You need an AI answer you can trust, but a leaderboard alone won’t tell you which platform is safest for your work. The lowest overall hallucination rate can still hide substantial risk when you ask about prices, local businesses, citations, statistics, or a niche subject.
The useful question is not simply, “Which AI hallucinates least?” It is, “How often does this platform fail on my kind of fact, my kind of prompt, and my response settings?” Once you separate those variables, you can choose tools more intelligently and build content that AI systems are less likely to misrepresent.
Start with the hallucination metric that matches your risk
A hallucination is a factual claim presented as true even though it was fabricated, contradicted by authoritative evidence, or attached to a citation that did not support it. Opinions, predictions, and writing mistakes are different problems.
Two rates matter, and they answer different questions:
- Response-level hallucination rate: the percentage of complete answers containing at least one false factual claim. This is the best starting point for estimating how often a user encounters a flawed answer.
- Claim-level hallucination rate: the percentage of individual factual claims that are false. This shows how much of the factual material inside all responses is unreliable.
Those numbers are not interchangeable. In September 2026, for example, ChatGPT had an 11.4% response-level rate and a 2.9% claim-level rate. That does not mean 11.4% of everything it said was false. It means 11.4% of complete responses contained at least one hallucination, while 2.9% of the factual claims examined were false.
Citation errors and refusals require separate columns. A response can reach the correct conclusion but cite a page that does not support it. A platform can also improve its apparent accuracy by declining difficult questions more often. If you measure only answered prompts or only citations, you can reward the wrong behavior.
For a controlled September 2026 comparison, 4,182 prompts were submitted to seven platforms in their default consumer settings. The prompt set covered nine subject areas and seven query types. Two reviewers checked each response against primary evidence, with a third resolving disagreements. That design makes the platforms comparable within this test; it does not establish a permanent error rate for every model, mode, language, or use case.
Rates also move as products change. From Q1 2024 through Q3 2026, the all-platform average fell from 31.7% to 16.0%. The improvement was not perfectly linear: ChatGPT rose in Q4 2024, and Gemini rose in Q2 2025, when new models attempted more difficult prompts instead of declining them. Treat any ranking as a dated snapshot, and record the platform, mode, and test date whenever you run your own evaluation.
The September 2026 platform ranking has an accuracy–refusal tradeoff
Here is the September 2026 platform comparison, including the measures that a headline ranking can obscure:
| Platform | Responses with a hallucination | False factual claims | Citation error rate | Declined to answer |
|---|---|---|---|---|
| Claude | 8.7% | 2.1% | 6.4% | 11.8% |
| ChatGPT | 11.4% | 2.9% | 9.7% | 6.3% |
| Google Gemini, including AI Mode | 13.9% | 3.4% | 11.2% | 5.1% |
| Microsoft Copilot | 14.6% | 3.7% | 12.8% | 4.6% |
| Grok | 17.3% | 4.6% | 18.9% | 2.7% |
| Meta AI | 21.8% | 5.8% | 23.4% | 3.2% |
| DeepSeek | 24.6% | 6.7% | 27.1% | 1.9% |
Claude produced the lowest response- and claim-level rates, but it also declined the largest share of prompts. DeepSeek declined the fewest and produced the highest hallucination rate. Across all seven platforms, greater willingness to decline tracked closely with greater accuracy.
That relationship does not make refusal inherently good or prove that refusal alone caused the accuracy difference. It tells you to evaluate two failures separately: a system that cannot complete the task, and a system that completes it with fabricated facts. A visible refusal can be routed to another workflow. A confident falsehood may pass unnoticed into a page, report, customer answer, or executive decision.
The citation column deserves particular attention. Grok, Meta AI, and DeepSeek each had a citation error rate above their overall response-level hallucination rate. Asking for links did not automatically make their answers safer. If your workflow depends on references, score whether each cited page actually supports the adjacent claim rather than awarding credit for the presence of a URL.
Choose a platform using the task you need completed. For high-volume drafting from supplied material, claim-level accuracy and adherence to the provided text may matter most. For autonomous research, current facts and citations matter more. For customer-facing answers, the response-level rate is crucial because one invented price, feature, address, or policy can make the whole answer operationally unsafe.
Topic and prompt design can matter more than platform choice
Facts that change frequently create the largest exposure
Hallucination rates varied more by subject than by platform. The nine-topic breakdown shows where a general platform average can understate your actual risk:
| Topic | Responses with a hallucination | Typical failure |
|---|---|---|
| Local business information | 26.4% | Outdated hours, addresses, and closures |
| Legal questions | 23.8% | Invented case law and statutes |
| Current events | 21.2% | Conflated or outdated events |
| Product and brand information | 19.7% | Invented features and pricing |
| Financial questions | 17.1% | Outdated rates and figures |
| Medical and health questions | 15.3% | Outdated or misapplied guidance |
| Software and coding questions | 12.4% | Invented functions and packages |
| Scientific and technical questions | 9.6% | Misattributed findings |
| Historical facts | 6.8% | Incorrect dates and figures |
The pattern is practical: frequently changing facts require a current retrieval step and stronger verification. If you manage a local company, do not apply the 6.8% historical-fact rate to questions about today’s opening hours. Your closer baseline is the 26.4% local-business rate. If you publish product information, the relevant benchmark is 19.7%, with prices and nonexistent features among the recurring errors.
Legal, medical, and financial answers carry consequences beyond content quality. Do not rely on a generated answer as the final authority for an individual decision. Verify it against current official material and use a qualified legal, medical, or financial professional where the decision requires professional judgment. The specific failure modes here—nonexistent law, outdated figures, and misapplied guidance—can sound entirely plausible.
