Which AI engine repeats its sources most depends on the category. Perplexity's repeated links are mostly a project management habit.

An earlier post on this blog found that Gemini repeats a cited domain in 65.8% of its responses, and that Perplexity's repeats are mostly the identical link listed twice. Both numbers were blended across every category we audit. We re-split the same 8,664 real responses by category to check whether those engine traits hold everywhere or belong to one kind of buyer question.

The engine ranking changes in every category

For each real, non-error response with at least one citation, we checked whether any domain appeared more than once in the citation list, excluding Gemini's opaque vertexaisearch redirect wrappers. The blended ranking from the earlier post was Gemini 65.8%, ChatGPT 44.4%, Claude 31.6%, Perplexity 30.5%. Split by the category of the audited brand, that ranking does not survive intact in any of the three.

The table shows the share of responses that repeat at least one domain, with the number of responses in each cell in parentheses. In project management, Gemini leads at 71.1% and ChatGPT is last at 27.8%. In customer support and HR software, ChatGPT is first, and Perplexity is last by a wide margin, 4.8% and 18.5%. ChatGPT alone spans 27.8% to 77.7% depending on the category. Claude is the steadiest of the four, staying between 23.7% and 35.4%.

The gap between raw citation count and distinct domains moves the same way. Gemini's raw count overstates its distinct domains by 16.4% in project management and 11.9% in HR software. ChatGPT's overstates by 25.2% in HR software and 7.0% in project management. A single per-engine number for this trait hides a spread that is often wider than the difference between engines.

Perplexity's identical-link repeats are a project management pattern

The earlier post split repeats into two kinds: a different page on a domain already cited, and the identical URL listed twice. It found that most of Perplexity's repeats were the second kind. Splitting that by category shows where they come from.

This is not a handful of panels driving the number. In project management, 167 of the 173 audited brand panels contain at least one Perplexity response whose only repeats are identical links. It also does not depend on how the prompt was phrased: the rate is 26.2% on organic prompts (338 of 1,291) and 33.2% on branded head-to-head prompts (130 of 392). Something about how Perplexity assembles citations for project management queries produces the duplicates, and customer support queries almost never do. We can measure the pattern but not explain its cause from response data alone.

Project management is 1,683 of Perplexity's 2,358 responses, 71% of its total. That weighting is why the blended figure read as a general Perplexity trait when it is closer to one category's behavior.

A caveat on ChatGPT's HR cell

ChatGPT's 77.7% in HR software is the highest single cell in the table, and it rests on 148 responses from 15 of the 39 HR panels. ChatGPT's live calls errored at a high rate during collection, mostly quota limits, and those errored responses are excluded. Its customer support cell, 528 responses across all 58 panels, and its project management cell, 503 responses across 57 of 173 panels, are better covered. We would read the HR figure as a strong direction and not a precise rate. The main finding does not depend on it: project management (27.8%) and customer support (50.9%) alone differ by 23 points on those two samples.

What this means for reading a citation list

If you are using an audit's citation table to judge how many independent sources an engine draws on, the right comparison is within a category, not against a blended average. A Gemini answer to a project management question is likely to include several pages from the same site. A Perplexity answer to a customer support question almost never repeats anything. Counting distinct domains instead of raw citations removes most of that variation, and any tool that ingests these lists needs URL-level deduplication for Perplexity in project management specifically.

Methodology

Data drawn from the same 270 live, search-grounded audit panels used in the earlier repeat-citations post (173 project management, 58 customer support, and 39 HR software brands), each run across four AI engines: ChatGPT with web search, Gemini with grounding, Perplexity Sonar, and Claude with web search. We used every successful (non-error) response, branded and organic prompts together, 8,664 responses in total, with category taken from each panel's own configuration. Category totals (5,589, 1,937 and 1,138) and the blended per-engine repeat rates (Gemini 65.8%, ChatGPT 44.4%, Perplexity 30.5%, Claude 31.6%) match the earlier posts exactly, confirming this is the same dataset re-split and not a new sample. Domain is the hostname of each citation URL after excluding Gemini's vertexaisearch redirect wrappers. A response has a "distinct page, same domain" repeat if two citations share a hostname and differ in path, and an "identical URL only" repeat if it has a repeated hostname but every repeat is the same URL (query string removed). The gap between raw and distinct counts is the raw count minus the distinct count, divided by the raw count. Results describe correlation within this dataset and are not a claim about why an engine builds its citation list the way it does. No development-rail or fixture data is included; all responses came from live engine calls. Data collected June to July 2026.