ChatGPT Research Depth Doubled: Multi-Query Fan-Out and Brand Citations

ChatGPT is reported to research an answer in more depth than it used to: it breaks a question into several sub-queries, consults, by those reports, roughly twice as many pages, and then makes targeted follow-up queries against the domains it has discovered. For brands the consequence is that being found is no longer enough. A brand's own pages are increasingly the place where a claim is checked before it is cited. This article describes the discover-then-verify pattern, what it appears to change for brand-owned domains and for intermediaries, and how to diagnose and fix a loss of citations.

A note on evidence. The change discussed here is an August 2026 shift in ChatGPT's research behaviour. The size of the effects people have reported for it comes from observation of the product from outside. Quasarank did not run these measurements and has not verified them, so apart from the one headline figure the title is named for, the reported doubling of the pages consulted, this article describes direction and mechanism and does not repeat the magnitudes.

From Single-Pass Retrieval to Multi-Query Fan-Out and Verification

Earlier generative engines stuffed the context window in a single pass: retrieve once, then write. An agentic engine splits the work in two. It first explores broadly, then verifies narrowly. When query classification detects a commercial comparison or another consideration-heavy intent, an intent-routing step expands the root query into several sub-queries, so the number of consulted documents per answer rises. The sub-queries run in parallel across a lexical index and a dense vector index, and the candidates are merged.

A Discover-Then-Verify Protocol

[Phase 1: Intent decomposition and fan-out] -> [Phase 2: Hybrid retrieval] -> [Phase 3: Domain-targeted verification] -> [Phase 4: Reranking and attribution]

1. Phase 1: Intent decomposition and multi-query fan-out. The agent reads the prompt $q$ and assesses how uncertain the answer is. For commercial or complex consideration intents, it expands $q$ into $K$ heterogeneous sub-queries $\{q_1, q_2, \dots, q_K\}$. More queries are issued per session than before.

2. Phase 2: Neural generative retrieval. Each sub-query runs against an inverted lexical index and against a quantised Hierarchical Navigable Small World (HNSW) vector index. The candidate set $\mathcal{C}$ is larger than under single-pass retrieval.

3. Phase 3: Domain-targeted verification. Broad retrieval identifies authoritative entities, and the engine then issues follow-up queries restricted to a specific domain with the site: operator, for example site:example.com, or to an institutional source. Reported behaviour is that this kind of site-targeted sub-query went from rare to a visible share of all sub-queries. For some top-of-mind entities the model may go straight to the domain from what it already knows, without a discovery query first.

4. Phase 4: Reranking and attribution. Candidate passages are fused with reciprocal rank fusion (RRF), which uses only rank positions, and may be adjusted with click-feedback signals. In the formula, $r_m(d)$ is the rank of document $d$ in channel $m$, $w_m$ is the channel weight and $k$ is a smoothing constant, commonly set to 60:

   $\text{RRF}(d) = \sum_{m \in \mathcal{M}} \frac{w_m}{k + r_m(d)}$

   The engine then writes the answer with inline citations. Citation concentration across brands can be summarised with the Herfindahl-Hirschman Index, where $s_i$ is the citation share of brand $i$ across a set of evaluation queries:

   $\text{HHI} = \sum_{i=1}^{N} s_i^2$

The channel weights are not published, so the formula is a conceptual model for reasoning about the pipeline and not a description of any one engine.

A conceptual objective for the fan-out step is to choose sub-queries that together add the most information about the candidate corpus $\mathcal{C}$, while penalising sub-queries that repeat each other. In that expression, $\mathcal{I}$ is the information a sub-query adds, $\mathcal{S}_{\text{redundancy}}$ measures overlap between sub-queries, and $\lambda$ trades one against the other:

$\mathcal{Q}_{\text{fanout}}(q) = \arg\max_{\{q_1, \dots, q_K\}} \sum_{k=1}^K \mathcal{I}(q_k; \mathcal{C}) - \lambda \cdot \mathcal{S}_{\text{redundancy}}(q_k, q_{\setminus k})$

What Changed for Brand-Owned Pages and Intermediaries

Depth does not scale equally across intents. Commercial comparison queries are reported to trigger the largest increase in the number of sources consulted, because competing brand claims call for several angles of verification. Informational queries follow, and transactional queries show the smallest increase of the three.

Two quantities have to be kept apart when reading a result of this kind. Sources retrieved are the pages the research drawer discloses, and inline citations are the links in the written answer. Consulted pages are reported to have risen faster than inline citations, which is consistent with the model compressing its context rather than citing every page it read.

The reported pattern for who benefits is asymmetric.

Brand mention rate across answers is reported to have moved only slightly, so the larger effect is on which pages are cited, not on how often a brand is named. The shift also appears to favour category leaders that have authoritative structured data, which makes visibility more concentrated rather than more evenly spread.

Measuring Citation Concentration and Displacement

The Herfindahl-Hirschman Index above gives a single figure for concentration: the sum of squared citation shares. When the index rises across a set of benchmark queries, citations are consolidating among fewer brands, a winner-take-most dynamic. It is a useful way to check whether deeper research helps challengers or entrenches leaders in your own category.

