What seems to drive carrier placement in LLMs.
This is a running observation log, not a study. Each entry is something we’ve seen move a carrier’s placement inside LLM answers across ad-hoc analyses on US insurance intents. Each has an impact label, a concrete carrier example, and — where available — links to the pages LLMs actually cite when they return the winning carrier, so you can read the reference material and see what the best-example content looks like.
Labels: Seems to have impact = cross-LLM replication at ≥3/5 on both LLMs tested. Small / conditional impact = replicates in some cases, not all. No observed impact = does not replicate across runs or across LLMs.
Seems to have impact
Property-type naming in the query (condo, luxury, historic)
In home insurance, the same carrier wins the same property-type query across independent LLMs. Condo → Nationwide at 5/5 on Perplexity AND 5/5 on Gemini. Luxury home and historic home → Chubb at 5/5 + 5/5. Consistent across two same-day sessions.
Emerging-category presence (EV in auto)
Electric-vehicle auto insurance → Travelers at 4/5 on Perplexity AND 5/5 on Gemini. Neither the EV-native insurtechs (Root, Lemonade) nor Tesla Insurance wins this query. Travelers captured the category editorially before anyone else.
Category editorial depth on a specialty use-case (even for a generalist)
In auto, rideshare → State Farm (3/5 + 5/5), SR-22 → GEICO (5/5 + 3/5). The actual market specialists — Progressive for rideshare, Dairyland for SR-22 — do not win these queries. The generalist carriers with deeper consumer-facing content have annexed the specialty surface.
The observation is about editorial depth, not about being the category specialist. Being the actual specialist is not enough on its own (see the No-observed-impact section below).
Regional monopoly — strictly line-specific
Newark → NJM at 4/5 + 4/5 on auto. Boston → MAPFRE 5/5 + 5/5 on home. Both carriers have overwhelming state-level market share in the specific line.
Caveat: the same carrier does not win across lines. MAPFRE owns Boston home at 5/5 + 5/5 but does NOT own Boston auto — both LLMs pick State Farm there. Line-of-business dominance is required.
Being cited on Perplexity-favorite comparison sites (NerdWallet, Bankrate, The Zebra)
Across 8 US cities × 4 LLMs in an earlier multi-city probe, roughly 58% of Perplexity's unique citations are consumer-comparison sites (NerdWallet, The Zebra, Bankrate, Policygenius, Insurify). Every winning carrier in the observations above is named on at least one of these sites in the matching category.
Small / conditional impact
Specialty-peril coverage — only where the category is saturated
High-value jewelry → Chubb at 4/5 + 4/5 (clean cross-LLM). Earthquake → GeoVera replicates on Perplexity and Claude in Stage 1 ablation. But flood and wildfire do not replicate cleanly — the specialty surfaces generalists by default on at least one LLM.
No observed impact
Being HQ'd in the query city
Tested across 5 HQ cities. Only 1 of 10 observations matched. Bloomington IL → State Farm surfaced only on Gemini; Perplexity gave USAA. Columbus OH (Nationwide HQ) → Cincinnati and Erie, not Nationwide. Boston MA (Liberty Mutual HQ) → MAPFRE on both LLMs. HQ address has no observed lift.
Being the actual market specialist — without matching editorial depth
Hagerty (largest US classic-car insurer) does not win the LLM classic-car query. Dairyland (largest SR-22 monoline writer) does not win SR-22. Progressive (rideshare-insurance category pioneer) does not win rideshare. Tesla Insurance does not win the Tesla query on Perplexity. In auto, being the specialist is not enough — editorial surface area wins.
This is the inverse of the specialty-use-case-editorial-depth observation. Specialists can recover their surface by closing the editorial gap, not by leaning on brand.
How these observations are made
Each entry above comes from an ad-hoc analysis: a set of buyer-shaped queries run against two LLMs (Perplexity Sonar Pro + Gemini 2.5 Pro, both with web-search / grounding enabled). Stable queries get 5 runs per LLM. An observation enters this list when the same first-named carrier returns with ≥3/5 stability on both LLMs. Reference pages are the actual URLs cited in the LLM’s answer. Raw data is committed to the Phidea repository.
Deeper session notes behind each observation: