How to check a fact, a company or a figure in five minutes
The answer arrives confident and with references, and then it turns out one leads somewhere else and a third does not exist. There will be no signal saying "I am not sure here". Below is why a search tool is built differently and the four techniques that make it usable for work.
The answer arrives confident and with references, and then it turns out one leads somewhere else and a third does not exist. There will be no signal saying "I am not sure here". Below is why a search tool is built differently and the four techniques that make it usable for work.
---
The problem is not that AI does not know. It is that it does not doubt
You ask and get a confident answer with references. It looks convincing. Then it emerges that one reference leads elsewhere, a second to a page saying something different, and a third does not exist.
A mistake looks exactly like a correct answer. There will be no signal saying "I am unsure here": a number looks like a number, a date like a date, a reference like a reference.
For an everyday question that does not matter. For somebody whose wrong fact costs their reputation — a lawyer, a doctor, a finance professional, a writer, an entrepreneur — it makes the tool unusable. Right up until a verification process appears.
The process exists and it takes minutes rather than hours.
Why search AI works differently from a chat
The key difference, from which everything else follows.
An ordinary model answers from memory. It does not search — it reconstructs from what it saw during training. Hence the invented references: there is nowhere to get them from, but the answer has to be complete.
A search tool searches first and answers second. It has sources and it works from them.
Hence a non-obvious consequence: long detailed requests work worse here than short ones. In an ordinary model, context helps. In a search tool it interferes: a long request throws off the search and you get an answer from the wrong selection.
The rule: separate the intent to search from the intent to output. First a short precise query to find the right thing. Then separately, what to do with what was found.
Four techniques
1. Precise search instead of general
The general web gives you what is best optimised rather than what is most accurate. A substantial part of the skill is being able to narrow the field: search specific sources, limit by time, go to databases rather than to articles about databases.
One properly narrowed query saves an hour of reading paraphrases.
2. The zero-trust protocol
The main technique and the main principle. Stated like this: make the tool prove every statement.
A direct quotation rather than a paraphrase. An indication of exactly where from. And separately, an analysis of contradictions: where the sources diverge.
The last is the most valuable. Sources agreeing proves nothing — they may have copied each other. A divergence always shows where to dig.
3. Separate facts from conclusions
Ask for the result in two parts: first a list of statements with their sources, then the interpretation separately.
That gets you a ready list of what needs checking instead of a wall of text where everything does.
4. Multi-stage research
A complex question does not get solved by one query. It gets solved by a chain: found the basics → saw what was missing → refined → found a contradiction → worked it out.
In that scheme you are the lead and the tool is the research team. Your job: setting direction and deciding what to do with what was found.
What verification looks like in practice
Three minutes per fact, no more.
Step one — isolate the statement. Not "check the text" but take one specific thing: a figure, a date, a number, a name. Checking wholesale is not possible; you check point by point.
Step two — find the primary source rather than a paraphrase. An official document, a register, an organisation's own site, the original publication. An article citing an article citing an article is not verification.
Step three — compare word for word. Not "roughly that" but is it exactly that figure, that period, that definition. More than half the errors here are not invented facts but correct facts about something else: the figure exists but relates to a different year or a different country.
Step four — record it. The link to the primary source, next to the fact, in your own document. Otherwise in a month you will be checking it again.
And the main discipline rule: whoever signs the result does the checking. If material goes out in your name, the responsibility for every figure is yours regardless of who found it.
What always gets checked
A short list worth holding in mind.
Figures and dates. All of them. Particularly the plausible-looking ones.
Numbers of rules, documents and standards. The most frequent category of invention.
References. Open them. Not "there is a reference" but go there and see what is claimed.
Quotations. Check the wording and the context.
Names and titles. Does it exist, is it that person, is it that company.
What need not be checked: the structure, the reasoning, the phrasing. Those you judge with your own expertise.
Three signs an answer cannot be trusted
Not a guarantee but grounds for suspicion before you start checking.
Figures too round. Real data is rarely even. "Around thirty per cent" may be true, but more often it is a sign the figure was assembled rather than found.
A reference with no specifics. "According to research" and "experts say" is the absence of a source. A real reference leads to a specific document with a specific page.
An answer that is too convenient. If what was found perfectly confirms what you wanted to hear, check it twice over. The tool adapts to how the question was phrased, and a leading question gives a leading answer.
Hence a technique: ask the same question the other way round. If the answer flipped, you have not a fact but a reflection of your query.
Where this gets used in practice
Checking a client before a deal. How long they have existed, whether there are traces of other projects, whether what the listing says matches what is visible from outside. A minute of checking against an hour of free "test work".
Analysing somebody's business before a conversation. Three precise observations gathered before the first message work better than any CV.
Checking data in your own material. Before putting a figure into a client's deck.
Competitive intelligence. What they do, what they stay silent about, where there is space.
Where to start
The "Perplexity as an investigative tool" express course — five modules on turning AI search into an analyst. Inside: the search-first approach and why short queries work better than long ones, precise search techniques and working with databases rather than the general web, structured synthesis, zero-trust verification protocols and multi-stage research.
Who it is for: people who need accuracy — experts, writers, specialists and entrepreneurs for whom a wrong fact costs reputation and manual re-checking costs time.
The result: you stop guessing whether a reference was invented and get material you can put to work with no redoing.
The price is $9. The course is included in the Basic plan, and registration is free and opens three days of full Basic access — meaning you can simply take it having paid nothing.
Alongside it are five more foundational courses, the ten-lesson sales module and chats with 140+ AI assistants across 14 categories, analytics and legal among them.
Do one thing today: take the last answer you got from a model and used somewhere. Find every verifiable statement in it — figures, dates, references, numbers. Check them against sources. If it all matches, you have a habit now. If not, you have just understood why a protocol is needed.
---
The next level: verification inside the process
The four techniques above are manual work. There is an approach with re-checking built in.
Agentic RAG from the automations archive works differently from ordinary search: it dynamically selects several sources, refines the query if the first result is weak, re-checks what it found and makes a second pass.
The phrasing from its description: ordinary RAG is like a student copying and pasting, one source and one attempt. Agentic is an agent working autonomously from the search to the finished text.
Plus module 4, the "logic and reasoning" group — how to make a model show its working so a mistake is visible in the process rather than only in the result.
Module 4 — Basic; agentic RAG — N8N and Full.
---
What to read next
[How to check what a model tells you](/en/blog/how-to-verify-ai-output) — the protocol in general.
[AI for lawyers](/en/blog/ai-for-lawyer) — the profession where this is critical.
[Finding trends](/en/blog/finding-trends) — another use for precise search.
[Analysing a competitor by figures](/en/blog/competitor-analysis-by-numbers) — verifiable data instead of impressions.
[Best open LLMs 2026](/en/blog/open-source-llm-guide) — the full guide to open models.