Stop writing content first. Look at what the AI actually searches before it answers.
The reason is mechanical: a customer asks one full question, the model expands it into several retrieval queries, searches on those, then builds an answer from the pages returned. The keywords in your spreadsheet are not the unit of competition. The expanded queries are.
What query fan-out actually is
Query fan-out is the step where a model turns a single user question into multiple search queries before generating an answer.
Take the question "how long does US customs clearance take". The model may simultaneously retrieve material on IOR number rules, Form 5106 requirements, demurrage and storage calculations, HS classification disputes, and AMS or ISF filing deadlines.
This process used to be invisible. On 14 October 2026, an author open-sourced a tool (chatgpt-fanout-queries, in the open-growth-skills repository) that calls the Responses API with web search enabled, stores the full raw response, then extracts queries, sources and url_citation entries into an offline HTML report.
One distinction in that report matters more than anything else: sources returned by search are not the same as sources cited in the final answer. Ranking in retrieval only buys the chance to be seen. Citation buys the chance to be used.
Three kinds of content gaps
Once you have the expanded queries, audit your site against them. Gaps usually fall into three groups.
The first is covered but shallow. A concept appears once in a sentence and cannot stand alone as an answer. The fix is to rewrite, covering definition, responsible party, applicable conditions and consequences of error.
The second is missing entirely. These are the foundational pages customers search for and competitors rarely write: filing forms, classification rules, cost breakdowns. They are the highest priority because intent is clear, competition is thin and reuse is high.
The third is structural. One intent is split across several queries and scattered through many articles, none of them complete. The fix is a topic page with internal links, not another standalone post.
Make the trail part of the deliverable
GEO services lose trust for a predictable reason: conclusions without process.
A workable delivery format puts a section at the front of the diagnostic report listing which queries the question expanded into, which sources were returned, which pages were ultimately cited, and whether a corresponding page exists on the client site. The client can check every line, and three months later, compare.
For freight forwarding, where repeat business decides survival, that paper trail is worth more than any ranking claim.
Where the boundary sits
A single run is a lead, not a conclusion. It says nothing about search volume or ranking difficulty and guarantees no citation. The defensible method is to run multiple real customer questions repeatedly, then weigh results against the existing keyword base, site performance and business relevance.
The white-hat boundary is equally firm: analyse publicly observable search behaviour to find content gaps, do not probe how AI systems evaluate brands, do not extract scoring or inclusion rules, and never promise rankings or guaranteed citations.
How this connects to AI adoption
Gap analysis, page comparison and report archiving are all process work. Turning "customer questions, expanded queries, gap categories, content calendar" into a monthly routine beats hiring someone to write ten articles on instinct.
It is also the easiest AI initiative in freight forwarding to justify: it does not change model capability, it changes how topics get chosen.
FAQ
Q1: Is GEO the same as SEO?
Not exactly. SEO optimises short-query ranking and page quality on search engine results pages. GEO focuses on being cited inside generated answers. They share the same site-quality foundation, but GEO adds a layer of query expansion analysis.
Q2: Can a query fan-out report predict rankings?
No. It reflects observable retrieval behaviour for one prompt at one moment and is not a ranking or volume signal.
Q3: Which pages should a forwarder write first?
Foundational pages that customers search for and competitors rarely cover: IOR numbers, customs forms, demurrage components, classification disputes.
Q4: How often should the analysis be run?
Monthly, with 10 to 20 real questions per run, keeping each result for comparison.
Q5: How do you verify a gap has been closed?
Re-run the same question set and check whether the new page appears among retrieval sources and whether its link is eventually cited.
References
- "GEO Open-Source Skill: ChatGPT Fan-out Queries Report", by Jeff Li, open-growth-skills repository, published 14 October 2026
- Methodological caveats (single runs as leads only, no volume or ranking implication, no citation guarantee) are as stated by the tool's author