Kojable research
How Predictable Are Gemini’s Fan-Out Queries? A 180-Prompt Study
Rephrasing a buyer question does not necessarily send Gemini down an unrelated retrieval path. Kojable analysed how grounded search behaviour changed across 180 B2B finance prompts with meaningful variation in wording and context.
Prompt similarity and grounding-query-set similarity had a strong positive relationship: r = 0.869. Similar prompts usually triggered similar retrieval paths, without making those paths identical.
Quick answer
Similar prompts often led Gemini towards similar subtopics
Fan-out query similarity is the degree to which semantically similar prompts trigger semantically similar sets of searches inside a grounded AI system. Kojable analysed 180 B2B finance prompts and recorded 1,620 grounding-query instances, an average of nine recorded search instances per prompt.
Prompt similarity and query-set similarity correlated at r = 0.869. The result supports organising prompt variations into measurable semantic clusters. It does not prove identical searches, identical citations, or identical brand outcomes.
Research at a glance
A pairwise view of prompt wording and grounded retrieval
The study compared semantic similarity in the original prompts with semantic similarity in the associated grounding query sets. The reported query count is treated as recorded query instances rather than a claim that every string was unique.
Prompts
Three B2B finance topic clusters with 60 prompts each.
Prompt pairs
Every unique unordered pair among the 180 prompts.
Within-topic pairs
Comparisons between prompts in the same topic cluster.
Cross-topic pairs
Comparisons between prompts in different topic clusters.
Query instances
Recorded grounding-query instances across the experiment.
Mean per prompt
An average of nine recorded searches, not a count for every prompt.
Prompt to query-set correlation
A strong positive relationship in this dataset.
Descriptive squared correlation
A descriptive summary, not a causal estimate.
Matrix permutations
Zero matched or exceeded the observed correlation.
Empirical result
Evidence against an arbitrary matching of the two matrices.
Topic-specific source titles
Associated with only one topic cluster.
Companion result
Prompt similarity versus final response similarity.
Fan-out queries
An observable part of a grounded retrieval path
A fan-out query is one of several related searches an AI system generates to gather the information needed for a broader user question. Google’s platform descriptions explain that a grounded model can analyse a prompt, decide whether search may improve its answer, generate one or more searches, process retrieved information, and return a grounded response with source information.
This study uses “fan-out queries” as practical shorthand for searches recorded in Gemini grounding metadata. They expose an observable part of the retrieval path: what the system searched for before generating an answer. They do not prove why a particular source was selected or reveal every internal model process.
Why this matters for AEO
Retrieval signals sit upstream of the answer
Answer-level measurements show whether a company appears, is cited, sits beside competitors, is described accurately, or is recommended in a particular order. Fan-out queries show the subtopics, reformulations, constraints, supporting evidence needs, and information gaps that can shape those downstream answers.
A question such as “How should a SaaS CFO improve free cash flow?” can lead to narrower needs around SaaS free-cash-flow benchmarks, working capital, accounts-receivable collection periods, forecasting, calculation methods, and finance software. A company may address cash-flow management broadly yet lack evidence for the specific needs the system is trying to resolve.
Analyse the prompt
The model interprets the user’s question and the information it may require.
Decide whether search may help
Grounding can be used when the model determines that external information could improve the response.
Generate related searches
One or more searches can explore the broader question through useful subtopics and constraints.
Process retrieved information
The model uses the retrieved material as part of preparing a grounded answer.
Produce the answer with source information
The final response is where citations, brand representation, and recommendation outcomes can be assessed.
Research question
When similar users ask differently, does Gemini search for similar information?
If minor wording changes produced unrelated retrieval behaviour, AEO teams would need to monitor a very large prompt universe. If related prompts produced overlapping retrieval paths, teams could cluster questions, identify recurring query families, and prioritise fewer evidence gaps.
The study tests that second possibility without assuming that similar retrieval paths settle every commercial question. Brand inclusion, citations, recommendations, factual accuracy, and vendor order remain separate outcomes to measure.
Experiment design
180 prompts across three finance topics
The prompt set was divided evenly between free cash flow, B2B payment automation, and transaction fraud detection. Each topic contained 60 prompts. Rather than creating superficial rewrites, the study varied persona, industry, geography, integration, intent, and prompt form.
