case study chatgpt query fan out

Reverse-engineering ChatGPT’s query fan-out | A B2B case study

André Pitì

23 de December de 2025

Chat GPT’s query fan-out is a behind-the-scene process that breaks down a single user prompt into multiple queries to determine what to answer in the output.

In short, it means running a series of variations of the main query, to address the semantic meaning and intent of it, before blending relevant chunks into an answer.

We’ll publish other deep dives on this process, including its rationale, purpose, and how it changes across different LLMs.

Executive Summary and disclaimer

In this experiment for a client in the domain reselling industry, we conducted a repeatable observation over a high-buying intent prompt to determine what web information influences the answer.

At a surface level, what we did is:

  1. Explored ChatGPT’s process after a BoFU query to compare domain service providers, observing its grounded search, thinking and comparison process, down to the user answer
  2. Used GPT model 5.1 to aggregate patterns and spot evident gaps between the fanout sequence – results and the brand content we want to optimize for.
  3. Isolated candidate topics, content structures, and quick technical wins that we could tackle client’s brand and presented them as actionable recommendations, then:

Disclaimer for this SEO case study

This case study documents the outcome of a repeatable fanout observation experiment on a specific BoFU query and a specific client context.

While the workflow is repeatable and scalable, the results should be interpreted as directional, not definitive.

It remains obvious that:

Disclaimer – TL;DR

This case study is best used as:

ChatGPT’s query fanout experiment: why we ran it

BoFU queries like “best X for Y” don’t behave like classic SEO queries anymore: in ChatGPT, the user often gets a comparison table + recommendations built from a short list of sources, with brands either included as “candidate platforms” or excluded entirely.

Our goal was to observe (and document) the full grounded workflow ChatGPT uses for that decision, then translate it into actionable levers for brand inclusion and stronger “row-level” positioning.

What we tested – methodology, rationale, components

We executed a controlled “fanout observation” run:

Combining human analysis with proprietary GPT projects trained on specific clients characteristics, to spot and surmise content gaps (derived from the fanout sequence).

Core rationale and underlying assumptions

The core idea is to treat ChatGPT’s answer as the end of a pipeline with two distinct stages:

  1. Round 1: autonomous web search + document set construction
  2. Round 2: synthesis into a user-facing comparison (table + narrative)

That separation matters because optimization opportunities differ depending on whether you’re trying to:

Now, there’s a very important factor we have to point out here.

The biggest impact we can hope to have is by acting to influence the process in Round 1, meaning, adapting our web resources to better match with the retrieval process of a specific query.

As David McSweeney nicely put it in an article about how Chat GPT works:

“The output is non-deterministic. You’re trying to “GEO” or “AEO” a cloud formation.

I’ll concede there is some value in tracking directionality (e.g., “Is the sentiment generally positive?”, “are we showing up more”). But trying to reverse-engineer the algorithm based on the specific adjectives GPT 5.2 chose to use on Tuesday vs. Wednesday is a waste of time. There’s far too much noise.

The only part of this chain that is deterministic, the only part you can reliably engineer for, is the retrieval. The search bit.

If Thinky one of ChatGPT’s model, editor’s note] doesn’t find you, you’re praying that the frontier model remembers you from a scrape 12 months ago. Which means, for the best chance of being cited, you need to rank in search. Gamble with your search rankings with GEO spam, gamble with your AI visibility.

How we analyzed it – methodology and pattern observation

We captured and reviewed the artifacts ChatGPT exposes in grounded mode, then classified influence:

Round 1 (retrieval)

Round 2 (synthesis)

What ChatGPT produced – output pattern)

For this BoFU query, ChatGPT returned a short intro + a comparison table labeled “Top Domain Reseller Platforms,” with per-platform “ why it’s good for resellers” rows. Openprovider was included with a concise positioning line.

This matters because it implies a row-construction mechanism: the model looks for “copy-pasteable” fragments, clear descriptors, numbers, and role-based benefits—that fit neatly into a table cell.

Key findings: what influenced inclusion and wording

In this section we break down the specific signals and source patterns that most directly shaped which brands made the shortlist, and how each one was described in the final comparison.

A) “Hard influence” sources were a mix of:

In the Round 1 “hard influence” list, we saw examples such as OpenSRS, Openprovider, Enom, ResellerClub, WHMCS community, LowEndTalk, and a Stablepoint article (Hexonet mention).

Takeaway – for BoFU comparisons, the model doesn’t only reward typical “best-of SEO pages.” It blends product truth (official docs), explainers, and crowd validation.

B) The two lever categories

Two areas of actions (levers) emerged, and we built our final insights and to-do list around them.

Lever 1: off-site levers (outside the client’s website)

ChatGPT seems to like:

Takeaway – these pages can decide whether a brand is even eligible to be mentioned or not.

Lever 2: on-site levers (on Openprovider.com)

ChatGPT looks for:

The bridge insight: why Round 1 vs Round 2 changes what you do

We explicitly mapped levers to pipeline risk:

Results: step-by-step actions list

What we actually did with the findings was the following.

Step 1: translate observations into an actionable recommendation set

We converted the two lever categories into a prioritized checklist, including:

Step 2: use GPT 5.1 to aggregate patterns and pinpoint gaps

After collecting the fanout artifacts (sources, table framing, “candidate pool”) classified as “event: delta”, we used GPT 5.1 to:

Example of the raw data:

event: delta

data: {"v": [{"type": "search_result_group", "domain": "www.openprovider.com", "entries": [{"type": "search_result", "url": "https://www.openprovider.com/blog/best-domain-registrars", "title": "The best domain registrars of 2025", "snippet": "Dec 30, 2024 — Best domain registrars · Openprovider · GoDaddy · Namecheap · Porkbun · CentralNic Reseller · OpenSRS · InterNetX.", "ref_id": {"turn_index": 0, "ref_type": "search", "ref_index": 15}, "pub_date": null, "attribution": "www.openprovider.com"}]}]}

Step 3: create content drafts from the same recommendation set (Make-Rev)

We fed the prioritized recommendations into SEOritmo’s Make-Rev to generate SEO-optimized draft pages and sections aligned to the observed BoFU comparison frame:

Step 4: deploy technical wins at scale (Tech Bender)

Finally, we used SEOritmo’s Tech Bender to generate schema markup at scale, ensuring pages are:

What this case study demonstrates (repeatable takeaway)

This experiment shows a practical way to do “AI search optimization” without guessing:

  1. Run a BoFU prompt in grounded mode under controlled conditions.
  2. Separate retrieval vs synthesis (Round 1 vs Round 2).
  3. Extract hard-influence sources and the candidate pool.
  4. Turn patterns into levers (off-site + on-site), mapped to where they affect the pipeline.
  5. Operationalize the recommendations into content and technical deployments using an SEO AI content tool(Make-Rev), and an automated technical SEO tool(Tech Bender).