Skip to content

reportAI GovernanceVendor Watch

Who searches for legal AI, what they ask, and who answers: the search-demand record, September 2026

Ten Semrush keyword sets, 998 rows, US and UK: brand and vendor-comparison queries dominate; incident, error-rate and dated-regulation queries go unserved.

HarveyLegoraLexis+ AI

Edited and verified by Cognesio LLP

Researched with AI assistance · sources verified by Cognesio LLP · How this was made ↓

This report records what people typed into Google about legal AI in the United States and the United Kingdom, as measured by Semrush on 5 September 2026, and what the first page of results gave them. It is built from ten keyword sets (998 rows, published as CSV at /data/search-demand-2026-09/), two Keyword Overview country splits, and same-day checks of the first page of results for eleven head queries. It covers demand, not opinion, and it treats this site’s own position as one data point among others.

The finding is a mismatch. The largest blocks of demand are product names and product comparisons, and they are answered by the products themselves. The queries a professional risk officer asks, about sanctions, error rates, court rules by jurisdiction and what a vendor actually documents, are smaller, cheaper to rank for, and answered mostly by vendors paraphrasing two primary sources. India now searches the head term nearly as often as the United States.

Key findings

  1. As of 5 September 2026 the seed “legal ai” returns 15,713 keywords with a combined 569,840 searches a month in the US database at an average difficulty of 41; the same seed in the UK database returns 6,444 keywords and 58,340 searches, roughly a tenth.
  2. Country split for “legal ai” (Keyword Overview, 5 September 2026): global 29.5K a month, United States 9.9K, India 8.1K, United Kingdom 2.4K, Australia 1.3K, Canada 1.0K, Germany 720. On 15 July 2026 this site’s keyword research recorded global 30.0K, United States 12.1K and India 6.6K: the India figure rose and the US figure fell in seven weeks.
  3. Brand navigation is the largest single block: “lexis” 60,500, “harvey” 49,500, “harvey ai” 27,100, “paxton” 22,200, “legora” 12,100, “lexis plus” 8,100. The vendor’s own site ranks first for each.
  4. For the head term “legal ai”, nine of the ten first-page results on 5 September 2026 are vendor or product pages (Cetient, Legora, LawConnect, LexisNexis, AILawyer, Darrow, Paxton, MyCase and LexLegis) and the remaining one is an American Bar Association page of ChatGPT tips; an AI Overview sits above them.
  5. The “harvey ai” seed carries 169 “vs” keywords, 85 “pricing”, 83 “review”, 58 “evaluation” and 41 “competitors”; the first page for “harvey ai review pricing vs legora” is seven competitor, reseller or affiliate pages that disagree with each other on price.
  6. Incident-shaped queries have measurable volume at low difficulty: “courts punish lawyers citing ai-generated fake cases in legal filings” 320 a month at difficulty 42, “california lawyer ai misconduct” 210 at 43, “former federal prosecutor reprimanded for submitting ai generated court brief” 140 at 33. The first page for “ai hallucination cases lawyers sanctioned” is the ABA Journal, one law firm, and five vendor pages styled as trackers.
  7. Dated variants of category terms are the cheapest demand in the record: “legal tech news december 2025” 480 a month at difficulty 7, “legal tech news october 2025” 320 at 7, “eu ai act news today 2026” 3,600 at 35 against “eu ai act news” 90,500 at 76.
  8. Twenty-six of the 100 “legal ai” question keywords, and five of the 100 broad-match keywords, carry no difficulty score at all. These are long strings with real volume (“expert witness discipline taxonomy legal ai” 4,400; “legal ethics ai professional responsibility uk” 2,400) that read as the sub-queries an AI answer engine issues, not as typed searches.

Why this question

A publication that records what courts and regulators do about AI in legal practice needs to know two things about its readers before it writes anything: what they are asking, and who is already answering. The first is measurable from the keyword database. The second is measurable from the first page of results. The vendors who hold this data use it to write comparison pages; this report is an attempt to publish the demand side in the open, with the rows attached.

The question also has a public-interest edge. If a lawyer facing a show-cause order searches for what other lawyers were sanctioned and finds a legal-AI vendor’s “tracker”, the incentives of the answer are not the incentives of the question. This report measures how often that happens.