Examples included “Free cash flow explained for B2B SaaS”, “How does cash flow forecasting work for a CFO?”,
“Best cash flow tools with NetSuite integration”, and “Generate a monthly free cash flow report”. Every prompt
was submitted to gemini-3-flash with Google Search grounding enabled.
Enabling grounding does not mean every prompt must trigger the same number of searches. The experiment recorded 1,620 grounding-query instances across 180 prompts: an exact mean of nine recorded query instances per prompt. The average does not imply that every prompt produced nine searches.
Embedding and similarity methodology
Semantic meaning is more useful than exact word overlap
Exact-word overlap would treat “How to calculate free cash flow” and “FCF calculation formula” as less related
than they are. The study embedded both prompts and grounding queries with gemini-embedding-001, converting
text into numerical vectors that represent semantic meaning.
Cosine similarity then compares how closely two vectors point in the embedding space. A higher value indicates greater semantic similarity. It is a similarity measure, not a probability and not proof that two strings are identical.
Symmetric best-match similarity
Compare query sets without rewarding longer lists
Query-set comparison is difficult because prompts can produce different numbers of queries, their order is not guaranteed to align, and position-by-position matching would be inappropriate. Comparing every query with every query would also favour a larger set simply because it has more possible matches.
The study therefore used symmetric best-match similarity. It measures semantic coverage between two sets instead of requiring the same wording, order, or list length.
Match Set A to Set B
For each query in Set A, find its most semantically similar query in Set B and average those best matches.
Match Set B to Set A
Repeat the same calculation in the reverse direction.
Average the two directions
One direction measures how well B covers A; the reverse measures how well A covers B. Averaging reduces the advantage of a larger set.
S(A, B) = 1/2 × [average best match from A to B + average best match from B to A]
Pairwise analysis
Two corresponding 180-by-180 similarity structures
The 180 prompts produced 180 × 179 ÷ 2 = 16,110 unique unordered prompt pairs: 5,310 within-topic pairs and 10,800 cross-topic pairs. For each pair, the analysis compared similarity between the original prompts with symmetric best-match similarity between their associated query sets.
This produced one similarity structure for prompts and another for retrieval-query sets. The central analysis asks how closely those two structures align, while recognising that the individual pairs share prompts and are not independent.
Finding 1
Similar prompts triggered similar query sets
The Pearson correlation between prompt similarity and query-set similarity was r = 0.869. Prompts that were closer in semantic meaning generally generated query sets that were closer in semantic meaning too. In practical terms, similar questions usually led Gemini to investigate similar subtopics.
The descriptive squared correlation was r2 = 0.755. Descriptively, about 75.5% of pairwise variation aligned with the linear relationship in this dataset. Because prompt pairs share observations, this is not a causal estimate or an independent-sample regression result.
Finding 2
Retrieval behaviour remained topic-sensitive
Free-cash-flow prompts tended to produce searches about cash flow, forecasting, working capital, and finance metrics. Payment-automation prompts tended to search for processing, integrations, reconciliation, and payment operations. Fraud-detection prompts tended to search for transaction risk, detection methods, controls, and fraud signals.
Within-topic pairs had higher query-set similarity than cross-topic pairs. All three clusters were still within B2B finance, so this is a practical control: Gemini did not simply reuse one generic finance-search plan for every topic.
Finding 3
The relationship survived a matrix permutation test
A conventional independent-observation p-value would be misleading because each prompt appeared in 179 pairwise comparisons. Pairwise observations share prompts, so treating all 16,110 pairs as independent would create pseudoreplication and overstate the effective sample size.
The study instead used a Mantel-style matrix permutation procedure. It calculated the observed correlation, randomly permuted corresponding row and column labels in one matrix, recalculated the relationship after breaking the original prompt-to-query alignment, repeated that process 2,000 times, and compared the result with the permutation distribution.
Observed relationship
The measured prompt-to-query-set correlation was r = 0.869.
2,000 permutations
None of the randomly permuted matrices matched or exceeded the observed correlation.
Empirical result
The reported result was p < 0.001, supporting a non-random association between the two matrices.
