Organic prompts share more of a source list across engines than branded ones. For ChatGPT, it flips.

An earlier post split the cross-engine domain overlap rate by category and found it essentially flat. It never checked the other obvious split: does an engine read a more shared source list when the prompt already names a brand ("Asana vs Jira") versus when it does not ("What project management software do you recommend?"). We recomputed the same domain-panel-pair overlap calculation, this time split by promptType across the 270 live panels.

Organic prompts pull a more shared source list, blended across engines

Same method as the earlier posts in this series: for each of the 270 live panels, we pulled every citation URL from every real, non-error response, normalized it to a hostname, and excluded Gemini's opaque vertexaisearch.cloud.google.com redirects. This time we split responses first by promptType (branded or organic, from panel.json, with the same brand-name-in-text fallback rule used for two legacy panels lacking the field) before building each panel's per-engine domain sets.

Both response counts (2,048 branded and 6,616 organic, summing to the full 8,664-response dataset) match the branded/organic split published in an earlier post exactly, confirming the same dataset re-joined on promptType, not a new sample.

The per-panel overlap share tells the same story with less noise: branded prompts have a median overlap share of 16.1% (P25 11.8%, P75 22.4%) across 261 panels with at least one branded citation, while organic prompts hold a median of 24.2% (P25 21.3%, P75 26.3%) across all 270 panels. The organic median lands almost exactly on the 24.0% blended figure published earlier, which makes sense since organic prompts are the large majority of the dataset. Branded prompts run a consistently thinner shared list, not a noisier one; the interquartile range is about the same width for both.

By category, organic runs higher than branded in project management (23.1% vs 17.5%), HR software (24.8% vs 19.2%), and customer support (24.0% vs 17.5%). The direction holds in all three categories, and the gap size barely moves between them, ruling out a single category driving the pattern.

But three of the six engine pairs invert that pattern completely

The blended number hides a split that only shows up once you look at individual engine pairs. We computed pairwise Jaccard overlap (shared domains divided by combined domains) separately for branded and organic prompts, for all six engine pairs.

The three pairs that do not involve ChatGPT (Gemini-Perplexity, Claude-Perplexity, Claude-Gemini) all show meaningfully higher overlap on organic prompts, from a fifth higher to nearly double. That is what drives the blended headline number. The three pairs that do involve ChatGPT (Claude-ChatGPT, Gemini-ChatGPT, ChatGPT-Perplexity) run the other way: their overlap is higher on branded prompts, by 37% to 61%. Every ChatGPT pairing sits below every non-ChatGPT pairing on organic prompts, and above at least one non-ChatGPT pairing (Claude-Perplexity, at 7.9%) on branded prompts. ChatGPT's reading list converges with the other three engines more when a prompt names two specific brands than when it asks an open recommendation question, the opposite of what the other three engines do with each other.

A caveat: ChatGPT's panel coverage is thinner in both splits

ChatGPT's real-world API error rate is the highest of the four engines, documented in earlier posts, and that shows up here as coverage: ChatGPT had usable citation data in 125 of 270 panels on branded prompts and 130 of 270 on organic prompts, versus 243-268 for Gemini, Perplexity, and Claude across both splits. A thinner, burstier sample of successful ChatGPT responses in both splits means fewer chances for its domains to overlap with anyone's, in either direction. It does not explain the sign flip, since the same coverage limitation applies to both the branded and organic ChatGPT numbers, and the branded number is higher despite branded prompts having slightly worse ChatGPT coverage (125 vs 130 panels) than organic.

One more caveat on the totals: unlike the earlier category split, which partitions panels cleanly (each panel belongs to exactly one category, so the category totals summed exactly to the blended post's published figures), promptType splits responses within a panel. A domain cited under both a branded and an organic prompt in the same panel counts as a separate domain-panel pair in each split. That is why 10,947 plus 22,934 (33,881) runs about 10% above the blended post's 30,912 total, while the overlap counts (1,934 plus 5,394 equals 7,328) land within 2% of the blended 7,426. The gap is the expected effect of double-counting shared prompt-type domains, not a data problem; the branded and organic figures reported above are each internally consistent and independently verified against the response-count baseline.

Why this matters for a visibility strategy

The cross-engine overlap posts in this series have consistently found that a source win on one engine is not a source win everywhere. This split adds a wrinkle specific to prompt framing: for three of the four engines, chasing the same generic "best of category" roundup articles is the more efficient path to multi-engine coverage, since those pages are what Gemini, Perplexity, and Claude already read in common on organic prompts. For ChatGPT specifically, that logic reverses. Its shared ground with the other three shows up more on branded, head-to-head comparison content than on generic buyer-intent roundups, meaning a ChatGPT-facing source strategy built purely around ranking in "best PM software" listicles is chasing the wrong overlap.

Methodology

Data drawn from the same 270 live, search-grounded audit panels used in the original cross-engine overlap post (173 project management, 58 customer support, 39 HR software brands), each run across four AI engines: ChatGPT with web search, Gemini with grounding, Perplexity Sonar, and Claude with web search. For every real (non-error) response, we pulled the citation URL list, normalized each to its hostname, and excluded Gemini's vertexaisearch.cloud.google.com redirect wrapper, matching the normalization used throughout this series. Each response was classified branded or organic using panel.json's promptType field, with a brand-name-in-prompt-text fallback for two legacy panels (teaser-gladly-live, teaser-kustomer-live, 43 responses) that predate the field, the same fallback used in an earlier branded-vs-organic post. For each panel, and separately within each promptType, we built the set of distinct cited domains per engine, then measured domain by domain how many engines cited it, and pairwise, the Jaccard similarity between each pair of engines' domain sets pooled across every panel where both engines had citation data in that split. Response counts (2,048 branded, 6,616 organic, 8,664 total) matched an earlier post's published branded/organic split exactly. Numbers recomputed fresh from results.jsonl and panel.json this session. No development-rail or fixture data is included; all responses came from live engine calls. Data collected June-July 2026.