The legal AI market is crowded with tools making similar-sounding claims — “built for lawyers,” “trained on legal data,” “citation-aware.” For a solo or small-firm lawyer without a procurement team or a research department to sanity-check a vendor’s claims, a short, specific checklist matters more than a long feature comparison.
1. Does the vendor publish an actual error rate?
Ask directly: what is your hallucination rate or false-verify rate, measured how, and updated how often? Independent research (the Stanford RegLab/HAI benchmark study) found meaningful hallucination rates even among premium legal research platforms — roughly 17% for Lexis+ AI and roughly 33% for Westlaw’s AI-Assisted Research in that study — which means “built for lawyers” alone is not evidence of a low error rate. Most vendors today, when asked directly, do not have a published number at all; treat that as informative in itself.
2. Does it separate drafting from verification, or blend them?
A tool that both drafts your brief and verifies its own citations has a structural conflict: the same model that hallucinated the citation is being asked to confirm it. Tools that separate research/drafting assistance from a distinct, deterministic verification step (existence, quote-match, proposition-match run against a primary source, not asked of the same generative model) are structurally more trustworthy for the verification function specifically.
3. What database does it actually check against?
Ask specifically whether verification runs against a real, current case-law database (CourtListener, Westlaw, Lexis) or against the AI model’s own training data, which is functionally circular for this purpose. A tool that verifies citations by asking a language model “is this real?” rather than performing an actual database lookup provides very little additional protection over not verifying at all.
4. What happens on an unmatched citation?
A citation that cannot be found should be flagged clearly as unconfirmed, not silently passed as verified and not silently treated as definitely fabricated. Coverage gaps are real (unpublished trial-court decisions, state-specific or agency decisions like BIA rulings are often outside general-purpose databases) and a tool that cannot distinguish “not found because fabricated” from “not found because outside coverage” will generate false alarms or, worse, false confidence.
5. What does it cost, and does the pricing model fit solo practice?
Premium bundled AI research products (CoCounsel, Lexis+ AI) are often priced for firm-level budgets, sometimes hundreds of dollars per user per month, frequently requiring an existing Westlaw or Lexis subscription. Purpose-built citation verification tools tend to be priced separately and more accessibly for solo use — compare the actual per-brief or per-month cost against your realistic filing volume, not against a firm-scale sticker price that assumes a different practice size.
6. Can you test it on a document with a known fabrication?
Before committing, run a tool against a document containing at least one citation you know is fabricated (you can construct one easily, or use a publicly documented example like the fake Varghese v. China Southern Airlines citation from Mata v. Avianca) and confirm it actually catches it. A vendor demo using only clean, real citations tells you nothing about false-negative performance, which is the risk that actually matters.
7. Does it fit into your actual filing workflow?
A tool that requires exporting your document, running it through a separate portal, and manually cross-referencing results back into your brief will get skipped under deadline pressure, regardless of how good its underlying checks are. Word add-ins or upload-based tools that work directly against the document you are about to file have a much better chance of actually being used consistently.
Reading between the lines of vendor marketing language
Certain phrases in legal AI marketing materials are worth specific skepticism: “trained on legal data” describes the training process, not a measured accuracy outcome; “built by lawyers” describes who built it, not how well it performs; “citation-aware” can mean anything from a genuine database-backed verification layer to simply formatting citations in Bluebook style without checking whether they exist. Ask the vendor to translate any of these phrases into a specific, falsifiable claim you could actually test, rather than accepting the marketing language at face value.
Negotiating a trial period tied to your own test document
Before committing to an annual contract, request a trial period specifically long enough to run the tool against several of your own past filings, including at least one you have deliberately altered to include a known error. Vendors confident in their product's accuracy should have no objection to this kind of structured trial; reluctance to support a real test against your own documents, as opposed to a scripted demo, is itself informative.
Considering total switching cost, not just monthly price
Factor in the time cost of learning a new tool's workflow, migrating any existing document templates, and retraining any staff who will use it, not just the subscription price itself. A slightly more expensive tool that fits cleanly into your existing drafting workflow may have a lower total cost than a cheaper tool that requires you to change how you work in ways you are unlikely to sustain once the novelty wears off.
Reference checks: talking to other solo practitioners who use the tool
Beyond a vendor's own demo and marketing materials, seek out other solo or small-firm attorneys who have used the specific tool you are evaluating, through bar association listservs, legal technology forums, or direct outreach. Unfiltered peer experience, particularly about how a tool performs on real, messy documents rather than clean demo material, is often more informative than anything a vendor will volunteer directly.
Contract terms worth reading closely before signing
Beyond price, review the contract for data retention and confidentiality terms (what happens to documents you upload), any liability limitation language specific to verification accuracy, and cancellation terms if the tool does not perform as expected. A vendor confident in its own accuracy should have little difficulty offering reasonable cancellation flexibility during an initial evaluation period.
A common question
Is it worth paying for a dedicated verification tool if I only file a handful of briefs a month?
For low filing volume, doing the checks by hand using CourtListener’s free existence and full-text search is a completely reasonable option and costs nothing but time — roughly two minutes per citation done carefully. A dedicated tool mainly buys speed and a documented, repeatable process; whether that is worth paying for depends on your own filing volume and how much your time is worth relative to the subscription cost.