Finding 4
The cited-source landscape was largely topic-specific
More than 80% of cited source titles appeared in only one topic cluster. Cash-flow prompts mostly cited cash-flow material; fraud prompts mostly cited fraud-related material; and payment-automation prompts mostly cited payment-system and operations material. Some broad finance publishers appeared across more than one cluster.
The cited-source landscape was largely topic-specific in this experiment. This is source-title specificity, not proof of domain specificity or URL specificity. A topic-specific title does not prove that its specificity caused a citation, and the study did not measure source-ranking weights.
Finding 5
Retrieval similarity nearly matched response similarity
The companion study measured prompt similarity versus final response similarity at r = 0.878. This study measured prompt similarity versus grounding-query-set similarity at r = 0.869. The difference is 0.009, indicating substantial semantic structure in both the retrieval-query sets and the generated answers in this dataset.
Read the companion 180-prompt response-similarity study. Neither result proves that brand mentions, citations, recommendations, or factual claims remain identical across prompt variations.
What this changes about AEO
Optimise for semantic query families, not 1,620 literal strings
Exact-query optimisation would reproduce the weaknesses of keyword stuffing. The useful unit is the semantic query family: a recurring entity, subtopic, constraint, evidence need, and decision context. Several searches about free cash flow may belong to one family: Free cash flow calculation and measurement for B2B SaaS.
A strong page for that family can include a direct definition, formula, worked example, SaaS-specific adjustments, common errors, sourced benchmarks, clear author information, an update date, and relevant internal links. The goal is not to repeat every search phrase; it is to become a clear, extractable, well-supported source for the information need behind the family.
Practical workflow
A seven-step fan-out optimisation playbook
Use the fan-out layer to move from a large set of literal strings to a smaller, decision-ready view of recurring information needs. The workflow keeps buyer context, evidence quality, and retesting in the same operating loop.
Define the buyer-question clusters
Start with commercially meaningful questions across discovery, education, implementation, integration, comparison, risk, pricing, and proof. Keep distinct buyer problems separate even where terminology overlaps.
Select representative seed prompts
Vary persona, use case, industry, geography, intent, product requirement, and expertise. Represent genuine buyer context rather than producing dozens of superficial rewrites.
Capture the complete grounded record
Keep the prompt, answer, grounding queries, source titles and URLs, cited material, and observable response context together so later checks remain comparable.
Cluster the fan-out queries semantically
Group queries by meaning rather than exact-match frequency. A query expressed once can still belong to a high-frequency family with many formulations.
Prioritise query families
Beginning with the top 15 families per topic is an operating starting point. Across three topics that would create 45 priority families, not a universal optimum or a confirmed coverage percentage.
Map each family to the information environment
The right action may be a product page, use-case page, technical documentation, comparison, FAQ, benchmark, or third-party evidence rather than a new article. Explain what to change, why, where, and how it will be retested.
Retest comparable prompts
Compare query families, sources, brand and competitor citations, answer language, recommendation order, entity accuracy, source-title specificity, and retrieval volatility. Use 7, 14, and 30-day checks where the category changes quickly.
Metrics AEO teams should track
Connect retrieval measurements to answer quality
No single metric is sufficient. Stable fan-out coverage can coexist with poor citations; high mention rates can coexist with inaccurate descriptions; and citations can come from weak or outdated pages.
Fan-out query coverage
The proportion of recorded queries represented by priority query families.
Query-set similarity
Semantic overlap between the search-query sets generated by different prompts.
Retrieval volatility
How much a fan-out query set changes between repeated tests.
Source overlap
The cited sources shared across prompts, clusters, or time periods.
Source-title specificity
The proportion of cited titles associated with one defined topic cluster.
Brand mention rate
The proportion of answers that mention the company.
Citation rate
The proportion of answers that cite a company-controlled or target source.
Recommendation rate
The proportion of answers that actively recommend the company.
Competitor co-mention rate
How often the company and named competitors appear together.
Entity accuracy
The accuracy of category, audience, capability, and company descriptions.
Share of answer
The company’s mention, citation, or recommendation presence relative to competitors across a defined answer set.
Citation quality
The relevance, credibility, freshness, and actionability of the sources used.
