Perplexity vs ChatGPT Search vs Gemini Deep Research: A Test You Can Run

Disclosure: no affiliate links in this post. All three tools have free entry points, though limits differ.

Nobody can tell you which of these three is best for your questions, and any post that does without showing its test is guessing. So this one does two things. It reports what the vendors' own pages said on 2026-10-23, including which pages I could not read. Then it gives you a five-question test, a scoring sheet and a small script, so you can produce your own answer. I did not run any of the three tools for this post. All scores below are empty templates.

What do the vendor docs say about each tool?

Only one of the three vendors' pages was readable for me.

One useful point from the readable page: Gemini's report mode is a different job from a quick search. Comparing a 10-minute report with a 20-second search answer is unfair unless you score for the job each one claims to do. The test below handles that by asking one question at two depths.

For a published head-to-head with a different set of tools, see Perplexity vs Claude vs ChatGPT for research. This post is not a rerun of it; it gives you the method.

How do you design a fair five-question test?

Pick questions where you can find the truth yourself. Use these slots:

1. A fact with a recent change. Something that changed in the last six months in a field you follow. You know the old and new answer.
2. A numeric lookup. A figure that lives on one official page (a fee, a limit, a date). Keep the URL private until scoring.
3. A comparison. Two things in your field, "which for whom". You need an opinion to judge it against, ideally from a primary source.
4. A local or India-specific question. Rules, prices or services that depend on country. Generic answers will show up here.
5. A trap. A question whose premise is slightly wrong. See whether the tool corrects you or plays along.

For each tool, use the same wording. For question 3, run it once in the quick search mode and once in the report mode (Deep Research or Research) to see what the extra minutes buy.

Rules to keep it fair:

• Same day, same prompts, logged in or out consistently. Record the plan you used and its model label if shown.
• Do not edit prompts between tools.
• Save the full answer and every link before you score.
• Score blind if a friend can strip the tool names.

How do you score the answers?

Check every cited claim against the cited page, not against your memory. Score each answer 0, 1 or 2 on four criteria:

This is a template. The weights are my judgement, not a standard; change them to fit your priorities. The weights are chosen so unsupported claims hurt most.

Put results in a CSV with the columns tool,question,claims_supported,sources_real,answers_question,recency, one row per answer. The script below sums weighted scores per tool.

import csv
import sys
from collections import defaultdict

WEIGHTS = {"claims_supported": 3, "sources_real": 2, "answers_question": 2, "recency": 1}

totals = defaultdict(float)
counts = defaultdict(int)
with open(sys.argv[1], newline="") as f:
    for row in csv.DictReader(f):
        tool = row["tool"]
        score = sum(WEIGHTS[k] * int(row[k]) for k in WEIGHTS)
        totals[tool] += score
        counts[tool] += 1

max_per_q = sum(WEIGHTS.values()) * 2
for tool in sorted(totals):
    pct = 100 * totals[tool] / (max_per_q * counts[tool])
    print(f"{tool}: {totals[tool]:.0f} / {max_per_q * counts[tool]} ({pct:.0f}%)")

I ran this on a four-row file of made-up numbers (tools named ToolA and ToolB, two questions each) only to check the arithmetic. Output:

ToolA: 26 / 32 (81%)
ToolB: 19 / 32 (59%)

Those numbers are dummy data and say nothing about any real product. Replace the CSV with your own scores.

What do you do with the result?

• If one tool wins on claims supported but loses on speed, use it for questions where being wrong costs you, and the faster one for orientation.
• If the trap question fooled a tool, do not use it unsupervised for anything where your own premise might be off.
• If the India-specific question exposes generic answers across all three, add the official source URL to your prompt and re-run once to see if it improves.
• Keep the CSV. Run it again after a major model or plan change; vendors change things often.

The honest limit of a five-question test is that it is small. It tells you which tool suits your kind of question, not which is better overall. Ten questions per tool would be steadier, and it costs only another half hour.