“Near Me” Finds a Lawyer. It Doesn’t Really Find Yours.

The search

You typed find a personal injury lawyer near me.

You meant something much more specific. Something like: a lawyer near me who has handled a rear-end collision involving a commercial vehicle, where the other driver’s insurer has already called me twice, where I’m four months into physical therapy and still not right, and who can tell me whether the thing I signed last week was a mistake.

It returned pages of results, every one of them advertising experience, expertise, and a dedication to fighting for you. None of them ordered by anything to do with what happened to you, because the search never asked questions — it just returned results.

This is the single most common way personal legal research begins, and it has a structural flaw that no amount of scrolling fixes: the prompt that’s easy to type is not the question you need answered, and the results aren’t organized in any shape a licensed attorney can use to help you.

Of course location matters — a lawyer has to be licensed in the right state. But that makes proximity a filter, not a relevance ranking. Once every name on the list clears your jurisdictional filter, the difference between four miles and thirty carries no information at all. “Near me” is just a sort order.

So let’s be straight about what this kind of prompt can and can’t give you. It won’t hand you the name of the lawyer who’s handled your specific case — no single search can, and by the end you’ll see why not. What a properly scoped agent can do is show you the two things that are actually inside your control: what you know before you contact a lawyer, and how good a fit you are for the practice you contact. Everything that follows is about those two things, and about what happens when you skip them without knowing you did.

“But I can just ask ChatGPT...”

You can. It’s free. And most people now do: Pew measured chatbot use at 49% of U.S. adults in February 2026, up from 33% two years earlier, with ChatGPT alone reaching 44%. One in five of those users already brings medical questions to a chatbot. Consequential personal problems aren’t an edge case for these tools — they’re a main use.

Legal research is going the same way. A 2026 survey by a legal marketing firm found roughly four in ten people would use ChatGPT to research an attorney, up from under three in ten the year before, while Google slipped from about 87% to 72%. Those are one vendor’s numbers. The direction is the same one Pew is measuring everywhere else: people are abandoning a list of links for something they can describe their situation to — because describing the situation was always the necessity.

So we ran a simple prompt experiment. On 23 August 2026 we put the question from the top of this post into ChatGPT and Google’s Gemini, on their default settings, three times each — twice word-for-word, once lightly reworded. Six runs total: Runs 1–3 are from ChatGPT, Runs 4–6 are from Gemini.

Nine runs by the end of this post; as an illustration, not a survey — too small to establish a true rate, but large enough to show a pattern that repeated every time we looked. Everything below is what we saw, stated as what we saw.

Starting with what they got right: neither agent made anything up. Every run said it couldn’t tell whether the signed document was a mistake without seeing it. ChatGPT asked us to upload it each time. That’s the correct answer.

The problem is elsewhere.

It never asked us anything

Across all six runs, neither assistant asked a single question before recommending lawyers.

None asked where the crash happened. Or who owned the truck. Or whether a police report exists, whether a recorded statement had already been given, whether money had changed hands, or what the signed document was actually titled — the one question that would have answered what was being asked, and an easy one to ask if you knew to ask it.

Those questions are the “first five minutes” of any real consultation. Six times out of six, the shortlist of attorneys came before the facts. Six times out of six, we were handed the answer before the agent understood the case.

It knew where we were — and it was the wrong “where”

We never included a location. Both agents supplied one anyway.

ChatGPT named our town, correctly, and built every recommendation around it — down to a map with pinned offices and star ratings. Some small grey text at the bottom showed the town beside an offer to make the location more precise. It never said where it got it. Gemini didn’t even do that: no town, no notice, just firms from the surrounding two counties as though we’d asked.

But a lawyer has to be licensed where your matter belongs, and what usually decides that is where the collision happened or where the other side can be sued — not where you’re holding your phone. Both systems guessed our whereabouts from an IP address and treated it as an answer about jurisdiction. That’s the same error the original search makes, one level deeper: not just proximity used as a ranking, but the wrong proximity, inferred without asking.

Ask the same question twice, get a different firm

We asked each agent the identical question twice, minutes apart. ChatGPT named around ten different firms across Runs 1–3; two appeared every time. Gemini named about nine across Runs 4–6; two appeared every time. Run 6 drifted into a different county.

