Clearbrief alternatives, and the question the shortlist never asks: what is each one actually checking?

Clearbrief runs inside Microsoft Word. Its own homepage leads with the line Cite facts, not fake cases and describes the product as AI for legal-grade accuracy in Word, trusted by serious litigators, in-house teams, and courts. The feature it names first is AI-powered fact-checking and hyperlinked evidence, and the mechanic underneath it is the one that matters: it takes the sentences in a draft and looks for the passage in the record or the authority that actually supports each one, then hyperlinks it so a reader can click through. It is distributed through the Microsoft Store, it claims SOC 2 Type 2 certification, and it names state and federal courts and the American Arbitration Association among its users.

Pricing is on the homepage rather than a separate page. The address clearbrief.com/pricing returns a 308 redirect to an anchor on the homepage, where a Solo plan reads 300 dollars per month per user and an Enterprise Unlimited plan reads Custom, with volume discounts available. That puts Clearbrief in a minority: most of the products it gets compared against publish nothing.

Now the part that makes the standard shortlist a bad instrument. Search for Clearbrief alternatives and you will be handed some arrangement of Paxton AI, Alexi, Midpage and Marveri, usually with Lexis and Westlaw somewhere underneath. Read each vendor's own homepage and they are not doing the same job:

• Marveri does not belong on the list at all. Its own homepage headline is AI Due Diligence for M and A, Venture Financings, and Corporate Transactions, and the workflow it describes is connecting a data room and producing diligence memos, request lists, schedules and tie-outs. That is transactional diligence. It has nothing to do with checking whether a cited case exists. Its pricing address returns a 404.
• Alexi is a research and drafting platform, not a checker. It describes itself as one AI platform for legal research, drafting, and analysis, and as private AI built for the enterprise of law. Its pricing address returns a 301 redirect to the homepage and no figure appears anywhere on the site.
• Paxton AI and Midpage are AI legal research products with a citator attached. Both publish prices, which is unusual and to their credit. Neither claims to audit a document you wrote somewhere else.
• Clearbrief is the only one on the list whose primary job is auditing a finished draft. That is a different purchase from a research subscription, and a firm that already pays for Westlaw or Lexis is not choosing between them.

None of that makes any of these products bad. It makes the word alternatives misleading, because four of the five are not substitutes for each other. What follows is the shortlist rebuilt around the only question that separates them, which is what each product actually checks.

Ask a vendor whether its product checks citations and every one of them will say yes. The word covers three separate operations, and the gap between them is where firms get into trouble.

One. Does this case exist? This is the failure everybody knows about, because it is the one that made the news. A model produces a case name, a reporter cite and a quote, and none of it corresponds to anything. Checking it is mechanical: look up the cite in a real database and see whether anything comes back.

Two. Is this case still good law? This is what a citator does. Shepard's, KeyCite, and the AI citators that Paxton and Midpage ship all answer this question. They read later decisions and report how those decisions treated the case: affirmed, distinguished, questioned, overruled. This is a hundred-year-old product category with an AI layer on top.

Three. Does this case actually say what the brief claims it says? This is the check nobody automated for a very long time, and it is the one Clearbrief is built around. A real, currently good case can be cited for a proposition it does not stand for. No citator catches that, because from the citator's point of view nothing is wrong: the case exists and it has not been overruled.

Three checks, three different products, and vendors who answer yes to all of it. Read the documentation rather than the homepage and the vendors themselves draw the line. Paxton's own help documentation, on the page describing its AI Citator, states plainly that a missing treatment badge means Paxton did not flag the case, not that the case has been affirmatively validated, and that the treatment indicator and the underlying analysis are starting points for your professional judgment, not substitutes for it. Its own FAQ adds that a case with no badge still warrants verification before you rely on it. That is a vendor telling you, in its own documentation, that check two is not check one and is not check three.

Hold that distinction in mind through the rest of this page. It is the reason a firm can buy an expensive, well-built, properly grounded research product and still end up in front of a judge explaining itself.

Read down the second column rather than the third. The price is the easy part; what each product checks is the part that decides whether you have bought the thing you thought you were buying. Every row was read on the vendor's own site or documentation on August 18, 2026, and every ownership entry that says no acquisition notice found is stating the absence of a finding rather than proving independence, because a company is under no obligation to publish its ownership on its homepage.

Two rows are worth pausing on. Marveri is the only product here that is not in this market, and it appears on these lists anyway, which tells you how these lists get assembled. And vLex is the only entry whose ownership changed the shape of the decision: it is now part of Clio, so for a Clio firm it is a consolidation rather than a diversification.