Method and data

Tool. Semrush, read live in a browser on 5 September 2026. Keyword Magic Tool in broad match, sorted by monthly volume, first page of 100 rows for each seed; the Questions filter for one seed; Keyword Overview for country splits and the cached first page of results for two head terms; Domain Overview for three domains. No API was used.

Seeds and databases. “legal ai” (US broad match, US questions, UK broad match), and “ai lawyer”, “harvey ai”, “ai hallucination”, “eu ai act”, “legal tech”, “chatgpt lawyer”, “ai court” in the US database. Ten sets, 998 rows after de-duplication within each set. Each row records the keyword, Semrush’s intent flag, monthly volume, keyword difficulty (KD, 0 to 100), cost per click in US dollars, the count of additional SERP features and Semrush’s result count.

Result-page checks. Semrush’s cached first page for “legal ai” and “eu ai act news” (US), and a same-day web search for “ai hallucination cases lawyers sanctioned”, “legal tech news”, “harvey ai review pricing vs legora”, “ai hallucination rate legal research tools benchmark”, “law firm generative AI policy survey adoption 2026”, “legal AI tools directory vendors security training data”, “court rules AI disclosure requirements by state”, “chatgpt legal advice accuracy study” and “AI note takers privilege confidentiality risk lawyers”. Intent counts below use Semrush’s first-listed intent flag; CPC medians are rounded half up. Each result was classified as vendor or product page, law-firm or bar page, news outlet, academic or primary source, or reference site.

Limits. Volumes are Semrush estimates from clickstream and search-engine data, not Google’s own figures, and are twelve-month averages. Only the first 100 rows of each seed were captured; the “legal ai” seed has 158 pages. KD is Semrush’s estimate of the referring domains needed to rank and is missing (“n/a”) where Semrush has not computed it. First-page results move weekly; the two Semrush lists are cached results whose lag behind the live page is unknown; the classifications are for 5 September 2026 only. The site’s own metrics come from the same Domain Overview and are reported without adjustment.

Disclosure. SafeLegalAI is published by Cognesio LLP, whose team also builds LegalAI Space, a governance product listed in this site’s tools directory with an ownership label. LegalAI Space is not counted, ranked or cited in this report; the report’s data would not change if it were.

The shape of the demand

Seed (database)Keywords in setVolume of set per monthAverage KD of setRows capturedVolume of the 100 rowsMedian volume of a row
legal ai (US, broad)15,713569,84041100319,4401,300
legal ai (US, questions)1,09110,000381007,54040
legal ai (UK)6,44458,3404610031,580170
ai lawyer (US)1,907220,0204599131,170590
harvey ai (US)2,050139,78045100118,330210
ai hallucination (US)3,64789,1305810061,130260
eu ai act (US)16,1212,138,57050992,037,900480
legal tech (US)5,956138,36028100100,730320
chatgpt lawyer (US)48740,6204410030,130210
ai court (US)1,47941,4204810033,980110

The “eu ai act” set is dominated by one row, “artificial intelligence” at 1,830,000 a month, which Semrush includes because the seed’s words are contained in it; the set’s legal-AI content is the other 98 rows. Setting the “eu ai act” set aside, the “ai lawyer” set is the second largest by volume, and it is consumer demand: “free legal advice online chat” 4,400, “talk to a lawyer for free 24/7 chat online” 1,600, “ask a lawyer online free” 590. The “legal tech” set has the lowest average difficulty of any seed (28) and the highest share of rows under KD 30 (36 of 100).

The intent flags tell the same story from a different angle. Semrush marks 62 of the 100 “legal ai” rows informational, 18 navigational, 10 commercial and 6 transactional; in the “harvey ai” set 32 of 100 are navigational; in the “ai lawyer” and “chatgpt lawyer” sets 21 and 22 of 100 are commercial. The Keyword Overview strategy panel splits the head term “legal ai” itself 92 percent informational and 8 percent commercial, while its sub-clusters “ai lawyer free”, “legal ai software” and “ai for lawyers” run 64 to 100 percent commercial.

