Kojable research · Decision 2 · Query type and citation exposure
AI Citations by Query Type: What a Large Cross-Platform Study Shows
A cross-platform analysis of how buyer-question type relates to visible AI citation exposure—and why the largest raw query-category difference becomes much smaller once retrieval need is compared on common support.
Key finding
Gemini shows the clearest query-type variation, but its largest raw route-level deficit shrinks from roughly 29 percentage points to about 5 points in the comparable Medium-retrieval-need cohort.
Qualification: Query type and retrieval need overlap substantially in the observed prompt inventory, so the raw route difference is not an independent query-type effect or causal retrieval mechanism.
- AI citations by query type
- Buyer-question type
- Gemini citation measurement
- 16 min read
Study overview
Executive summary
Does the kind of question a user asks change the likelihood that an AI answer contains visible citations? Across a multi-month dataset of more than 55,000 responses from ChatGPT, Gemini and Perplexity, the answer is yes—but not as a universal rule.
Query type was associated with citation exposure most clearly on Gemini, only weakly on ChatGPT and hardly at all on Perplexity because Perplexity cited at close to saturation.
Gemini’s Educational route initially appeared to sit about 29 percentage points below the platform baseline: around 61% compared with about 90%. But Educational queries were strongly entangled with retrieval need. In the supported Medium-need comparison, the remaining difference was about 5 points.
The raw ~29-point Educational gap shrinks to roughly ~5 points in the comparable Medium-retrieval-need cohort.
Query type is useful for identifying where AI citation exposure differs, especially on Gemini, but the largest raw query-type difference is heavily entangled with retrieval need and repeated prompt design. Buyer-question categories should be treated as diagnostic segments, not causal retrieval mechanisms.
Research lineage
How this differs from Decision 1
Decision 1 asks whether retrieval need predicts visible citation exposure. Decision 2 asks whether buyer-question type adds useful structure once retrieval context is considered.
Decision 1 owns retrieval need. Decision 2 owns query and buyer-question type. The papers remain separate because the classifications answer different analytical questions even when they overlap in the observed prompt inventory.
The next stage asks whether broad source absence explains the remaining citation misses. Read Decision 3 on AI citation misses and source availability.
Analytical definition
What “query type” and “query route” mean
Query type is an analyst-created label describing the kind of information a prompt asks for. It is not an observed internal routing decision made by ChatGPT, Gemini or Perplexity.
Representative public examples include:
- Vendor evaluation
- Product capability
- Operational workflow
- Integration / ecosystem
- Security / governance
- Educational / informational questions
These sit alongside several narrower query categories.
The major buyer and implementation categories account for most of the observed responses and have broader prompt-template coverage. Several narrower routes should be interpreted more cautiously. Response volume and template breadth are not the same thing. Repeating a small number of prompt designs does not provide the same generalisability as observing many distinct prompts.
Method
What this study measures
Study scale
More than 55,000 response rows across ChatGPT, Gemini and Perplexity, using approximately 1,000 labelled prompt templates.
Route coverage
More than ten broad buyer-question categories with unequal response volume and prompt-template breadth.
Outcome
A binary visible-citation indicator recorded whether the final answer contained at least one detected citation.
Comparisons
Within-platform exposure, difference from platform baseline, template-level sensitivity checks and comparable retrieval context.
Citation exposure is a measurement outcome, not proof of retrieval. A response may retrieve information without displaying a citation, and a visible citation does not establish source authority, relevance, recency or support.
Platform results
Citation exposure varies by query type mainly on Gemini
| Platform | Approximate overall exposure | Query-type sensitivity |
|---|---|---|
| Gemini | ~90% | Meaningful observed variation |
| ChatGPT | ~95% | Comparatively small variation |
| Perplexity | ~100% | Near-ceiling exposure |
These rates must be interpreted within platform. A route at 95% carries different information on a platform whose baseline is about 95% than on one whose baseline is about 90%.
Gemini shows the clearest variation
The Educational route is the standout result at around 61%, roughly 29 percentage points below the platform baseline. Smaller route differences should be interpreted cautiously where template breadth is narrow.
ChatGPT differences are compressed
ChatGPT’s better-supported route rates remain close to its baseline. Large samples can make small differences detectable, but their practical association remains limited.
Perplexity is near a ceiling
Perplexity cites in almost every observed response, leaving little variation for a binary citation-presence outcome to explain. This does not mean Perplexity retrieves identically for every query type. Source choice, citation count, source family, domain diversity, source authority and claim support are more useful next-stage measures in a saturated environment.
Central result
The raw Gemini gap shrinks sharply on common support
| Gemini comparison | Approximate citation exposure |
|---|---|
| Overall | ~90% |
| Educational, raw | ~61% |
| Medium-need Educational | ~85% (mid-80% range) |
| Other Medium-need routes | ~90% |
The raw Educational deficit is approximately 29 points. The comparable Medium-need difference is approximately 5 points.
Low retrieval need and Educational query type overlap heavily in the observed Gemini data, so their independent effects cannot be cleanly separated. The comparable Medium-retrieval-need cohort shows a much smaller Educational difference than the raw route-level comparison.
Defensible interpretation
The raw ~29-point Educational gap shrinks to roughly ~5 points in the comparable Medium-retrieval-need cohort.
The data do not support attributing lower citation exposure to Educational query type alone.
Sensitivity analysis
Template weighting tells the same story
The Educational route is internally heterogeneous. Its response-weighted Gemini rate is much lower than its equal-template interpretation.
Template-level sensitivity checks show that the low aggregate Educational result is not evenly distributed across every prompt design. A narrower cluster contributes disproportionately to the route-level average. Many Medium-need Educational templates behave much more like other Medium-need routes.
