The 2026 NYC AI Recommendation Index: 13,368 Extracted Firm Citations Across 8 Markets

What the index measures

The NYC AI Recommendation Index compares how frequently named service providers appear when buyer-style questions are sent to four assistants. The current edition covers divorce, personal injury, immigration, plaintiff-side employment, estate planning, cosmetic dentistry, IVF, and cosmetic surgery.

Across a consistent in-memory rebuild of the live extraction corpus, Atlas counted 13,368 ranking-eligible firm citations among 3,172 ranked firm records. The assistants were OpenAI GPT-4o, Anthropic Claude Sonnet 4.6, Google Gemini 2.5 Pro, and Perplexity Sonar Pro. The unit is an extracted, resolved firm appearance—not a click, displayed source citation, client, or estimate of consumer usage.

The index asks three questions: how much recommendation volume the leaders hold, how many assistants recognize those leaders, and how deep the named competitive field is. Those variables describe different market structures and should not be compressed into one score.

Concentration ranges from 23.6% to 57.1%

Plaintiff-side employment is the most concentrated category in the reconciled panel. Its top five firms receive 57.1% of ranking-eligible citations. Immigration follows at 47.5% and divorce at 44.6%.

IVF sits at 37.7%, estate planning at 37.1%, cosmetic surgery at 29.7%, cosmetic dentistry at 25.3%, and personal injury at 23.6%. The expansion of citation-only entity extraction materially widened several fields compared with the earlier exported snapshots.

The top-five share in plaintiff employment is about 2.4 times the personal-injury figure. That does not make one category “better” for GEO. It describes concentration under one consistent extraction revision.

Consensus creates a second market map

Concentration does not tell us whether assistants agree. Among each category’s top ten, nine IVF firms had four-model coverage. Personal injury had eight, cosmetic surgery seven, estate planning six, divorce four, immigration four, plaintiff employment four, and cosmetic dentistry two.

The model-level leaders make the split clearer. None of the eight categories had the same provider lead all four assistants. Estate planning and cosmetic dentistry produced four different leaders. The full assistant-by-assistant comparison shows why an aggregate winner should never be described as a universal one.

A concentrated market with low consensus is not the same as a concentrated consensus market. The former can be dominated in aggregate while remaining vulnerable to provider choice. The latter has an incumbent set reinforced across retrieval systems.

The eight-category scorecard

Plaintiff-side employment: 792 citations, 186 ranked firms, 57.1% top-five share, and four four-model firms in the top ten.

Immigration: 731 citations, 202 firms, 47.5% top-five share, and four four-model firms.

Divorce: 912 citations, 191 firms, 44.6% top-five share, and four four-model firms.

IVF: 5,302 citations, 677 firms, 37.7% top-five share, and nine four-model firms.

Estate planning: 1,225 citations, 584 firms, 37.1% top-five share, and six four-model firms.

Cosmetic surgery: 2,003 citations, 516 firms, 29.7% top-five share, and seven four-model firms.

Cosmetic dentistry: 833 citations, 412 firms, 25.3% top-five share, and two four-model firms.

Personal injury: 1,570 citations, 404 firms, 23.6% top-five share, and eight four-model firms.

How to use the index

A firm should compare itself with its category’s fifth- and tenth-place thresholds, then inspect model coverage and prompt-stage performance. A five-percent share may be meaningful in one field and ordinary in another. A two-model position may be an opening in immigration but a conspicuous weakness among IVF leaders.

Use the index to choose the competitive question, not to prescribe content automatically. The missing asset might be a service page, practitioner record, proof page, location signal, or credible third-party evidence. The benchmark field note translates these market shapes into target-setting guidance.

The next editions should add repeated snapshots, normalized diversity, and source provenance. Those additions will show whether a market structure persists, which assistants create it, and what evidence accompanies the recommendations.

Method and limitations

The index selects the most complete run on each active prompt set, prioritizing full bucket and model coverage over recency. It excludes three superseded divorce methodologies and a 135-response, two-bucket IVF retry. Rankings are rebuilt in memory from live citations so older and newer extraction revisions are not mixed.

The current-state panel contains 7,637 responses across all buckets and 6,101 ranking-eligible responses. The underlying runs were collected June 26 through July 13 and reconciled August 11, 2026. Entity extraction expanded substantially after the original exports and still requires ongoing junk-entity QA. See the Atlas methodology for ranking details.