines the decision; the analyst runs the query and fills the ledger; the reviewer checks the critical claims and signs the release. One person can hold more than one role in a low-risk exploration, but for high-impact decisions the same person who asked the question should not be the only reviewer of the answer they want to use.

Before the first query, write a one-sentence intended use. “Find the official source for a current threshold” is a safer purpose than “tell me what to do.” Define what evidence would be sufficient before seeing the answer. This prevents a persuasive summary from changing the bar after the fact.

After the query, freeze the first response. If a citation is wrong, capture it as a failure. Then ask a follow-up such as “which sentence is unsupported, and what passage would disprove it?” This can help uncover a problem, but it is not an independent check; a model may produce a second confident answer. The reviewer still opens the page and records the exact supporting passage.

Once the ledger is complete, apply the decision matrix. Low-risk exploratory notes can remain provisional and clearly marked. Customer-facing claims need correction and sign-off. Critical claims need an authoritative record and an accountable owner. If a workflow writes directly to a CRM, sends a customer message, or triggers an eligibility action, insert the verification gate before that action, not after an error report.

Keep the version record. A product update, mode change, region, or selected model may change retrieval behavior. If an organization wants a longitudinal comparison, repeat a stable fixture at a defined interval and record the visible build information. Avoid a claim that “the tool improved” based on two runs with different prompts or different policy questions.

What this article can and cannot establish

The public materials reviewed here support a few bounded conclusions. The vendor describes citations as a route to source material and tells users to validate. Research distinguishes source quality from answer correctness. A current private-sector report documents errors on one narrow set of civic questions, while a regulator-sponsored qualitative study documents how a small interview sample experienced generative search. The sources do not establish a universal accuracy rate for Perplexity or prove that the same issues affect every task.

The 16-case fixture is original as an operational asset, but it has not been executed against Perplexity or another product. It provides a repeatable way to inspect future answers and report outcomes without claiming generalization. A team may publish its own measurements only after it runs the fixture, records the exact product conditions, defines its adjudication rules, and states the sample and exclusions. Its count should be described as the observed result for that run, not a population estimate.

The protocol also does not certify a source as unbiased. Source class and institutional accountability are not the same as truth. Government tables can be revised, company studies can be methodologically narrow, academic work can be preliminary, and user reports can be sincere but unverified. The reviewer must read limitations and decide whether the evidence is fit for the question.

The best operational conclusion is straightforward: citations make a generated answer inspectable, but only claim-level review makes it verified. For routine discovery, that distinction lets teams use an answer to find useful evidence without surrendering judgment. For a decision, require the exact passage, scope, date, and source class before the sentence moves downstream. Keep the raw answer, record the correction, and let the evidence determine what can be said.

Sources and review notes

• Perplexity Support, What is Pro Search?, product documentation last modified 26 October 2026; describes modes and recommends users validate answers through linked sources.
• Searchable, AI Misinformation Report: US Midterm Elections, updated 7 October 2026; company-produced study, fixed US political question set and methodology as summarized above.
• Jin et al., SourceBench: Can AI Answers Reference Quality Web Sources?, arXiv, 18 February 2026; cited-source quality benchmark, not an answer accuracy test.
• Ofcom, User experiences of Generative Artificial Intelligence search, 26 September 2025; 24 qualitative interviews, not a representative accuracy estimate.
• McKinsey, Implementing generative AI with speed and safety; older governance discussion, used for process context only.
• Product demonstration: Perplexity fact-checking interface walkthrough (demonstration only; not evidence of accuracy).

Reviewed on 31 October 2026. The article makes no ranking promise, reports no original model test result, and treats source access, citation support, and answer correctness as separate questions.