ChatGPT's query fan-out increasingly narrows its background searches to a single named domain using the site: operator, and the domains it points at are overwhelmingly first-party. According to a July 2026 analysis by AI-visibility platform Peec AI, the site: operator appeared in about 43% of ChatGPT 5.6 (codename Luna) query fan-outs, up from roughly 0.004% in ChatGPT 5.5, and 84% of those site: domains were branded — the site of the product or company the question was about. This reference lays out what the fan-out actually targets, with every figure attributed and dated.
Last reviewed: 24 July 2026. All figures below are from Peec AI's analysis of ChatGPT 5.6 query fan-outs, published by researcher David Konitzny (22 July 2026), unless noted. It is a single-vendor, directional dataset from one analytics vendor's prompt sample, so treat the exact percentages as a snapshot of one model version and cite "Peec AI, July 2026" with the date when you reuse a figure. For the underlying mechanics of the fan-out step itself, see how query fan-out works.
How much more does ChatGPT 5.6 use the site: operator?
Dramatically more than the prior version. The site: operator restricts a search to one domain, and its use inside ChatGPT's fan-out jumped by orders of magnitude between versions. Peec AI's term-frequency comparison of the sub-queries each model generates:
| Term in fan-out sub-queries | ChatGPT 5.6 (Luna) | ChatGPT 5.5 |
|---|---|---|
site: prefix |
43.2% | 0.004% |
| Year mention (2025 / 2026) | 26.5% | 6.2% |
| "official" | 21.9% | 1.4% |
| "best" / "top" | 5.4% | 8.5% |
| "pricing" / "cost" | 3.0% | 0.51% |
| "review" | 1.1% | 1.02% |
| "vs" / "versus" | 0.03% | 1.30% |
The pattern: Luna leans heavily on domain-restricted, "official," and recency-dated sub-queries, and away from open "best/top" and "vs" comparisons. The fan-out is behaving less like a broad keyword sweep and more like targeted verification against specific, current, first-party sources.
Which domains does the site: operator target?
Mostly one domain at a time, and mostly the brand's own. Peec AI reports that in 88.66% of fan-outs using the operator, only a single site: domain appears; two or three domains show up in roughly 9% of cases combined. And the domains split sharply between first-party and secondary sources:
| Domain type | Share of site: domains |
What it means |
|---|---|---|
| Branded (first-party) | 83.63% | The model targets the product/company's own site |
| Non-branded (secondary) | 16.37% | Community, review, and institutional sources |
That 84%-branded finding is the mechanistic companion to ChatGPT's shift toward brand-name fan-out sub-queries: the model has often already decided which brand the question is about and points a site: probe straight at that brand's domain. The consequence for AI visibility is direct — if your first-party pages hide their facts behind JavaScript or images, the fan-out narrows to your domain and finds nothing citable.
What terms appear alongside the site: operator?
Product attributes lead. When Peec AI clustered the terms that appear with a site: domain, the most common were product specifications, followed by "official" and reviews:
Term cluster alongside site: |
Share of site: occurrences |
|---|---|
| Product specs (size, model, colour) | 11.72% |
| Official | 9.29% |
| Reviews | 4.71% |
| Support | 3.17% |
| Pricing | 3.14% |
| Recommendations | 0.88% |
The read: the operator is used to pull highly specific facts from a trusted domain — the exact spec, the official detail, the price — not to browse. This is why structured, extractable product facts on your own pages matter so much for the branded majority of these searches.
Which non-branded domains does the fan-out reach for?
Reddit, by a wide margin. Within the 16% of site: domains that are non-branded secondary sources, one platform dominates:
| Non-branded domain | Share of non-branded site: targets |
|---|---|
| reddit.com | 71.5% |
| trustpilot.com | 5.92% |
| linkedin.com | 2.34% |
| g2.com | 2.27% |
| capterra.com | 1.75% |
Reddit alone is about 12.28% of all site: operator domains — roughly one in eight. Grouped by category, the non-branded targets break down as Social & forums 76.6%, Review & comparison 10.9%, and Public institutions 4.1%. Within Social & forums, Reddit is 93.4% of the category; within Review & comparison, Trustpilot leads at 54.5% (then G2 20.9%, Capterra 16.1%). Note the important nuance from the cross-engine citation data: being the top retrieval target is not the same as being cited — ChatGPT rejects the vast majority of Reddit pages it pulls in.
Public institutions are more spread out (bbb.org 34.6%, fda.gov 22.8%, mayoclinic.org 17.7%, pubmed 8.6%), with arXiv appearing at about 7.1% of that category — echoing the silent-source pattern where research repositories are retrieved for context without always being cited.
Does the operator target whole domains or specific pages?
Whole domains, mostly. Peec AI found the site: operator points at the bare root domain in 86.19% of cases and at a specific path or URL in only 13.81%. But the intent differs by depth:
| Term cluster | Root-domain site: |
Domain + path site: |
|---|---|---|
| Product / specs | 12.37% | 32.52% |
| Pricing | 5.41% | 12.31% |
| Official / verification | 10.17% | 8.58% |
| Reviews / ratings | 7.27% | 0.55% |
| Customer service | 4.41% | 0.63% |
| Comparison | 1.84% | 0.05% |
Root-domain searches are used to identify a trusted source; path-level searches appear when the model already has a precise need, such as a specific product detail or price.
In plain terms: broad informational intents (reviews, official, service, comparison) skew to the whole domain, while product- and price-specific queries reach for a specific page. The fan-out is a structured retrieval strategy — first identify the trusted domain, then narrow to the exact fact.
How should you use this dataset?
As a directional reference, not a target to game:
- Cite the source and date. Attribute figures to "Peec AI, July 2026 (ChatGPT 5.6 analysis)" and hedge them as single-vendor and version-specific — Luna's behaviour can change with the next model release.
- Serve your first-party facts cleanly. With 84% of
site:targets being branded, your own pages must expose specs, pricing, and official details as server-rendered text the model can lift. - Know your non-branded destinations. For your category, the external validation the fan-out reaches for likely lives on Reddit, Trustpilot, G2, or Capterra — but weight that against whether each engine actually cites those sources.
- Measure retrieval and citation separately. Being targeted by a
site:probe is upstream of being cited, which is upstream of being recommended.
Tracking what the fan-out points at for your buyers' questions — and whether those retrievals turn into citations across ChatGPT, Gemini, Claude, and Google — is exactly what Buffy Intel is built to measure, engine by engine, over time.