Section 01Executive Summary
The four frontier assistant families do not behave like one channel. On the same commercial prompts they name different brands, lean on different sources and fail in different ways. A brand optimising for "AI" in the aggregate is optimising for an average that no single assistant exhibits.
This study evaluates a 50,000-prompt commercial subset of the KernelX Benchmark Corpus — best-of, comparison, alternatives and purchase-advice intents only — issued identically to the GPT, Claude, Gemini and Perplexity assistant families. Metrics follow the methodology paper exactly; all figures are representative benchmark findings expressed as ranges.
Key findings
- GPT recommended the widest range of established brands — the broadest recommendation diversity of any family, with a measurable tilt toward well-known incumbents.
- Claude relied most heavily on official documentation and long-form canonical content, and was the most conservative recommender: fewer named brands per answer, more hedging.
- Gemini showed the strongest alignment with its surrounding knowledge ecosystem — knowledge-graph presence and structured data correlated with recommendation more strongly than for any other family.
- Perplexity cited external sources far more frequently than the other assistants, with the flattest trust curve across source classes.
- Reddit influenced recommendations most strongly in consumer software and electronics — visible in every family, strongest in GPT and Perplexity.
- Cross-family agreement on the top recommendation was roughly 40–55% depending on intent — confirming that visibility must be managed per model.
Model families are compared at the assistant-product level under default consumer settings during the collection window. Version churn means absolute numbers age quickly; the behavioural signatures below have so far proven more durable than any point estimate.
Key takeaways
- Each assistant family has a distinct recommendation signature.
Business implications
- A single "AI visibility" number hides material per-model risk.
Recommendations
- Measure and optimise per model family, prioritised by your audience's assistant mix.
Section 02Methodology & Research Questions
The panel design, entity registry, recommendation scoring, consistency sampling and hallucination checks are inherited unchanged from the flagship methodology paper. This study restricts the corpus to the four commercial intent classes where brand selection is explicit, and adds per-family breakdowns of every metric.
Research questions
- How does recommendation breadth and diversity differ across assistant families? (RQ1)
- Which source classes does each family prefer when grounding commercial answers? (RQ2)
- How do families differ in consistency across repeated sampling and paraphrase? (RQ3)
- How do failure rates — hallucinated entities and attribute errors — compare? (RQ4)
- Where do families agree and disagree, and what does disagreement imply for brands? (RQ5)
Key takeaways
- Identical panels per family make the comparison like-for-like.
Business implications
- Differences below reflect model behaviour, not prompt selection.
Recommendations
- Replicate with your own category panel before generalising to your brand.
Section 03Recommendation Behaviour
The families differ visibly in how many brands they name, how concentrated those names are, and how willing they are to commit to a single recommendation.
| Metric (panel medians) | GPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| Brands named per commercial answer | ~4–6 | ~2–4 | ~3–5 | ~5–7 |
| Unique brands across a category panel | Highest | Lowest | Mid | High |
| Top-3 concentration of recommendation weight | ~50–55% | ~65–70% | ~60–65% | ~45–50% |
| Commits to a single "best" when asked | Often | Sometimes, hedged | Often | Usually, with sources |
| Incumbent tilt (share going to top-5 category leaders) | High | High | Mid–high | Mid |
Table 1 — Recommendation-behaviour signatures by assistant family on the commercial panel.
The visual communicates normalised diversity of recommended entity sets: Perplexity and GPT surface the broadest brand mix; Claude concentrates on a narrower, more established set.
Challenger brands have asymmetric opportunity: retrieval-led assistants (Perplexity, GPT) are the realistic first surface to win, while Claude's narrower set makes it the strongest trust signal once earned. Sequence targets accordingly.
Key takeaways
- Breadth: Perplexity ≈ GPT > Gemini > Claude; concentration runs the opposite way.
Business implications
- Where a brand can realistically break in first differs by family.
Recommendations
- Challengers: lead with retrieval-heavy surfaces. Incumbents: defend concentration in Claude/Gemini.
Section 04Citation Behaviour
Grounding rates and source preferences separate the families sharply. Perplexity cites in nearly every answer; the chat-first assistants cite selectively; and each family's source mix reveals its retrieval posture.
| Source class (share of cited answers) | GPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| Answers containing citations at all | ~55–60% | ~50–55% | ~60–65% | ~90–95% |
| Official brand sites | ~78–82% | ~85–88% | ~80–84% | ~72–76% |
| Official docs & long-form canonical | Mid | Highest | Mid | Mid |
| Community (Reddit et al.) | ~40–44% | ~28–32% | ~34–38% | ~44–48% |
| Encyclopedic (Wikipedia) | ~30–34% | ~34–38% | ~36–40% | ~28–32% |
| Knowledge-ecosystem surfaces (graphs, panels) | Low–mid | Low | Highest | Low–mid |
| News & trade media | ~22–26% | ~20–24% | ~24–28% | ~30–34% |
Table 2 — Citation signatures by family. Claude anchors on canonical documentation; Gemini on knowledge-ecosystem surfaces; Perplexity spreads citations widest.