The general explanation of rear-end liability stayed steady throughout. What moved was the list of lawyers you’d actually call. Ask Tuesday, get one list; ask Wednesday, most of the list has been swapped — with nothing on screen suggesting the first list was any less confident.

The deadline that was critical, then it wasn’t, then it disappeared

New Jersey has a statutory protection covering certain insurance releases, with a short window to cancel.

Run 1 raised it and called the timing potentially critical — don’t wait. Run 2, on the same facts, raised the same rule and concluded it likely didn’t apply. Run 3 never mentioned it. Same product, same week; the only variable was phrasing.

One version sends you to a lawyer that afternoon. The other doesn’t. Neither tells you the other ever happened.

What it was reading and why

Almost everything sourced was law firm marketing — practice-area pages, case-result pages, the copy firms write to be found. Advertised settlement figures got passed through as reasons to rank one firm above another. Those numbers come from the firm’s list of cases. They aren’t matched to your story, audited, comparable, or checkable by you.

ChatGPT at least showed its sources. Gemini showed none — cards with names and star ratings, nothing to click.

No questions asked

New Jersey is a no-fault state. Drivers elect, on their own policy, between a limited and an unlimited right to sue — a choice made once at signing and rarely thought about again. In a New Jersey auto injury it can determine whether a pain-and-suffering claim exists at all.

Six runs. Two products. A dozen firms. Not one mention of it.

That’s the gap. It isn’t that these tools misrepresent on purpose — in our runs they didn’t. They answer the question you asked instead of the case you’re in, and and without asking questions, they can’t tell the difference — or flag it for you.

“Then just tell it to ask”

There’s an obvious objection here, and it’s a good one. A prompt isn’t only a question — it can carry instructions. If the problem is that these tools don’t run an intake, then tell them to run an intake.

So we did. One run each, the same sentence as before with five words in front of it: Treat this like a legal intake. Nothing else changed. No location typed, same default settings.

ChatGPT complied, in a sense. It produced a section headed Preliminary intake — six bullets, crisply formatted, exactly what you’d hope for. Five of them were our own sentence handed back to us. The sixth read: Jurisdiction: assuming the crash occurred in New Jersey. It then built the statute of limitations, two New Jersey statutes, and every firm recommendation on top of that assumption. Disclosing the guess is to its credit. But the guess is still doing all the work, and one question to us would have settled it. It asked none.

It also went further than any earlier run in the direction we’d already flagged. Its top recommendation arrived under the heading my strongest factual match, and the match was a $375,000 settlement the firm reports on its own case-results page. An advertised number had been promoted to a fact about our case. To its credit, that is at least closer to case relevance.

Gemini didn’t comply at all. It opened on five firm cards with star ratings — no preamble, no sources, no sign it had read the instruction. One firm appeared twice. When we accepted its offer to help organize things, what came back was a checklist: get the police report, get the carrier’s name, get the exact date, get a copy of what you signed. Every item on it was a question it could have asked and instead assigned back to us.

Nine runs now, across two products. Still not one mention of the limited or unlimited right to sue.

What that experiment establishes is more useful than another round of the same complaint. The format of an intake is imitable in five words. The behavior isn’t. Asking before answering is not a tone you can request — it’s a decision about what an agent is permitted to do with a fact it doesn’t have. A tool built to answer on the first turn will produce something intake-shaped when you ask for one, because producing things is what it does. What it will not do is stop and wait, evaluate and turn, because stopping and waiting is the single move it was built not to make from a simple prompt.

And the cost of that is not obvious, which is what makes it worth naming. An assumed fact and an established fact look identical three paragraphs later. Everything that follows after assuming New Jersey inherits the assumption, and nothing on the page marks which parts are load-bearing. Establishing a deadline, a jurisdiction, or a shortlist from a premise nobody confirmed isn’t a good prompt. It’s a different design.

Run 7: By design, the same question, we asked back

We put the same sentence into our own intake. What came back wasn’t a firm list or a settlement range. It was the question that mattered: where did the accident happen, and what state do you live in?

That’s the difference, and it isn’t cleverness. One kind of system is built to answer immediately. The other is built to find out what it doesn’t know yet — because everything downstream is wrong if the first fact is an assumption.

Then you reach out for help

Those runs showed the tools can’t tell what your case is. Here’s the part that should worry you more: the humans on the other end make the same judgment, on the same thin information, in about the same amount of time.