To tell whether a change in which brands appear is real, compare it with ordinary turnover. Citation lists reshuffle from one window to the next even when nothing changes. A null baseline measures that churn between two windows with no known change, and the volatility that is attributable to a real shift is the turnover observed in the test window beyond that baseline. If the top-brand retention in the test window is lower than the null baseline suggests, the displacement is genuine and not noise.

This is also a reason to measure your own category with a fixed prompt set over time. Background on how ChatGPT sources its answers is in Where Does ChatGPT Get Its Data, and on how a query is split into sub-queries in Query Fanout Architecture in AEO.

Troubleshooting Own-Domain Citation Loss

When a generative engine runs a site-directed sub-query, its crawler evaluates the raw HTTP response, not the browser-rendered page. A single-page application that returns an empty HTML shell without pre-rendering can lead the verification agent to log a timeout and drop the entity. Server-side rendering or pre-rendering that returns complete semantic HTML, with sensible cache validators such as ETag and Last-Modified from RFC 9110, restores a page that a crawler can traverse.

The vocabulary of this domain is multi-query fan-out, the discover-then-verify paradigm, neural generative retrieval, vector quantisation, HNSW, reciprocal rank fusion, the Herfindahl-Hirschman Index, null baseline normalisation, generative engine optimisation (GEO), answer engine optimisation (AEO), retrieval-augmented generation, sub-query decomposition and domain-targeted verification.

An Own-Domain Citation Recovery Protocol

[Phase 1: Verification ingestion audit] -> [Phase 2: Entity disambiguation] -> [Phase 3: Context-window hardening]

1. Phase 1: Verification ingestion audit and edge delivery. Replay site-restricted queries with a synthetic crawler and confirm that your edge returns the full semantic HTML quickly, with working ETag and Last-Modified headers. Remove client-side hydration dependencies for core product specifications, comparison tables and compliance pages, using server-side rendering or incremental static regeneration.
2. Phase 2: Semantic graph disambiguation. Audit existing schema markup for entity ambiguity, where a brand name collides with another entity, and add a W3C JSON-LD 1.1 @graph that anchors the brand with unambiguous canonical identifiers through sameAs links.
3. Phase 3: Context-window attention hardening. High-density facts buried under navigation menus or long paragraphs can be pruned when an engine compresses context. State technical claims as short atomic fact capsules, roughly 30 to 70 words, with a comparison table immediately under the heading.

A conceptual score for citation likelihood combines authority, verification and schema features with a truncation penalty, where $\mathbf{w}$ is a weight vector, $\phi$ are feature scores and $\theta_{\text{truncation}}$ is a penalty for context truncation:

$\mathcal{P}_{\text{citation}}(d) = \sigma\left( \mathbf{w}^T \left[ \phi_{\text{authority}}(d), \phi_{\text{verification}}(d), \phi_{\text{schema}}(d) \right] - \theta_{\text{truncation}} \right)$

Structured Data and Institutional Signals

Engines that prefer verifiable primary sources reward pages that can be checked mechanically. The foundation is a W3C JSON-LD 1.1 @graph using the Schema.org vocabulary, with an unambiguous @id for each entity, and HTTP behaviour that follows RFC 9110. The payload below illustrates the structure. All hosts and names are placeholders.

{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "TechArticle",
      "@id": "https://www.example.com/research/chatgpt-research-depth#article",
      "headline": "Example Headline for a Research Article",
      "inLanguage": "en-US",
      "mainEntityOfPage": "https://www.example.com/research/chatgpt-research-depth",
      "datePublished": "2026-08-15T08:00:00Z",
      "dateModified": "2026-10-04T12:00:00Z",
      "author": {
        "@type": "Organization",
        "@id": "https://www.example.com/#organization",
        "name": "Example Research Group"
      },
      "publisher": {
        "@type": "Organization",
        "@id": "https://www.example.com/#organization"
      },
      "about": [
        {
          "@type": "Thing",
          "name": "HTTP Semantics",
          "sameAs": "https://www.rfc-editor.org/rfc/rfc9110"
        }
      ]
    },
    {
      "@type": "Organization",
      "@id": "https://www.example.com/#organization",
      "name": "Example Corp",
      "url": "https://www.example.com",
      "sameAs": [
        "https://www.example.com/entity/example-corp"
      ]
    }
  ]
}

Institutional and regulatory domains are also reported to be cited more after the change. A reasonable reading is that an engine verifying a commercial claim prefers a primary, checkable source over an unverified secondary summary. A conceptual authority score might add a standards-compliance indicator, a similarity to institutional language and a measure of provenance depth, with weights $\alpha$, $\beta$ and $\gamma$:

$\mathcal{S}_{\text{authority}}(d) = \alpha \cdot \mathbb{I}_{\text{standards}}(d) + \beta \cdot \text{Sim}_{\text{cosine}}(\mathbf{v}_d, \mathbf{v}_{\text{institutional}}) + \gamma \cdot \text{Depth}_{\text{provenance}}(d)$

For the product side of this change, see How Do I Show Up in ChatGPT and How to Increase AI Search Citations.

References

• JSON-LD 1.1 — W3C, 2020. https://www.w3.org/TR/json-ld11/
• HTTP Semantics (RFC 9110) — IETF, 2022. https://www.rfc-editor.org/rfc/rfc9110