The visual communicates the grounding gap: Perplexity behaves like a citation engine; the chat-first families cite selectively, reserving citations for factual anchors.
The same brand asset earns citations for different reasons per family: documentation depth wins Claude, structured data and entity panels win Gemini, community corroboration wins GPT, and being retrievable at all wins Perplexity.
Key takeaways
- Grounding rate and source mix are family-specific signatures.
Business implications
- Content investments map to specific families — documentation ≠ structured data ≠ community.
Recommendations
- Match asset strategy to the family that your buyers actually use.
Section 05Community Influence: The Reddit Effect
Community corroboration is now a first-class recommendation input, but its weight varies by category. Measuring the share of recommendations accompanied by community-sourced justification ("users report…", cited threads) shows where the effect concentrates.
The visual communicates where community consensus moves recommendations: strongest in consumer software and electronics, weakest in regulated categories where all families revert to canonical sources.
Community presence cannot be faked at measurement time. Sustained, authentic participation — answered threads, honest comparisons, responsive maintainers — is the only community signal that survives across sampling runs.
Key takeaways
- Reddit-class consensus moves consumer categories most; regulated categories least.
Business implications
- Community strategy is a visibility strategy, not just a support channel.
Recommendations
- Consumer brands: treat the top 5 community threads about your category as a landing page.
Section 06Consistency & Failure Modes
Repeat sampling and paraphrase variants expose how stable each family's answers are — and registry checks expose how often they fail. The families trade off differently between commitment and caution.
| Metric (commercial panel) | GPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| Stability of top recommendation across 5 samples | ~72–78% | ~80–85% | ~74–80% | ~68–74% |
| Paraphrase robustness (same answer, reworded prompt) | Mid | Highest | Mid–high | Mid |
| Hallucination-class failures (entities/attributes) | ~4–6% | ~2–4% | ~3–5% | ~3–5% |
| Stale-fact errors (outdated pricing, features) | Mid | Mid | Low–mid | Lowest |
| Refusal / heavy hedging on advice | Low | Highest | Low–mid | Lowest |
Table 3 — Consistency and failure signatures. Claude trades breadth for stability; Perplexity trades stability for freshness.
Failure rates are lower bounds: only errors checkable against the entity registry are counted. Subjective mischaracterisations ("X is best for enterprises" when it is not) are outside the detector's scope.
Key takeaways
- Stability and failure profiles differ as much as recommendation profiles.
Business implications
- Brand-safety exposure (being hallucinated about) is family-specific too.
Recommendations
- Monitor your brand's failure modes per family; corrections target different root causes.
Section 07GEO Implications & Best Practices by Family
The signatures compose into a practical playbook. The table below summarises the highest-leverage asset per family, given the observed retrieval and citation posture.
| Family | Behavioural signature | Highest-leverage assets | Primary risk |
|---|---|---|---|
| GPT | Broad, incumbent-tilted, community-aware | Comparison hubs; wide third-party footprint; community corroboration | Being outframed by better-known incumbents |
| Claude | Narrow, canonical, stable, cautious | Deep official documentation; precise entity facts; long-form evidence | Never entering the narrow consideration set |
| Gemini | Ecosystem-aligned, structured-data sensitive | Knowledge-graph completeness; schema coverage; consistent entity data | Entity inconsistency silently capping retrieval |
| Perplexity | Citation-dense, freshness-weighted, flat trust curve | Current, crawlable, data-rich pages; original research | Losing answers to fresher third-party sources |
Table 4 — Per-family GEO playbook derived from the benchmark signatures.
The families reward different layers of the Entity Authority Model: Gemini rewards Identity, Claude rewards Evidence, GPT rewards Consensus, Perplexity rewards Freshness. A complete GEO programme covers all four layers — which is why shortcuts optimised for one assistant routinely fail on the others.
Key takeaways
- One playbook per family; four layers overall.
Business implications
- Budget allocation should follow audience assistant mix and current funnel leaks.
Recommendations
- Run the AI Visibility Funnel per family; invest where your deepest per-family leak is.
Section 08Conclusion & References
Treating AI assistants as a single channel is the most common measurement error in early GEO practice. The four frontier families exhibit distinct, stable behavioural signatures across recommendation breadth, citation posture, community sensitivity, consistency and failure modes. Brands that measure per family — and sequence their investments against each family's signature — convert the divergence from risk into opportunity.
References
- KernelX Labs Research. We Analysed One Million AI Responses. Foundational methodology: corpus design, metrics, frameworks. Read the methodology paper.
- KernelX Labs Research. The State of GEO — Q3 2026. Longitudinal trends on the same benchmark framework. Read the quarterly.
- KernelX Labs Research. The AI Visibility Index 2026. Illustrative rankings computed with the Visibility Score. View the rankings.
- Public documentation and system cards of the evaluated assistant families (versions as of the collection window).
Key takeaways
- Per-model measurement is the foundation of credible GEO work.
Business implications
- Aggregate dashboards without family breakdowns will mislead investment decisions.
Recommendations
- Adopt the shared methodology so results are comparable across teams and quarters.