Perplexity vs ChatGPT

Perplexity is built as a pipeline with a writing step on the end. Your question becomes a search, the search returns a set of pages, and a language model writes a summary of those pages. The sources are not a courtesy added afterwards; they are the material the answer was made from, which is why they can be listed at all.

ChatGPT is the other way round. It is a general model wrapped in a chat interface, and searching the web is one of the things it can reach for when it judges that it should. The point is not that it cannot search — it can — but that retrieval is a decision taken quietly mid-answer, and the finished text reads exactly the same either way. Recalled from training and fetched thirty seconds ago arrive in the same confident voice.

So the real difference is disclosure, not intelligence. With one tool you know roughly where the words came from. With the other you usually do not, and the tone gives you nothing to go on. Everything else on this page follows from that.

Grounding an answer in retrieved text is a genuine constraint. A model writing about pages in front of it invents less than a model writing from memory — that is the whole idea behind retrieval-augmented generation. It is a constraint, though, not a guarantee. Reducing invention is not the same as removing it.

A citation proves three useful things: a page exists, it was retrieved, and it was in view when that sentence was written. That is meaningfully more than a plain chatbot offers, and it kills the single worst failure mode of AI research — the reference that sounds perfect and turns out never to have been published.

Here is the longer list of what it does not prove.

• That the source is any good. A retrieval system ranks pages; it does not grade them. A content farm rewriting a press release is cited as readily as a peer-reviewed paper, and both look identical in a footnote.
• That the source says what the summary says. Misreading happens, and when it does, the citation is still real. You get a correct link attached to an incorrect claim, which is far harder to spot than an obvious error.
• That the claim was the point. A sentence pulled from an aside, a caveat, or a paragraph describing what the author is about to argue against can end up presented as the source's conclusion.
• That it is current. Old pages rank well. A superseded figure or a repealed rule cites just as cleanly as a live one.
• That nothing was dropped. You see what was used, not what was retrieved and set aside. A source that contradicted the answer leaves no trace.

The uncomfortable part is that a visible citation changes reader behaviour in the wrong direction. It looks like verification has already happened, so people stop checking at exactly the moment the interface implies they need not. The cure is small and boring: open one link before you believe anything, and open every link before you repeat it. How to Fact-Check AI Output sets out which claims deserve that time first, and Everyday Research with AI is the lesson that turns it into a habit.

A summary has to pick a voice. Fluent prose needs one narrative, and two sources that contradict each other cannot both be that narrative. Something has to give, and what usually gives is the conflict: the tool resolves it silently and writes one calm paragraph where the truthful answer was that the question is open.

How it resolves the conflict is the part worth knowing. Repetition carries a lot of weight on the open web. A claim copied across twenty sites has more textual presence than one careful correction published once, so the majority view of the internet can beat the better-evidenced view of a single expert. Volume is not evidence, but it looks like consensus from the inside of a summary.

This matters most precisely where people most want a quick answer: health, law, money, anything under active research, anything political, and any comparison of two products whose users each insist the other is worse. In all of those, the shape of the disagreement was the information you needed, and it is the first thing a summary removes.

You can partly undo it by asking for it back. "Which sources here disagree, and about what exactly? Give me the strongest version of each position and say who holds it." Both tools answer that noticeably better than they answer a bare question, because you have asked for a structure that has room for conflict in it. The compare two options objectively template does the same job for decisions.

There is a test that settles most of these choices in a couple of seconds. Ask whether the answer has a truth condition outside the conversation — something in the world that makes it right or wrong, independent of whether it reads well.

If it does, favour the tool that shows its working. What year the rule changed, whether a product still supports a format, what a study actually found, what happened last month: these have answers, you may not know enough to notice a wrong one, and a link is worth more than a paragraph.

If it does not, favour the tool that writes better and holds more of your instructions. Rewriting your own paragraph, structuring an argument, naming a feature, explaining a concept you will recognise as right or wrong the moment you read it, generating code you are about to run — there is nothing to cite, and citations would only be decoration. Code is an interesting case: it comes with its own verification, because it either works or it does not. AI tools for developers covers where that holds and where it stops holding.

The interfaces push you into different rhythms, and this is more noticeable in daily use than any capability difference.

Perplexity suits a thread of narrowing questions. Ask broadly, read the answer, ask the sharper question the answer just revealed. Each step is a new retrieval, so the thread behaves like a research session: you are converging on what to read rather than producing anything. That is a real strength for orientation, and a poor fit for a long piece of work.

ChatGPT suits an accumulating session. You set the role and audience, paste in the material, get a draft, say what is wrong with it, and get a better one. The catch is drift: in a long conversation, constraints you set early quietly stop being applied, and the model does not announce that it has forgotten them. Restating the important ones every so often is not superstition, it is maintenance — the habit taught in Treat Your First Prompt as a Draft.

Put plainly: finding out is a series of short questions, and producing is one long session. Most people who dislike one of these tools were using it for the other job.

The shared weaknesses are more consequential than the differences, and neither tool is close to fixing them.

Both can be confidently wrong. Neither has a visible signal for guessing. A well-supported fact and an invention arrive in the same register, at the same speed, with the same steady tone — and sources do not remove this, they relocate it from invented facts to misread ones. AI Hallucinations Explained covers why this is structural rather than a bug awaiting a patch, and the hallucination entry is the short definition.

Neither reads for you. A summary is orientation. It tells you what the territory looks like and what to ask next; it does not put the source's argument in your head. For anything that matters you still click through and read, which cancels part of the time saving — and being honest about that is the difference between using these tools well and using them quickly.

Neither knows your situation. Your company, your documents, your codebase, your constraints: none of it is in there unless you supply it, every time.

Free limits arrive mid-task. Both have usage ceilings on their free tiers, and they tend to land at the least convenient moment, which is when you are deep in something.

Both are the wrong place for confidential material. Neither should receive anything your employer has not cleared for AI tools, and a research question can carry more sensitive context than it first appears to. What Not to Share with AI goes through where the line sits.

There is no winner here, and any page that names one is describing a moment rather than a rule. These are different kinds of product that happen to accept the same kind of question. Match them to the job.

• Studying an unfamiliar subject. Use the answer engine to map the territory and collect the reading list, then the assistant to have the hard parts explained at your level. See how to use AI for studying for the study habits that survive both.
• Writing something other people will read. Draft with the assistant, then take every factual claim in the draft to the answer engine one at a time. Drafting and checking are separate jobs and go badly when merged.
• Anything recent. The answer engine, or an explicit instruction to the assistant to search and show its pages. A model's training has an end date; the web does not.
• Code, structure, planning, rewriting. The assistant. There is nothing to cite and plenty to iterate on.
• A question where experts disagree. Either, but only if you ask for the disagreement explicitly. Left alone, both will hand you one tidy paragraph and hide the argument you were looking for.
• If you will only learn one. Pick by the majority of your week. Mostly making things: a general assistant, and ChatGPT vs Claude vs Gemini compares the main ones. Mostly finding things out: the answer engine.

One caveat that applies to every line above: features, limits and pricing on both products change often, and the official sites are the only authority on what is true today. Treat this page as a way of thinking about the choice, not as a specification.