Seven clusters

Brand navigation

KeywordVolumeKDIntent
lexis60,50083navigational
harvey49,50080navigational
harvey ai27,10089navigational
paxton22,20064informational
legora12,10048informational, navigational
lexis plus8,10053navigational
lexis+6,60051navigational
alexi5,40042informational
lexis ai5,40048commercial
eve legal2,90052navigational
lexisnexis ai2,40050commercial
legora ai1,90034navigational
paxton ai1,90029navigational
cetient1,30030navigational, commercial
legalfly (UK)59024navigational

The brand rows are the top of every set that contains them, and the search engine sends them to the brand. What is reachable behind a brand is the “what is” and “news” tail: “what is harvey ai” 480 (KD 71), “what does harvey ai do” 140 (KD 69), “what is legora” 140 (KD 40), “harvey ai news” 720 (KD 62), “legora product updates” 1,300 (KD 33), “luminance legal ai news” 1,900 (KD 14). This site’s vendor hubs exist for that tail; the Domain Overview shows none of them ranking yet.

News category terms

KeywordVolumeKD
legal tech news33,10061
legal tech news today8,10059
legal tech ai news5,40045
legal ai news2,40046
ai law news2,40045
legaltech news1,90047
ai legal tech news1,60050
legal tech ai news today1,30037
legal tech news updates1,30052
ai legal news88045
law firm technology news72028
legal tech news 202648028
legal tech news december 20254807
legal tech news november 202539034
legal tech news october 20253207

The first page for “legal tech news” on 5 September 2026 was Artificial Lawyer’s homepage, Legal Dive’s legal-technology topic page, the LegalTech Fund’s insights page, Law360 Pulse’s legal-tech section, Legal Tech Monitor, Law360’s technology section and a legal-tech market-news page from New Market Pitch, a vendor. Artificial Lawyer’s Domain Overview the same day: Authority Score 36, 9,000 organic visits a month (down 3.1 percent), 4,300 organic keywords, 4,500 referring domains and a 23 percent share of the traffic Semrush attributes to the category. Its AI Search panel showed visibility 14 and 31 mentions, with reddit.com and lawnext.com among the sources most often cited alongside it.

Two facts about this cluster are new since July 2026. The compound forms with “ai” in them (5,400 + 1,600 + 1,300 + 880) now carry more combined volume than “legal ai news” itself. And the month-stamped forms have appeared with difficulty scores of 7: a searcher typing “legal tech news december 2025” wants an archive page for that month, and almost nobody has built one.

Vendor due diligence

Within the “harvey ai” set, the word groups are: vs 169 keywords, pricing 85, review 83, evaluation 58, comparison 56, valuation 52, features 53, competitors 41, adoption 24. The rows:

KeywordVolumeKDCPC (USD)
harvey ai pricing2601711.27
harvey ai vs casetext vs lexis+ ai vs cocounsel210n/a0.00
harvey ai competitors1401511.70
harvey ai cost1401911.03
how much does harvey ai cost1401611.68
harvey ai reviews1102212.40
legal ai companies480348.99
legal ai startups (recorded in the study, outside the published rows)26029not captured
legal ai platforms (recorded in the study, outside the published rows)2606327.84
legal ai software1,0005829.92

The first page for “harvey ai review pricing vs legora” was: Spellbook’s “Legora vs Harvey” brief, Contracko’s comparison, Bindlegal’s “Top 7 Harvey AI alternatives”, Layer3labs’ comparison, two Vaquill pages and Toolradar. Every result is a competitor, a reseller or an affiliate directory. They disagree on the central fact: one (Vaquill, read 5 September 2026) states Harvey at USD 1,200 to 2,000 a seat a month with a 25-seat, 12-month minimum and Legora at about USD 3,000 a user a year with a 10-seat floor; another (Layer3labs, same day) gives Legora at USD 200 to 500 a user a month. These are analysts’ estimates of unpublished enterprise pricing, re-reported by pages that cite each other. None links to a vendor’s own pricing page, because, as this site’s tools directory records, most vendors do not publish one: of 47 verified independent tools, 8 publish a price.