The wording of the question changes the odds
The same platforms produced sharply different results across seven query types:
| Query type | Responses with a hallucination | Risk versus a simple lookup | What to do |
|---|---|---|---|
| Summary of a provided document | 4.3% | 0.6x | Tell the model to stay within the supplied material and identify anything the material does not establish. |
| Simple factual lookup | 7.2% | 1.0x | Check consequential facts against the current primary page. |
| Multi-step reasoning | 14.8% | 2.1x | Separate the problem into stages and inspect intermediate factual claims. |
| Request for statistics or figures | 22.6% | 3.1x | Require a date, unit, denominator, geography, and primary origin for each number. |
| Request for citations or sources | 27.9% | 3.9x | Open every citation and confirm that it supports the exact claim. |
| Question about a niche topic | 31.4% | 4.4x | Supply defining documents, terminology, and boundaries before asking for conclusions. |
| Question containing a false premise | 46.7% | 6.5x | Instruct the model to test and correct the premise before answering. |
A false premise was the most dangerous prompt pattern. In many of those answers, the platform accepted the faulty assumption and built additional claims on top of it. This matters in competitive research and reputation monitoring, where a user may ask why a company “removed” a feature it never offered or “closed” a location that remains open.
Requests for citations were also riskier than ordinary lookups. A citation request gives you something to inspect; it does not guarantee the citation is real, current, or relevant. Treat citation checking as its own task: open the page, find the supporting passage, confirm the entity and date, and make sure the claim has not been broadened beyond what the page establishes.
Retrieval lowers hallucinations, but verification still matters
Changing the response setting produced a material difference. When a 1,100-prompt subset was rerun under alternative settings, retrieval had a larger effect than extended reasoning alone:
| Response setting | Responses with a hallucination | Citation error rate | Change from default |
|---|---|---|---|
| Default mode | 16.0% | 15.6% | Baseline |
| Web retrieval | 10.9% | 9.3% | 31.9% lower |
| Extended reasoning | 13.2% | 14.1% | 17.5% lower |
| Web retrieval with extended reasoning | 8.6% | 7.8% | 46.3% lower |
| Uncertainty instruction | 11.7% | 12.4% | 26.9% lower |
Retrieval cut the response-level rate by nearly a third. Combining retrieval with extended reasoning reduced it from 16.0% to 8.6%, a 46.3% relative reduction. That is a meaningful improvement, not a guarantee. Even the strongest setting in this comparison still produced flawed responses, and its citation error rate remained 7.8%.
Extended reasoning alone lowered the response-level rate to 13.2%, but citation errors moved only from 15.6% to 14.1%. More reasoning can help a model organize a problem without giving it better evidence. If the task depends on changing facts, let retrieval obtain current material before asking the model to reconcile or analyze it.
A useful instruction is: Use web retrieval. Treat the premise as unverified. Prefer current first-party pages. If credible pages conflict, identify the conflict. Do not invent a missing fact, statistic, or citation. You should still inspect any claim that affects publication, pricing, compliance, health, money, or legal exposure.
Key takeaways
- Use a platform’s overall hallucination rate as a baseline, not as a prediction for every task.
- Keep document summaries bounded to the material you supplied; this was the lowest-risk query type tested.
- Use retrieval for current facts, then use reasoning to analyze the retrieved evidence.
- Tell the model to challenge the premise and admit uncertainty instead of completing a plausible story.
- Audit citation support independently from answer correctness; a URL is not proof.
Make your brand facts easier for AI systems to represent correctly
Platform accuracy is only half of the AI visibility problem. Product and brand information produced a 19.7% response-level hallucination rate, and errors appeared most often when a brand’s own website conflicted with third-party pages. If your site, business profiles, resellers, directories, and old announcements disagree, an AI system has several competing versions of the same fact.
Use this workflow to reduce ambiguity and measure whether the cleanup works:
- Inventory facts that can change. Include locations, hours, service areas, prices, product versions, feature availability, contact details, policies, and status changes. These deserve more frequent review than stable company history.
- Choose a canonical first-party page for each fact. Put the answer in visible copy, state its region or version where relevant, and show an effective or updated date when recency matters. A user and a retrieval system should not have to infer which announcement superseded another.
- Align visible content and structured data. Use the appropriate Organization, LocalBusiness, Product, or Offer markup where it accurately describes the page. JSON-LD should confirm the visible fact, not introduce a different price, address, availability state, or product name.
- Resolve external contradictions. Check major profiles, directories, distributors, support pages, and archived campaign pages. Correct the places you control. Where you cannot change an external page, make the current first-party statement unambiguous and easy to cite.
- Build a fixed evaluation set. Test direct lookups, requests for figures, requests for citations, niche terminology, comparisons, and deliberately false premises. Include the questions customers actually ask, not just branded prompts that are easy to answer.
- Score separate failure types. Record whether each response contains any hallucination, which individual claims are false, whether each citation supports its claim, and whether the platform declined. Do not compress those outcomes into one vague “accuracy” score.
- Save enough context to reproduce the result. Keep the prompt, platform, visible model or mode, retrieval setting, date, response, cited URLs, verdict, and the page that needs correction. Retest after meaningful model changes or after you update the underlying fact.
This turns AI visibility monitoring into a content-quality process rather than a collection of screenshots. When an answer is wrong, you can distinguish a platform-wide weakness from a prompt problem, an outdated third-party listing, a missing canonical page, or contradictory structured data.
Begin with the brand facts that change most often and carry the greatest consequence when wrong. Publish one clear version, align the surrounding ecosystem, and test it with the prompt patterns most likely to fail. The goal is not to make hallucination impossible. It is to make the correct answer easier to retrieve, easier to support, and easier for you to verify before anyone acts on it.
References
Leave a Reply