Perplexity vs Claude vs ChatGPT for Research: Who Actually Cites Real Sources

I run client research weekly. Vendor comparisons. Regulatory updates. Market sizing. Competitive scans. All three tools claim to be the research tool. They are not equivalent.

I tested the same five real questions in each tool. Here's what I found.

What I tested

Five questions across categories I actually research:

1. "What FINRA rules apply to AI-generated marketing materials in 2026?" 2. "Compare pricing of Snowflake vs Databricks vs BigQuery for a 5TB warehouse" 3. "How did the EU AI Act amendments in late 2025 change vendor classification?" 4. "Which CRM tools have native MCP server support as of May 2026?" 5. "What's the realistic per-token cost of Claude Opus 4.7 in production right now?"

Three of these require recent data. Two require synthesis across multiple sources. All five would burn 30-60 minutes of manual research.

How each tool handled them

Perplexity returned answers fastest and with the most citations. It's the only one that consistently linked to specific source pages, not just "according to industry reports." On the FINRA question it pulled the actual rule text and the most recent guidance memo. On the EU AI Act question it pulled commission press releases and three analyst summaries.

It also confidently fabricated one citation on question 4 (claimed a CRM had MCP support that I verified does not). When I clicked the source it linked to a Reddit thread that did not say what Perplexity said it said.

Claude (with web access on) returned the most careful answers. It said "I don't have current information on..." for the per-token Opus pricing question (correctly — pricing changes monthly and the API page is the only ground truth). On questions where it did answer, it was more cautious about confidence and clearly distinguished what it had verified from what it was inferring.

It cited fewer sources but the citations were more accurate per click. No fabricated citations across the five.

ChatGPT (with web search on) returned answers that were structurally good but the citations were often summaries of summaries. It would cite an analyst report that cited the original ruling, instead of linking to the original ruling. When I asked it to find the primary source it eventually did, but only after a follow-up.

On the EU AI Act question ChatGPT gave outdated information twice (the late-2025 amendments) and only corrected when I prompted "verify with sources from the last 6 months."

The actual differences

Perplexity is a search tool with an LLM on top. It's optimized for getting you to source pages fast. The summaries are summaries — you should treat them as a starting point, not an answer. Click the citations. Read the actual pages.

Claude is an LLM with optional web access. It's optimized for thinking carefully about what it knows and doesn't know. The answers are more synthesized and less link-dense. It treats your question as a research prompt, not a search query.

ChatGPT is a productivity tool with optional web access. It's optimized for giving you an answer that feels complete. The completeness is sometimes a feature and sometimes a problem (when it confidently fills gaps with plausible-sounding inferences).

Where each one breaks

Perplexity breaks on questions that require synthesis across many sources. It'll give you a quick answer with citations but the synthesis is shallow. For a real strategy memo you'll outgrow Perplexity in 30 minutes.

Perplexity also breaks on fast-moving topics if its index is stale. It's better than ChatGPT here but worse than reading the source directly.

Claude breaks on questions that require very recent (last week or two) information if web access is off. With web access on, it's the most reliable. Without, it'll tell you it doesn't know.

ChatGPT breaks on questions where the right answer is "I don't know." It'll generate confident wrong answers more often than the other two. The fluency is the problem.

My actual workflow

For a new research question I open Perplexity first to get the lay of the land and the primary sources. 5 minutes to scan.

I open the primary sources directly. Read them.

If I need synthesis I paste the source content into Claude (or use Claude's web access on a specific URL) and ask for the analysis. Claude's strength is reasoning across pasted content.

ChatGPT I use mostly for shaping output — turning the research into a deck or a memo with a specific tone. It's the best for prose shaping. I do not use it as the primary research tool.

The one that matters for client work

For client-facing research, Claude is the safest default. It's the one most likely to admit uncertainty. The cost of confident wrongness in a client memo is higher than the cost of admitting "I'd need to verify this."

For internal research where speed beats accuracy, Perplexity is the fastest path to the primary sources.

ChatGPT is the worst default for research and the best default for turning research into output. Use it second, not first.

What changes this

Both Claude and ChatGPT are catching up on the search side. Perplexity is catching up on the reasoning side. The gap is narrowing every quarter.

The thing that won't change is the underlying philosophy. Perplexity will keep being source-first. Claude will keep being careful-first. ChatGPT will keep being fluent-first. Pick the one whose philosophy matches the work.

My one rule

Never cite an AI tool as the source. The tool is the bridge to the source. The source is what you cite. If you can't get to a citable primary source, you don't have an answer yet.

That rule has saved me from publishing several confidently wrong claims. Adopt it before you need it.