The cost-per-click column is the market’s own estimate of what this demand is worth: USD 11 to 30 a click for the pricing and platform queries. The vendors bidding on those clicks are the same vendors whose comparison pages rank organically.

Hallucination: definitions are taken, measurement is not

KeywordVolumeKD
ai hallucination9,90083
ai hallucinations3,60092
ai hallucination examples1,30074
what is ai hallucination1,00074
llm hallucination1,00062
ai with lowest hallucination rate39043
grok ai hallucination rate 202621024
how often does ai hallucinate17071
hallucination detector17024
chatgpt hallucination rate17076
ai hallucination rate14069
ai hallucination rates14053
what is it called when ai makes up facts14081
sullivan and cromwell ai hallucinations590n/a
sullivan cromwell ai hallucination21026

Forty-three of the 100 “ai hallucination” rows have difficulty 70 or above; the definitional head is owned by IBM, Google, Wikipedia and the model vendors, and this site’s keyword rules exclude it. The measurement tail is different: the word groups “rate” (499 keywords), “benchmark” (130), “detection” (104) and “lowest” (95) describe a searcher who already knows what a hallucination is and wants a number. The first page for “ai hallucination rate legal research tools benchmark” was LLRX and AI Law Librarians (a February 2026 explainer), the Magesh et al. paper in the Journal of Empirical Legal Studies (paywalled; its open preprint reports hallucination in 17 percent of Lexis+ AI responses, 33 percent of Westlaw AI-Assisted Research responses and 43 percent of GPT-4 responses on the authors’ 202 queries), a general hallucination-rate aggregator and three arXiv preprints from May and June 2026 on legal hallucination detection, claim-level retrieval benchmarks and pro se legal reasoning. The primary study dates from 2024 to 2025; nobody had re-run it against 2026 tools on that page.

A named incident now carries its own volume. “sullivan and cromwell ai hallucinations” (590) and “sullivan cromwell ai hallucination” (210, KD 26) appear in the hallucination set five months after the bankruptcy-court filing they refer to.

Incident-shaped queries

KeywordVolumeKDSet
courts punish lawyers citing ai-generated fake cases in legal filings32042chatgpt lawyer, ai court
lawyers warned about presenting ai-generated false material in legal arguments26058chatgpt lawyer
california lawyer ai misconduct21043chatgpt lawyer
steven schwartz lawyer21020chatgpt lawyer
former federal prosecutor reprimanded for submitting ai generated court brief14033chatgpt lawyer, ai court
chatgpt lawsuit woman fired lawyer fake case law14020chatgpt lawyer
ai lawyer in court14056chatgpt lawyer, ai court
lawyers using ai14056chatgpt lawyer, ai court
lawyer sanctioned11046chatgpt lawyer
lawyer uses chatgpt in court11069chatgpt lawyer
georgia prosecutor suspended for submitting fake ai citations in court110n/aai court
can chatgpt be used in court11017chatgpt lawyer
us federal judges ai use39038ai court
ai judge39054ai court
ai legal defense evidence39023ai court
ai judges26034ai court
ai perpetuating bias and discrimination in court21049ai court
ai of deceased in court cases makes statement17024ai court
ai transcription accuracy vs human transcription court cases1705ai court

Several are headlines typed from memory. The first page for “ai hallucination cases lawyers sanctioned” was the ABA Journal (“sanctions ramping up in cases involving AI hallucinations”), Norton Rose Fulbright (“update on Gen AI sanctions in 2026”), and then five pages from legal-AI vendors presenting themselves as trackers: GC AI, Vaquill, Legal AI Governance, HAQQ (“the 1,598-case sanctions tracker”) and TheLawGPT. Each paraphrases Damien Charlotin’s database of court decisions. The database itself, damiencharlotin.com, had Authority Score 27, 155 organic keywords and 287 visits a month on the same day.

The same pattern holds for court rules. The first page for “court rules AI disclosure requirements by state” was Spellbook, Trace.law, Legal AI Compliance, Claude for Lawyers, Hintyr (“300+ rules”) and AI Vortex, all of which cite one law-firm tracker of United States judge-level orders. The site’s disclosure map is built from a different and larger dataset; on 5 September 2026 it did not rank.