Most writing on this subject cites the same one case and then reaches for adjectives. There is a better instrument. Damien Charlotin maintains a public database called AI Hallucination Cases that tracks legal decisions in which generative AI produced hallucinated content. It publishes a CSV. We downloaded it on August 18, 2026 and counted it ourselves rather than quoting anyone's summary of it.

The site's own headline number on the day we read it was 1,922 cases identified so far, last updated August 16, 2026. The CSV export we pulled carried 1,923 rows. Here is what is in it, and some of it runs against the way this subject is usually sold.

• 1,314 of the rows are United States cases. The rest are spread across forty other jurisdictions, with Canada at 211 and Australia at 98 the largest non-US contributors.
• The curve is the actual story. US rows by year: 11 in 2023, 37 in 2024, 525 in 2025, and 740 in the first seven and a half months of 2026. Whatever the profession did in response to the 2023 cases, the count kept climbing.
• The majority of it is not lawyers. Across US rows the party tag reads Pro Se Litigant 787 times and Lawyer 508 times. If you have seen this subject framed as a profession-wide epidemic of attorney misconduct, the database does not support that framing, and we would rather say so than inflate the number we are writing about.
• 508 US rows tagged Lawyer is still a lot. 150 of them carry a parsed monetary penalty. The median of those is 2,209 dollars. The largest is 110,204 dollars, in a District of Oregon matter dated March 23, 2026.
• Judges are in the database too. The party tag reads Judge on 28 rows across all jurisdictions, 8 of them in the US. That is not a comforting detail for anyone assuming the check happens downstream of the filing.

Then there is the breakdown that settles the argument this page started with. The database classifies what went wrong in each case, and counting the CSV at the case level gives 1,602 cases involving fabricated material, 800 involving misrepresented authority, 519 involving false quotes, and 33 involving outdated advice. Our counts reproduce the site's own published filter numbers to within one on the largest category, which is a useful check that we are reading the file the way its author intends.

Look at what those four buckets correspond to. Fabricated material is check one, does this case exist. Misrepresented authority and false quotes are check three, does this case say what you claimed it says. Outdated advice, at 33 cases out of 1,923, is check two, the citator's question. The thing every research vendor is selling you a citator for is the rarest failure in the record by a factor of roughly fifty. The two failure modes that dominate are the one a database lookup catches and the one that requires somebody to actually read the case, and neither is what a treatment signal is for.

Two honesty notes about this dataset, both taken from the maintainer's own description of it. First, it tracks decisions where a court found or implied that a party relied on hallucinated content, so it explicitly does not track the wider universe of every fake citation ever filed. The maintainer describes the database as necessarily an undercount for that reason. Second, it is a record of what courts said, not an audit of what tools did. We report it as what it is, which is the best public instrument available, not as a census.

The natural conclusion from the section above is to buy something serious and let the vendor's grounding solve it. The same dataset tests that conclusion, because it has a column recording which AI tool was involved.

We filtered every row in the database whose tool field names a paid legal research or verification product rather than a general-purpose chatbot. Twenty-eight rows name one. The named products include Westlaw and Westlaw Precision, CoCounsel, LexisNexis, Lexis+ AI and Protege, Fastcase, vLex and Paxton AI, and one row's tool list includes Clearbrief alongside four other products. Twenty-six of the twenty-eight involve a Lawyer party rather than a pro se litigant, which is what you would expect, because these are products lawyers pay for.

Three things must be said about that number immediately, because it is the kind of finding that gets misused.

• The database itself warns against reading it as blame. The tool column carries a note from the maintainer stating that the mention of a specific tool does not necessarily mean that tool was responsible for the hallucinations in question. In many of these matters a lawyer used several tools and the court never established which one produced the bad cite.
• Nine of the twenty-eight are marked Vendor Disputed. Every disputed row in the entire 1,923-row database sits inside this group of 28, which makes sense: vendors only file a response when they are named. In one, the database records that Thomson Reuters contends none of the alleged hallucinations originated in Westlaw. Those disputes are recorded, not resolved, and we are not resolving them either.
• The point is not that these products are bad. The point is narrower and harder to argue with. Even among lawyers who had paid for a grounded, citator-backed research product, the verification step still failed often enough to appear twenty-eight times in a public record of court findings. Whatever a firm buys, somebody still has to check, and that somebody is inside the firm.

Read the orders rather than the headlines and the fake-case story turns out to be the simple half of the problem. Take a recent one we pulled and read in full.

In Chakma v. Sushi Katsuei, Inc., No. 23 Civ. 7804 (KPF), Judge Katherine Polk Failla of the Southern District of New York issued an opinion and order on May 19, 2026 sanctioning defense counsel. The opinion records that at a hearing, counsel acknowledged she had used LexisNexis's AI tool to help construct multiple submissions to the Court and then failed to review the final versions for correctness, and that certain of the citations resulted from AI hallucinations. On the AI portion the court ordered 1,710 dollars in opposing counsel's fees and imposed a separate 1,000 dollar monetary penalty.

But the sentence worth the whole page is elsewhere in that opinion. Reviewing three of counsel's letters, the court noted that all cited cases seem to be irrelevant to the points they were offered for, and observed at the hearing that there was no case cited that actually stands for the proposition. Only some citations were entirely hallucinated. The rest were real cases, presumably still good law, that did not support the argument they were attached to.

A citator would have passed every one of those real cases. Shepard's would have reported them valid. KeyCite would have shown no negative treatment. The check that fails there is check three, and it is the check almost nothing in the market automates, which is precisely the space Clearbrief occupies and precisely why treating it as interchangeable with a research subscription is an error.

The original case is worth reading for the same reason. In Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Judge P. Kevin Castel imposed a 5,000 dollar penalty jointly and severally on two attorneys and their firm on June 22, 2023, paid into the registry of the court. The opening paragraph of that opinion is the most quotable thing any judge has written on this subject and it is not the part people quote: there is nothing inherently improper about using a reliable artificial intelligence tool for assistance, but existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings, citing Rule 11. The court did not have a problem with the tool. It had a problem with an unowned gatekeeping step.

The most useful reading in this category is not the homepage. It is the documentation and the fine print, where the same vendors are considerably more careful than their marketing.

Paxton AI is the most honest of the group and it costs them nothing. Its citator help page tells users that a missing treatment badge means Paxton did not flag the case, not that the case has been affirmatively validated, and that outcomes such as Affirmed, Praised, Followed and Neutrally Cited all display with no badge, so the case still warrants verification before you rely on it. On accuracy, Paxton publishes a real number against a real benchmark, which almost nobody in this category does. Its post reports 93.82 percent average accuracy across a random sample of 1,600 tasks drawn from the roughly 750,000 tasks in the Stanford legal hallucination benchmark, and an average non-hallucination rate of 94.7 percent. Note two things. The headline on that same page reads 94 percent plus while the table underneath reads 93.82 percent. And a 93.82 percent accuracy rate means roughly one task in sixteen was not accurate, on a benchmark that includes checking whether a case exists. That is a good score. It is not a number a firm can file behind.

LexisNexis published the strongest claim anyone has made here and then quietly stopped making it. The October 25, 2023 press release announcing Lexis+ AI is still live on lexisnexis.com under the headline Hallucination-Free Linked Legal Citations, and its body says the product checks all citations against Shepard's to ensure citation validation. The same paragraph, in the same document, says the product minimizes the risk of invented content. Hallucination-free in the headline and minimizes the risk in the body are not the same claim. The current Lexis+ AI product page does not use the phrase hallucination-free at all; it describes Shepard's Verify as helping ensure authority and reliability. We are not accusing anyone of anything. We are pointing out that the strongest version of the claim lives on a three-year-old press release that is still indexed, still reachable, and still the first thing a buyer finds.

The one independent benchmark in this category did not test any of this. The Vals Legal AI Report, dated February 27, 2025, evaluated four tools against a lawyer control group across seven tasks: data extraction, document question and answer, document summarization, redlining, transcript analysis, chronology generation, and EDGAR research. Citation verification is not on that list. The report also records that Lexis+ AI was initially evaluated but withdrew from the sections studied. So the most cited independent evaluation in legal AI is silent on the exact capability this page is about, and one of the largest vendors is not in it.

And one ownership fact belongs in any 2026 shortlist. vLex, whose Vincent AI shows up on most of these comparison lists as an independent option, now identifies itself in its own page title as vLex, part of Clio. Its own site describes Vincent Studio as Clio's no-code workflow builder. If your firm runs Clio as its practice management system, choosing Vincent is deepening a relationship you already have rather than diversifying away from it. That may well be the right call, since integration with software you already own is genuinely easier. It should be a decision you made rather than one you did not notice, and it is the same trap this site documented on the agency side when two products on one shortlist turned out to share a parent.

Pricing opacity is the norm in legal software and it is worth recording precisely rather than complaining about. We requested each vendor's pricing address on August 18, 2026 and wrote down what came back.