Suppose none of the above deters you. You take a list, pick six firms, and reach out. This is the second thing you control — and the one people underrate, because it doesn’t feel like how you reach out should matter.

Hennessey Digital’s 2025 lead-form response-time study submitted inquiries through the website contact forms of 1,333 U.S. law firms during business hours. 26% never responded at all. Of those that did, the median reply came in 13 minutes — the field splits sharply between firms with a real intake process and firms with none. Worth knowing that the firms it sampled skew large and established, which are the ones most likely to staff an intake desk. Twenty-six percent is a floor, not a midpoint.

Two conclusions, and the first matters more than it looks.

Silence is not an assessment of your case. Every one of those inquiries was clean, ordinary, and scripted, and roughly one in four vanished anyway. If you send five messages and hear back from two, nothing has been decided about your situation. Send more.

And practice fit is doing enormous work. The firms answering in minutes are running triage — someone scanning inbound inquiries and sorting them into “this is ours, this isn’t, this needs a callback today”. So the question that determines what happens to your message is: how long does it take a triage person to figure out what case you have and whether they can or want to handle it?

Most outreach fails that test. Not because the problem is weak, but because a paragraph written at 11pm by someone anxious and exhausted — no dates, no jurisdiction, no sense of what’s already happened — takes work to decode before anyone can answer it. The triage person has forty of those in the inbox on an ordinary day. The one they can read and place in twenty seconds gets the callback.

That’s the part you control.

The case research, and what it can do

Now describe what happened in your own words, as a conversation. A properly scoped agent turns that into a structured case record — not a chat transcript, but an actual reference document meaningful to an attorney: a plain-language summary of your story; key facts separated from impressions; required documents named specifically; the typical stages a matter like yours moves through; critical deadlines that may already be running; your rights in context; the questions worth asking; and the case type and jurisdiction where the incident occurred. The shape of your case.

That structure isn’t decoration. It’s the same set of things an attorney has to establish in the first ten minutes of any consultation before they can discuss anything meaningful with you — and if they’re establishing it live, they’re spending your consultation on data gathering instead of strategy.

It’s a living document, not a one-time export. It revises as your matter moves — when you upload the letter the insurer sent, or the notice that just arrived, the record updates around it, and what matters next moves to the top.

What “thirty to fifty hours” takes from you

We put case research at the equivalent of roughly 30 to 50 hours of manual work, produced in about five minutes. That deserves unpacking, because “saves you time” is what every system claims.

Doing the research yourself, with an agent, one question at a time, that same ground is roughly 100 to 200 research and verification steps. Most are boring, several are confusing, and any one of them can be silently wrong in a way you won’t discover for months waiting between attorney meetings. Done in evenings, after work, across dozens of separate searches, organizing it takes weeks. Most people give up.

Here’s precisely what that time saving buys. Not a faster lawyer lookup. A faster arrival at the point where you can contact one usefully. That matters most for the deadline problem: learning in your first ten minutes that a notice period may be running, and what documentation you need to gather, and what relevant questions to ask, is categorically different from discovering what was needed weeks later. What it doesn’t compress is the attorney’s evaluation and strategy, the court’s calendar, or how long anyone takes to call you back. Preparation was the part that took weeks. That’s the part that now takes minutes.

First contact matters

Here’s where the two halves meet. From that case record, the agent drafts your outreach to a prospective attorney, built to survive the triage you just read about.

Seven things should go in:

The case type and jurisdiction, up front — the two facts that determine whether this is a matter they take at all.

The situation, meaningfully articulated in plain prose, drawn from the structured summary rather than retold from scratch. This is what an agent does best.

The key facts, stated as facts and separated from impressions.

Where things currently stand — what’s already happened, what’s been filed or received, what the other side has done.

Any deadline that appears to be running, flagged as needing verification.

Any notices you’ve received, also flagged as needing verification.

The specific important questions and why, integrated into the prose so they read as a conversation rather than from a form.

It’s written from your perspective, in first person — never as “for my client,” never anything implying a representative is writing on your behalf. And it’s deliberately open-ended. The agent includes enough detail to earn a reply from a busy attorney and not enough to compromise your position. Those can pull against each other, and the resolution is editorial judgment rather than a document dump. Forty pages won’t get reviewed at this stage; and “I was in an accident, can you help” makes no impression at all.