Regulation: dated and enormous

KeywordVolumeKD
eu ai act news90,50076
eu ai act news today40,50063
eu ai act6,60087
ai regulation3,60084
eu ai act news today 20263,60035
eu ai act updates2,90078
eu ai act update april 20261,30051
eu ai act implementation news1,00037
eu ai act news 20261,00032
eu ai act summary1,00070
eu ai act timeline infographic88036
eu ai act deployer obligations log retention 6 months72025
eu ai act enforcement april 202672037
eu ai act risk categories39045
eu ai act fines penalties 202639038
eu ai act compliance32039

The two largest rows, “eu ai act news” and “eu ai act news today”, were flagged in this site’s July 2026 research as almost entirely US-recorded and possibly inflated by automated query traffic; they are reported here as Semrush reports them. Fifty-eight of the 99 rows in the “eu ai act” set are six words or longer, and 38,120 of the set’s non-generic volume sits in those long rows, almost all of them month-stamped. The first page for “eu ai act news” (Semrush cache, 5 September 2026) was artificialintelligenceact.eu, which Semrush credits with 17,500 visits a month from that one URL; the European Commission’s regulatory-framework page; IAPP; Travers Smith’s “state of play”; IBM; Euronews’ tag page; euaiact.com; a YouTube explainer; White & Case’s global tracker; and the Commission’s policy page. One reference site and one law-firm tracker take most of the demand, and the legal-practice slice (what a deploying law firm must log, from when, under which article) is on none of those pages. It is exactly what the 720-a-month string “eu ai act deployer obligations log retention 6 months” asks.

Queries with no difficulty score

Semrush now reports long strings with real volume and no KD. In the “legal ai” broad set, five of 100 rows; in the questions set, 26 of 100.

KeywordVolumeKD
expert witness discipline taxonomy legal ai4,400n/a
legal ethics ai professional responsibility uk2,400n/a
harvey luminance netherlands case study law firms2,40031
gdpr legal ai system ireland data privacy solicitor client1,30013
confidence scoring explainability ai recommendation legal1,00017
thomson reuters legal ai product news past 7 days880n/a
ai ethics legal compliance operational governance720n/a
ai in legal practice evidence integrity72044
in-house legal contract triage playbook ai approvals (UK)720n/a
how to sell ai legal document analysis software to clients72022
streamline ai in-house legal team size (recorded in the study, outside the published rows)590n/a
ai chatbot legal liability (recorded in the study, outside the published rows)39021
ai note takers legal risks (recorded in the study, outside the published rows)320n/a
how ai overviews change legal local search behavior3901

Some of these are noise (the site’s keyword rules already exclude “case study law firms” and “with no manual intervention” shapes as synthetic). Others read like the questions a governance lead types into an assistant, and their shape is the shape of the sub-queries that Google’s AI Mode, ChatGPT search and Perplexity issue when a user asks a long question. They cannot be ranked for in the ordinary sense. They can be answered: a page with a dated table, a named primary source per row and a plain answer paragraph is what an answer engine quotes. On 5 September 2026 this site’s AI Search visibility was zero; the pages that would change that are the ones these strings describe.

Questions

The Questions filter on “legal ai” returns 1,091 keywords with 10,000 searches a month between them; the median row in the top 100 is 40 a month. The shape matters more than the volume.

KindExamples (volume, KD)Site’s position
Career anxietyare lawyers going to be replaced by ai (140, 26); will ai take over lawyer jobs (90, 40); how will ai affect the legal field (140, 63)Excluded by rule
Reliabilitywhat are the limitations of ai legal drafting (110, 19); what are the risks of using ai for legal drafting (50, 26); how accurate are ai generated legal citations (30, n/a); can ai give legal advice (40, 17); can chatgpt be used in court (110, 17)In lane
Practice, descriptivehow are law firms using ai (70, 61); do lawyers use ai (30, 58); what tools do lawyers use (110, 10)In lane
Buyingwhat is the best ai for lawyers (70, 57); which ai is best for legal research (50, 53); what is the best ai tool for legal drafting (30, 42, CPC 26.34)Excluded by rule
Vendorwhat is legora (140, 40); what does legora do (50, 29); how much does lexis ai cost (30, 27)Vendor hubs

