Where Firms Disappear from AI Answers: 7,637 Responses Across the Buyer Journey

One category contains several recommendation markets

A person learning what probate means is not asking the same commercial question as a person validating three estate-planning firms. A patient in a post-procedure crisis needs different help from one choosing a cosmetic surgeon. Blending those prompts into one visibility score removes the decision stage that explains much of the answer.

Atlas organizes prompts into eight buyer modes: situation, crisis, learning, advice, constraints, decision, validation, and follow-up. Viclaro selected one most-complete active run per category, covering 7,637 responses, and counted live extracted firm-citation records per response in each bucket.

This analysis includes diagnostic buckets that do not all contribute to the public rankings. Its purpose is to describe when assistants name providers, not to recalculate leaderboard share.

Four case-driven legal markets peak at the decision stage

Decision prompts produced the highest extracted-citation rate in divorce at 2.79 per response, plaintiff employment at 2.67, personal injury at 4.39, and immigration at 2.06.

Learning and advice prompts were generally lower. Divorce learning prompts averaged 0.06 citations per response and advice prompts 0.13. Plaintiff-employment learning averaged 0.08 and advice 0.14. Personal injury was less sparse after extraction reconciliation—1.11 on learning and 0.75 on advice—but still well below decision prompts.

The assistants were not necessarily failing. On learning and advice questions, a useful answer may explain rights, procedures, deadlines, or evaluation criteria without naming a provider. Recommendation absence must therefore be interpreted against prompt intent, not automatically labeled poor visibility.

Constraints reshape the legal shortlist

Constraints nearly matched decision prompts in divorce at 2.78 citations per response. They averaged 1.76 in plaintiff employment, 1.92 in immigration, and 3.96 in personal injury.

A constraint adds facts such as geography, budget, language, urgency, case complexity, or representation type. Those details can make a generic category leader less relevant and create openings for specialists whose public evidence matches the narrower situation.

This is the buyer-side counterpart to Google’s query fan-out. One visible question can trigger retrieval around several subtopics. Our query-fan-out field note explains why the content response should be an evidence portfolio rather than a page stuffed with the head term.

Medical categories follow different paths

IVF produced extracted firm citations broadly across the journey. Situation prompts averaged 6.66 citations per response, constraints 6.03, crisis 5.81, decision 5.25, and learning 4.29. These high rates also warn that expanded entity extraction can identify more names than a human would treat as a clean shortlist.

Cosmetic surgery ranged from 2.38 on constraints and 2.16 on advice to 1.01 on validation. Cosmetic dentistry peaked on constraints at 2.93 and fell to 0.50 on learning.

Estate planning broke from the other legal categories. Constraints led at 2.26 citations per response, followed by crisis at 1.66 and situation at 1.64; decision averaged 0.76. Category labels therefore do not determine a universal funnel.

Visibility strategy should name the stage

A firm should not ask only “How visible are we?” It should ask where it appears: when buyers describe the situation, add constraints, choose a provider, validate a shortlist, or ask what to do next.

A learning-stage citation without a recommendation can still establish useful authority. A decision-stage recommendation without owned evidence may rely on directories or prior model knowledge. A constraint-stage absence may reveal that the site never explicitly supports the specialization buyers are asking about.

Measure each bucket using the same prompt set and denominator across snapshots. Then choose a page or evidence fix tied to the missing stage. The guide to building an AI visibility prompt set explains how to sample buyer situations without disguising a keyword list as research.

Method and limitations

The analysis selects one most-complete run per active taxonomy and joins its 7,637 responses to live firm-citation extraction. It reports extracted citation records divided by responses within each category and bucket. Repeated or noisy extracted entities, canonicalization, different prompt counts, model behavior, and category-specific wording affect the rates.

The result does not measure conversion intent, answer quality, or whether naming a firm was appropriate. It is descriptive evidence that provider-naming behavior changes materially across buyer modes. Comparisons are strongest within a category; cross-category differences should be treated as hypotheses for further controlled study.