Ask a question, get an answer you can check

You need one fact for a slide or an email: how much of the world's electricity came from solar last year, when a rule starts to apply, what a study actually found. A chatbot will answer in seconds, and its answer will sound the same whether it's current, out of date or invented. This page is about a different kind of answer, one built only from sources you can open, with every point linked, that says plainly when those sources don't settle it. It walks through a real question we asked twice: once when it came back "Not enough evidence", and once when it came back with a figure and a caveat worth reading.

Answers that never admit doubt

The trouble with a confident answer isn't that it's sometimes wrong. It's that you can't tell when. In March 2025, researchers at Columbia's Tow Center for Digital Journalism tested eight AI search tools by giving them excerpts from news articles and asking them to identify the source: the headline, publisher, date and link. Across 1,600 queries, the tools gave incorrect answers more than 60% of the time.

The more telling finding was about tone. The researchers found the tools "presented inaccurate answers with alarming confidence, rarely using qualifying phrases." One of them signaled a lack of confidence in just 15 of its 200 responses and never declined to answer. More than half the responses from two of the tools cited fabricated or broken links.

An answer you can use has to be able to say "the evidence doesn't show this". If it can't, its confidence tells you nothing.

How a question is answered

Choose Ask a question, or type into Ask anything and CiteJury works out that you're asking a question (it asks you if it isn't sure). The question is broken into a few sub-questions, and each one is searched across several independent search indexes. The search looks first for direct, primary answers such as official statistics, annual reviews, studies and regulators, and also for evidence that would point to a different answer.

The best passages are locked into a sealed evidence pack. In a Standard check, Claude, ChatGPT and Grok each answer from that pack alone, with no web search, and every quote they cite is checked word for word against its source. A final review writes the result: a direct answer in a few sentences, the points behind it with a link on each, and every source used with a ruling on how it bears on the question.

If the pack doesn't answer the question, the result says Not enough evidence. It won't piece a figure together from partial data, and it won't fall back on what a model remembers. You also see the sources that were set aside and why, with reasons like "Couldn't open" or "Site blocked us".

When there is an answer, you can ask follow-up questions on the finished check. They answer only from the same sources and cost about 1 to 2 credits.

The solar question, asked twice

We asked a question that sounds easy: how much of the world's electricity came from solar in 2025? This is the answer that came back.

“How much of the world's electricity came from solar in 2025?”

Answered

Claude agrees. ChatGPT said Not enough evidence. Grok said Sources disagree12 sources usedTook 3 minutes

Solar was over 8% of world electricity in 2025 (IEA), about 2,778 to 2,800 TWh.

• Solar's share of world electricity in 2025

  Over 8%, on IEA figures.

• Solar generation in 2025

  About 2,800 TWh (IEA) and 2,778 TWh (Ember).

• New solar capacity in 2025

  605 GW, from one article's headline citing the IEA.

• Who publishes the figures

  The IEA and Ember agree on volume. Only the IEA gives a share.

Safe way to say it

Say "over 8% in 2025, per the IEA". The share comes from a single passage reporting IEA figures.

09pv-magazine.comAnswers it

11gmk.centerAnswers it

05carbonbrief.orgAnswers it

CiteJury broke the question into four sub-questions: solar's share of world electricity, total solar generation, new capacity, and who publishes the figures. It searched 78 results and locked 12 sources into the pack. The answer: solar supplied over 8% of the world's electricity in 2025, according to the IEA. In volume that was around 2,800 TWh on the IEA's count and 2,778 TWh on Ember's, roughly 30% more than in 2024.

Now look at the line under the verdict. The AIs didn't agree. Claude answered, ChatGPT said there wasn't enough evidence and Grok said the sources disagreed. The final review read the same pack, found the figures and wrote the answer, but it didn't hide how thin the share was. The safe wording says to attribute "over 8%" to the IEA, because only one passage in the pack gave a solar share. Ember's pages gave generation in TWh, not a solar-only share, so the two couldn't be compared on that point.

That caveat holds up. Pew Research, drawing on Ember's data, puts solar at 9% of the world's electricity in 2025, which sits inside "over 8%". If your slide needs the precise figure, the answer has already told you where to get it: a primary source for the share, not a report quoting one.

The first time we asked, a day earlier, the result was Not enough evidence. The pages that carried the figure, Ember's and the IEA's, block automated access, and the passages the check could read stopped short of the number. That result said so, listed the blocked pages as "Site blocked us" and "Couldn't open", and didn't invent a figure to fill the gap. On the second run the search found reports that quoted those figures directly.

Both results were honest about what they could read. That's the behavior you want from a sourced answer: a figure when the sources give one, with a note on how much weight it bears, and a plain "not enough evidence" with the blocked pages listed when they don't. A tool built to always answer would have handed you a number both times, and nothing to tell you whether it was this year's figure, last year's or a guess.

Ask a question yourself

1. Ask one thing at a time. "What share of electricity came from solar in 2025, and is it growing faster than wind?" is two checks.
2. Name the place, the period and the measure. "Solar's share of global electricity generation in 2025" beats "how big is solar now".
3. Choose a tier. Quick, about 2 credits, uses one AI with no final review. Standard, about 8 credits, uses Claude, ChatGPT and Grok plus the review, usually in one to three minutes. Deep, about 32 credits, searches wider and lets the AIs run their own searches while sources are gathered.
4. Read the answer, then the points. Each point links to the passage it rests on. Open the ones you'll repeat.
5. If it says Not enough evidence, read the set-aside list. Pages marked "Couldn't open" or "Site blocked us" are often where the answer lives. Open them yourself, or narrow the question to a period with more coverage.
6. Follow up from the same sources. Ask what a figure covers or why two sources differ.

Questions that get good answers

• Allow for publication lag. Annual statistics for a year usually appear months after it ends, and coverage builds after that. A question about last year asked in January is more likely to come back Not enough evidence.
• Point at a publisher. "What does the IEA estimate for 2025" gives the search a clearer target than "what is the true figure".
• Separate facts from judgments. "Is X better than Y" isn't a question of fact. For that, Compare options gives you a sourced side-by-side table.
• Use a brief for big topics. If you need what's settled, what's contested and what's still open, a sourced research brief fits better than a single answer.
• Check the answer you already have. If a chatbot has already given you an answer, paste it in as a claim instead. See how to fact-check an AI answer.
• Treat Not enough evidence as a to-do. It tells you the answer needs a primary source, and often which pages to open first.

What it won't do

• Read every page. Pages behind paywalls or logins, and some sites that block automated access, can't be read, as the solar example shows. They're listed so you can open them yourself.
• Read charts. It checks text, not images, charts or video, and a lot of statistics are published as charts first.
• Do the arithmetic for you. It won't derive an answer the sources don't state.
• Cover breaking news. Very recent events may have little coverage, which often means Not enough evidence.
• Stand in for a professional. Answers to legal, financial or medical questions aren't legal, financial or medical advice.
• Always be right. Answers can be wrong or incomplete. Every point links to its source so you can check what you rely on.

A sourced answer takes a minute or two longer than a chatbot's. What you get for that time is an answer you can trace, and an honest note when the sources it could read don't give you one.

Questions people ask