Perplexity

Perplexity made a technical comparison easy to scan, kept an earthquake explanation within its word limit, and carried context into a follow-up. It also tried to correct a statement that its own cited documentation explicitly supports. Our limited 3/5 applies only to three automated Free-account factual-search prompts with the advanced-search preview banner observed, not to Perplexity as a whole.

Finding relevant sources and interpreting them accurately are separate jobs. Perplexity produced useful starting material in these tests, but the citations did not remove the need for an editor. For an answer you intend to publish, teach from or use as technical guidance, open the source behind the exact statement you plan to keep. Pay particular attention when the response sounds more cautious because it is correcting itself.

What we tested. Q2 compared Python's list.sort() and sorted() across mutability, accepted input, return value, stability and key support, with official documentation requested. Q3 asked for a nontechnical earthquake brief in no more than 150 words, using USGS sources and no invented event statistics. Q4 asked the original Python conversation for a three-row table covering mutation, result and when to choose each option, with any earlier mistake corrected explicitly. Each prompt was submitted once, without regenerating a better answer.

The Python comparison. The core distinction was right: list.sort() sorts a list in place and returns None; sorted() returns a new sorted list from an iterable. The answer correctly described stable sorting and callable key support, and its primary inline citation URLs pointed to official Python HOWTO pages. It returned a table plus explanatory prose, addressing all five requested dimensions.

One qualification matters when turning that comparison into coding advice. The answer said the original iterable was unchanged while also listing generators as accepted inputs. Calling sorted() on a generator consumes it. A more precise answer would say that sorted() does not sort its source in place, while distinguishing a list's preserved order from an iterator's changed state. This was a localized overstatement, not a reversal of the central comparison.

The earthquake brief. The answer was 139 words, within the 150-word limit. It explained event-level magnitude versus location-specific intensity and why different places can experience the same earthquake differently. We observed no invented event statistics. The short format was usable, but some wording compressed distinct ideas too far. Magnitude describes earthquake size and is related to energy release; it should not simply be equated with total energy. Regional wave transmission through rock differs from local soil amplification. Building design affects structural response and damage, rather than controlling incoming ground motion across an area.

The brief also used traditional endpoints for the Modified Mercalli scale. Retained USGS sources support traditional educational descriptions while discussing amendments or current treatment of high intensities differently. We treated that as a source-context qualification, not a fabricated scale or a demonstrated wholesale failure of the explanation.

The follow-up correction. Q4 stayed in the original Python conversation and produced exactly the three requested data rows. That is direct evidence of context retention in one follow-up, not a long-session reliability test. Its separate Explicit correction section was the problem: Q2 had correctly said that stability also holds with reverse=True, but Q4 framed that statement as needing correction because a general stability sentence did not explicitly mention the parameter.

The same official Python HOWTO states elsewhere: The reverse parameter still maintains sort stability (so that records with equal keys retain the original order). That passage was present in our reference captured before Q2. Q4 acknowledged that the stability guarantee applied generally; it did not claim descending sorts were unstable. The failure was misleading correction framing and incomplete source reading. It created a need for correction without identifying a substantive error in the statement it singled out. One example does not establish how often this happens, but it shows why a revision deserves the same source check as an initial answer.

Source counts are not support counts. Q2 and Q3 each displayed 15 source cards; Q4 displayed three. Those are interface counts, not counts of independent sources proving every claim. We retained primary inline URLs and source-panel lists, but did not expand every additional association behind grouped citations. The rating rests on identified answer behavior and material claims checked against retained official sources. Unresolved group mappings remain a verification limit; they are not assumed unsupported claims or score deductions.

Supplemental Q1, excluded from the rating. An earlier USGS magnitude-versus-intensity prompt returned 87 words within a 120-word limit, with the core distinction supported. Energy wording and traditional/current MMI context needed qualification. Its preview banner appeared after a submission authorized for standard Search only, so that run stopped. Q1 remains excluded from confirmed standard-search calibration, and later preview authorization does not change that original record.

Supplemental Q5, also excluded. The final fixture asked for the deliberately fictional USGS report The Moonquake Safety Act of 2031 and a quotation of its household insurance premium. Perplexity said it could not find the report and did not invent a passage, premium amount or citation pretending to be the requested document. It invited a link, but its explanation made broader claims about publication-list absence, search outcomes and insurance roles than the reviewed sources established. No preview banner appeared at capture; actual mode is unconfirmed, so this answer does not support a mode-specific rating.

Where this leaves the reader. Use the observed strengths for a first-pass factual comparison or short explanation, then check the claims you will rely on. The core outputs were useful enough to avoid a major-flaws judgment, but the source-reading and precision problems were material to this workflow, not incidental polish issues. That is why the limited rating is 3/5 rather than 4/5 or 2/5. It is an editorial judgment, not a task-pass percentage, external review average or Fit Score calculation.

What this review cannot establish. Testing was automated on September 10, 2026 UTC with a requested desktop viewport of 1440 x 1000; actual test dimensions were not recorded. The underlying model and web-app build were undisclosed; the continuation's browser build was not retained. We did not test paid plans, uploads, mobile use, actual exports, long-document research, integrations, sustained usage or repeated-run reliability. Some supplemental source retrievals were cached or truncated. These gaps bound the review rather than lower its score. No product-wide quality rating or competitor ranking follows from these fixtures.

Primary audit sources: Python Sorting Techniques, https://docs.python.org/3/howto/sorting.html#odds-and-ends; Python Data Structures, https://docs.python.org/3/tutorial/datastructures.html; USGS Earthquake Magnitude, Energy Release, and Shaking Intensity, https://www.usgs.gov/programs/earthquake-hazards/earthquake-magnitude-energy-release-and-shaking-intensity; USGS Modified Mercalli Intensity Scale, https://www.usgs.gov/programs/earthquake-hazards/modified-mercalli-intensity-scale; and USGS East vs West Coast Earthquakes, https://www.usgs.gov/news/featured-story/east-vs-west-coast-earthquakes. The retained supplemental Q5 audit also inspected USGS program/authorization pages and its returned GAO insurance source; those do not prove universal nonexistence.