• Clearbrief: clearbrief.com/pricing returns a 308 redirect to an anchor on the homepage, where the number is published. Solo, 300 dollars per month per user. Enterprise Unlimited, Custom, volume discounts available.
• Paxton AI: paxton.ai/pricing returns 200 with real figures. Individual, 499 dollars per user per month, or 2,999 dollars per user per year, which the page describes as saving 50 percent against monthly. Enterprise is custom volume-based pricing.
• Midpage: midpage.ai/pricing returns 200 with real figures and it is the cheapest entry point on this page by an order of magnitude. Starter, 30 dollars a month billed monthly. Pro, 80 dollars a month billed annually, which the page annotates as saving 240 dollars against the monthly option, and which includes 25 dollars a month of credits for its PACER tools and then pay-per-use. Legal Engineer is Custom, annual plans only.
• Alexi: alexi.com/pricing returns a 301 redirect to the homepage. No figure appears on the site.
• Marveri: marveri.com/pricing returns a 404. No figure appears on the site.
• Lexis+ AI and CoCounsel: neither publishes a number. The Lexis+ AI product page states that pricing varies based on the size of your organization, the capabilities required, and the scope of content access, and offers customized pricing.

Three of five publish something. That is a better disclosure rate than most categories this site has measured, and it is worth noticing that the two cheapest published prices belong to the two products that are most straightforwardly comparable to each other.

One arithmetic note for a firm sizing this, using published numbers only. Twenty timekeepers on Clearbrief Solo at 300 dollars per user per month is 72,000 dollars a year. The same twenty on Paxton Individual annual at 2,999 dollars is 59,980 dollars a year. Neither figure is what an enterprise buyer would actually pay, because both vendors quote custom above the solo tier, but it does establish the shape: at any real headcount this is a five-figure annual line, and it sits on top of the Westlaw or Lexis contract the firm already has. We are not aware of any published enterprise rate for either product and we are not going to invent one.

This site sells commissioned AI builds, so the useful thing we can do here is be exact about when that is the wrong answer.

Buy Clearbrief if the job is auditing finished drafts inside Word. That is a real product doing a real job that almost nobody else does, it is sold at a published price, and the integration point is where the work already happens. Rebuilding a Word add-in that maps assertions to record evidence, to avoid a 300 dollar per seat subscription, is the expensive mistake in this category and we would tell you so on the call.

Buy Paxton or Midpage if the job is research. Both publish prices, both ship a treatment citator, and Midpage's entry tier costs less per month than an hour of associate time. Neither replaces the check.

Do not buy Marveri for this. It is a good-looking product for a different problem. If you are doing transactional diligence, it is on the right list. It is not on this one.

The narrow case for commissioning something is not a better checker. It is the part no product on this page sells, because it is not a product: the verification step as an enforced, recorded control inside your own filing workflow. Every tool above is a thing an individual lawyer opens. None of them can tell a managing partner which filings went out last quarter without a documented check, who ran it, what it was run against, and which ones came back with an exception that somebody waived. That is a workflow and a record, sitting on the firm's document management system and its filing calendar, and it is the thing a court asks about after the fact. Judge Starr in the Northern District of Texas made the point structural back in May 2023 with a mandatory certificate requiring counsel to certify that no portion of any filing was drafted by generative AI, or that any AI-drafted language, including quotations, citations, paraphrased assertions, and legal analysis, was checked for accuracy by a human being before submission. A certificate is a control. Somebody has to be able to answer for it.

We have not shipped a commissioned citation-verification system for a law firm. Our legal work has been on intake, conflicts, knowledge retrieval and document workflow, and we would rather write that sentence than imply a track record we do not have. This page is research on what the vendors publish and what the courts have recorded. Where the right answer is to buy Clearbrief and write a two-page policy around it, that is what we will say on the call, and it is a shorter call.

Related reading.

On the rest of the legal stack: what we would build for a law firm, the consultants mid-market firms actually shortlist, and the wider list of AI consultants for law firms. On the research layer this page sits next to, Harvey against a custom build, CoCounsel against Harvey and the citator question underneath it, Legora against Harvey, the Harvey alternatives for a mid-market firm, and the Spellbook alternatives plus Spellbook against a custom build on the drafting side. On the systems a firm already runs, the Clio playbook, which is the one to read alongside the vLex ownership note above, the iManage playbook, the NetDocuments playbook and the Litify playbook, and where Clio's own AI stops. On the neighbouring comparisons, Filevine and case management, EvenUp and demand drafting, and Smith.ai and the intake layer. On governance, which is where the control described above actually lives, bar rules and malpractice, answering client AI questions in outside counsel RFPs, and governance without enterprise theater. On process and cost, hiring versus commissioning, what this costs for a firm, how to choose an AI consultant for law, and the rent versus own arithmetic. Our forward view is in the 2027 law firm benchmark, the maturity assessment places your firm before any of it, and the law firm resource and its diagnostic are the shortest route to a real conversation. The rest of the series sits on the comparisons hub, process covers how we work and pricing covers our own fee bands. If you want a read on your own filing workflow rather than a product recommendation, talk to us.