Interpretation boundary
Route differences are diagnostic, not causal
Route-level citation differences identify where to investigate. They do not reveal a platform’s hidden retrieval policy.
The observed differences may reflect retrieval need, repeated prompt-template structure, uneven route breadth, platform interactions, the evidence environment or combinations of those factors. The analytical label does not demonstrate that an internal query router exists.
SEO, AEO and GEO measurement
Use query type as a diagnostic segment
Buyer questions are not merely keyword variations—they represent different evidence needs.
The useful unit of analysis is platform × buyer-question type × retrieval context, not a single blended visibility score. A company can perform differently on vendor selection, product capability, workflow, integration, security or educational and informational questions. Strong overall visibility can therefore conceal weakness in a commercially critical question category.
-
Segment by buyer question
Track commercially important routes separately in AI citation monitoring.
-
Keep platforms separate
The observed relationships do not transfer cleanly across systems.
-
Investigate evidence before prescribing content
Inspect available and selected sources, brand presence, evidence quality and prompt composition. A low citation rate does not identify the remedy by itself.
-
Prioritise template breadth
Add distinct buyer questions before collecting many more repetitions of a narrow route.
Supported conclusions
What Decision 2 establishes
-
Query type is useful context
Citation exposure differs across categories, particularly on Gemini.
-
There is no universal route rule
The relationship is platform-specific.
-
The largest raw difference is confounded
The roughly 29-point raw gap becomes about 5 points in the comparable Medium-need cohort.
-
Template breadth matters
Large response volume cannot substitute for diverse prompt designs.
-
Citation presence is only a first observable stage
It does not measure source choice, claim support, authority, recency or source diversity.
Diagnostic research model
Follow the Decisions 1–8 diagnostic sequence
The sequence is retrieval need → query type → source availability → selected sources → content access → passage relevance → finalist quality → answer readiness. It is a diagnostic research model, not the confirmed internal architecture of AI platforms.
-
Decision 1
-
Decision 2
-
Decision 3
-
Decision 4
-
Decision 5
-
Decision 6
-
Decision 7
-
Decision 8
Study boundaries
Limitations
-
Observational design
Associations do not identify causal retrieval or citation mechanisms.
-
Overlapping classifications
Low retrieval need and Educational query type overlap heavily in the observed Gemini data, limiting independent attribution.
-
Template dependence
Repeated responses from the same prompt templates create clustering.
-
Uneven breadth
Several routes remain too narrow for broad generalisation.
-
Binary outcome
Visible citation presence does not measure retrieval, source choice, support quality or accuracy.
Method and governance
Research and reproducibility details
- Collection period and scope
-
The study used a multi-month collection in a consistent English-language market setting.
- Collection system
-
An AI monitoring system identified the included ChatGPT, Gemini and Perplexity product environments. Exact underlying model names and versions were not available.
- Taxonomies
-
Query route and retrieval need are rule-based classifications derived from prompt content, not internal platform decisions.
- Citation detection
-
A binary visible-citation indicator recorded whether the final answer contained at least one detected citation.
- Public materials
-
This publication reports approved rounded findings and representative analytical labels. Identifying collection details and reconstructive statistical materials are withheld.
Conclusion
Query type adds context, not a causal routing rule
Query type is useful for identifying where AI citation exposure differs, especially on Gemini, but the largest raw query-type difference is heavily entangled with retrieval need and repeated prompt design. Buyer-question categories should be treated as diagnostic segments, not causal retrieval mechanisms.
The central public result is the attenuation from an approximately 29-point raw Educational deficit to an approximately 5-point difference in the comparable Medium-retrieval-need cohort.
For AI search, SEO, AEO and GEO measurement, use platform × buyer-question type × retrieval context, then investigate the evidence behind any observed deficit before prescribing a content or publishing response.
FAQ
Frequently asked questions
Does query type affect AI citation exposure?
Yes. Query type is associated with citation exposure in the observed data, most clearly on Gemini, but the relationship is platform-specific and does not establish causation.
Which platform shows the strongest query-type variation?
Gemini shows the strongest observed variation. ChatGPT's category-level differences are comparatively small, while Perplexity is near citation saturation.
Why does Gemini's Educational category appear to have lower citation exposure?
Its raw citation exposure is about 61% against a Gemini baseline of about 90%. Low retrieval need and Educational query type overlap heavily in the observed data, and a narrower cluster of repeated prompt designs contributes disproportionately to the aggregate.
Why does the raw ~29-point Gemini gap shrink to about ~5 points?
The raw comparison is about 61% versus 90%. Within the comparable Medium-retrieval-need cohort, Educational is about 85% and other routes are about 90%, leaving a difference of about 5 points.
Does this prove Educational queries cause fewer Gemini citations?
No. The study is observational, query type and retrieval need overlap substantially, and repeated prompt designs also influence the route-level average.
What does “query type” mean in this study?
It is an analyst-created label for the kind of information a prompt asks for. It is not an observed internal routing decision made by an AI platform.
Why should query-type effects be measured within platform?
Platform baselines differ: Gemini is about 90%, ChatGPT about 95% and Perplexity about 100%. The same route rate therefore has a different meaning on each platform.
Why is Perplexity difficult to analyse using citation presence alone?
Perplexity cites in nearly every observed response, so binary citation presence has little remaining variation for query type to explain. Source choice, diversity and claim support are more informative next measures.
Why does prompt-template breadth matter?
Repeating a small number of prompt designs does not provide the same generalisability as observing many distinct prompts, even when both approaches produce many response rows.
How does Decision 2 connect to Decisions 1 and 3?
Decision 1 studies retrieval need, Decision 2 studies buyer-question type, and Decision 3 asks whether broad source availability explains the remaining citation misses.