Kojable interpretation
AEO is a measurement and evidence-management problem
Kojable applies fan-out analysis through Monitor, Diagnose, Improve, and Verify. Monitor captures prompts, answers, grounding queries, citations, and competitor appearances. Diagnose identifies recurring query families, source patterns, missing evidence, outdated claims, and topic-specific gaps. Improve explains what should change, why it matters, where it belongs, and how to carry it out. Verify retests comparable prompts and measures changes in retrieval paths, citations, and AI representation.
The underlying question is: “Which recurring information needs are shaping buyer-facing AI answers, where is our evidence missing, and what changes when we improve it?” Piush Vaish, Kojable’s founder, brings approximately a decade of data-science and machine-learning experience to that measurement-first perspective.
Limitations
What this study does not prove
- Similar prompts always produce identical searches; r = 0.869 is strong but not perfect.
- Similar query sets guarantee identical sources, citations, brand inclusion, vendor ranking, sentiment, recommendations, or factual accuracy.
- The relationship is causal rather than an observed association.
- The result generalises beyond Gemini with Google Search grounding, B2B finance, or this point in time.
- Source-title specificity establishes domain specificity, URL specificity, or source causality.
- The permutation result removes every statistical limitation.
Models, indexes, sources, and retrieval systems change. Brand-level outcomes and citation consistency require their own measurements even when two retrieval-query sets are highly similar.
Bottom line
Structured enough to measure, not fully predictable
Across 180 prompts, 16,110 prompt pairs, and 1,620 recorded grounding-query instances, prompt similarity correlated with query-set similarity at r = 0.869. None of 2,000 matrix permutations equalled or exceeded the observed relationship, and more than 80% of cited source titles were associated with only one topic cluster.
Group buyer questions into meaningful clusters. Observe recurring search-query families. Identify the pages and evidence needed to answer them. Prioritise commercially important gaps. Retest comparable questions after the information environment changes.
The fan-out process was not fully predictable, but it was structured enough to measure.
FAQ
Questions about Gemini fan-out query similarity
What is a fan-out query?
A fan-out query is one of several related searches an AI system generates to gather information for a broader user question. It represents a subtopic, constraint, or supporting information need within the original prompt.
What is fan-out query similarity?
It measures how closely the search-query set generated for one prompt resembles the set generated for another. This study used semantic embeddings and symmetric best-match similarity rather than exact word overlap.
How many fan-out queries did Gemini generate in the study?
The experiment recorded 1,620 query instances across 180 prompts, an average of nine per prompt. The published analysis calls them query instances rather than claiming that all 1,620 strings were unique.
How strong was the relationship between prompts and fan-out queries?
The correlation between prompt similarity and query-set similarity was r = 0.869. The squared correlation was approximately 0.755 and is presented descriptively because the pairwise observations were dependent.
Why was symmetric best-match similarity used?
Different prompts generated query sets of different sizes. Symmetric best-match similarity compares each query with its closest semantic match in the other set in both directions, reducing bias from unequal list lengths.
What did the permutation test show?
None of 2,000 random permutations matched or exceeded the observed correlation of 0.869. The study reported an empirical result of p < 0.001.
Did similar prompts cite the same sources?
The study found strong topic specificity in source titles, with more than 80% associated with only one topic cluster. It did not prove that every similar prompt cited the same URL.
Does this mean a company only needs one prompt per topic?
No. One prompt may represent broad semantic behaviour, but teams still need variations covering personas, intents, integrations, geographies, and commercial contexts. Brand outcomes also need separate measurement.
Should content teams optimise for the exact fan-out queries?
Use the queries to identify recurring entities, subtopics, and evidence needs. Do not repeat exact strings unnaturally; build clear, useful, well-supported pages for the underlying information need.
How does Kojable use fan-out query analysis?
Kojable uses fan-out queries within its Monitor, Diagnose, Improve, and Verify process to identify recurring retrieval patterns, source gaps, content priorities, and the changes to retest.
Next step
See which information needs are shaping AI answers about your company
Use evidence-backed Answer Intelligence to monitor representation, diagnose the sources and gaps shaping it, improve the information environment, and verify what changes.