The United Kingdom

The UK set is a tenth of the US set by volume and differently shaped. Head terms: “legal ai” 2,400 (KD 69), “ai lawyer” 1,900 (55), “ai and law” 1,300 (61), “ai in legal” 1,300 (61), “ai law” 1,300 (70), “artificial intelligence and law” 1,300 (58), “law ai” 1,000 (64). Then “legalfly” 590 (24), “ai in law firms” 390 (53), “legal ai tools” 390 (63), “ai solicitors” 170 (41), “legal ai uk” 110 (49), “ai lawyer uk” 110 (47), “clio ai” 110 (22), “legal it insider” 140 (13). The word groups include “ethical” (98 keywords), “accuracy” (95) and “risk” (89) where the US list shows “generator” and “free”. The academic phrasing is stronger and the consumer-chat demand weaker. Sixty-seven of 100 rows are informational and 21 commercial.

Where the site stands

Domain, 5 Sep 2026Authority ScoreOrganic visits / monthOrganic keywordsReferring domainsAI Search visibility
artificiallawyer.com369,000 (−3.1%)4,3004,50014 (31 mentions)
damiencharlotin.com272871552,50014 (1 mention)
safelegalai.com204 (+33%)360

This site’s four ranking keywords are all one Australian citation query, and its 36 referring domains are link-farm noise the site does not count. The comparison is the point of the table: the canonical hallucination database ranks for 155 keywords and draws fewer than 300 visits a month while its vendor paraphrases rank above it, and the category’s leading news site takes most of its traffic on its homepage for one term. Neither fact is an argument for chasing the head terms. Both are arguments for the long tail this report has measured: dated variants at KD 7 to 40, incident and measurement queries at KD 5 to 45, and answer-engine strings that reward a table with sources.

What to watch

The quarterly refresh will re-run the same ten seeds and report deltas. Three movements to watch: whether India overtakes the United States on the head term, which the July-to-September movement (India 6.6K to 8.1K, the United States 12.1K to 9.9K) would produce within a quarter if it continued; whether the “n/a” difficulty rows grow as a share of the question set, which would measure the shift of professional queries from the search box to the assistant; and whether the vendor-tracker pages hold the first page for incident queries once the underlying database and this site’s sanctions ledger are cited more often than they are paraphrased.

Three sentences journalists can quote

On 5 September 2026 the Semrush database recorded 15,713 keywords and 569,840 monthly searches around “legal ai” in the United States, and the first page of results for that term was nine product pages and one bar-association page of ChatGPT tips.

India searched “legal ai” 8,100 times a month against 9,900 in the United States; in July 2026 the figures were 6,600 and 12,100.

For the query “ai hallucination cases lawyers sanctioned”, five of the seven first-page results were legal-AI vendors presenting paraphrases of a single academic’s database as their own tracker.

Appendix A: the ten keyword sets

Published as CSV at /data/search-demand-2026-09/, one file per set, with the columns keyword, intent, volume, kd, cpc_usd, serp_features_extra and results. Per-set statistics (top 100 rows):

SetVolume of rowsMedian volumeMedian KDRows with KD n/aRows under KD 30Rows KD 70+Rows with CPC above 0Median CPC where above 0 (USD)Rows of 6+ words
legal ai, US broad319,4401,300525910748.1611
legal ai, US questions7,540404226261477.1872
legal ai, UK31,580170522131904.562
ai lawyer, US131,170590601612948.446
harvey ai, US118,3302104922325749.596
ai hallucination, US61,130260672943580.489
eu ai act, US2,037,9004804751420163.6158
legal tech, US100,73032037.52361607.287
chatgpt lawyer, US30,130210500142847.739
ai court, US33,9801105421112445.2117

Appendix B: sources

Appendix C: changes to this report

None since first publication. Next edition: quarterly, with deltas per set; add the India, Australia and Canada databases for the head term’s long tail; add the seeds “ai governance”, “legora”, “ai legal research”, “ai contract review” and “ai solicitor”.