The draft also discloses itself. Every outreach draft closes with a line stating it was drafted with AI assistance from input the sender provided, and that facts should be independently verified. An attorney should know what they’re reading, and a public agent that hides its own involvement in a professional communication is not one you should trust with anything else.

And it’s sent by you, not the agent. The draft opens in your own email client, subject and body fully editable. You choose the recipient, add your name, change whatever you want. Nothing is transmitted to any attorney on your behalf, at any stage.

And follow-up is a necessity

Getting a reply is the first challenge. Staying prioritized over months is the second, and it quietly costs people more when it’s left to someone else’s priorities. Once a matter is ongoing, the same record turns the other direction: it carries a timeline, so a check-in needs to draft against the current stage — here’s where things stand, here’s what I understand I should be focused on, here’s what I’d like to cover. That’s a materially different message from “any update?”

These messages are also drafted on demand, not on a schedule. No automated sequence, no drip campaign, nothing that contacts an attorney without you deciding to send it. Software that automatically generates recurring pressure on lawyers would be a bad thing to build, so we don’t do that.

What this feature does not do, and it would be easy to imply otherwise: it does not make a firm respond. Nothing overrides an understaffed intake desk, and in full transparency, we have no data showing better-structured inquiries get answered more often — nobody has run that study. The argument is mechanical, not empirical. Firms triage; triage rewards clarity, strong relevance, and urgency; a structured inquiry supplies all three where a hastily written paragraph supplies none. That’s the claim, and we’ll leave it there.

The prompt exposes your case publicly

Read the query at the top of this post the way a stranger would: a commercial-vehicle collision, an insurer that has already called twice, four months of physical therapy, and a document signed last week that the writer suspects was a mistake.

That is not a question. It’s a statement about an active dispute, containing something that functions as an admission, volunteered in writing by the person standing to lose from it. If opposing counsel could commission one document from you, it would look a great deal like that sentence.

Three things are true about typing it into a public assistant. It isn’t privileged — courts have treated user-chatbot exchanges as ordinary third-party communications, discoverable like email or Slack. Deleting it may not be your decision — in the New York Times’ litigation against OpenAI, a May 2025 preservation order required the company to preserve output log data that would otherwise have been deleted, and a litigation hold outranks any privacy policy. And “de-identified” is weaker than it sounds — a collision on a known date with a specific treatment duration and insurer contact history are signals that can be correlated.

The difference isn’t that CC My Attorney is careful and public AI is careless. It’s architectural. A general assistant has no separation between what you tell it and what it searches with, because it was never built to have one: your paragraph is the prompt, the query, and the log entry at once. You are now the payload. A CC My Attorney workflow splits those roles — the conversation and your uploaded documents go to Anthropic, which handles intake and analysis under terms that do not permit training on our inputs, while attorney-matching and comparable-case research run at OpenAI as a derived payload, not your entire prompt, consisting of case type, your city and state, summary, and key facts, assembled after a PII prescreening. The research half of the workflow does not know your story. It only knows the abstracted shape of the problem.

And on the conversation side, Anthropic operates under a zero-data-retention arrangement: it does not retain inputs or outputs after processing, except where retention is required by law or where content is flagged by its automated safety systems, in which case it may be kept for a limited period. OpenAI, which never receives the conversation, may retain the derived research fields for a short period. It’s a contractual arrangement rather than a prediction about anyone’s future legal obligations, but the practical consequence is straightforward: data that was never retained isn’t sitting anywhere waiting to become relevant to the opposing side of your case. And the split holds independently of it — the half of the system doing the searching never received your story to begin with.

What free tools are genuinely good for

Being fair, because overstating this would be disingenuous. Public assistants are good at explaining what a term means, giving a rough orientation to an unfamiliar area, and helping you draft something. Run 2 was a better answer than most people would write for themselves.

What they cannot do is hold your matter across sessions. They don’t know your jurisdiction unless you volunteer it, won’t compute representative deadlines or stages, won’t read what you signed or hold a place for it later, won’t verify the firm they named has handled anything close to your situation, and can’t tell you what happens to what you typed.

The point

Go back to the original prompt. Rear-end collision with a commercial vehicle. An insurer that has already called twice. Four months of physical therapy. A document signed last week.

Every run in this post gave that prompt an answer. Not one of them asked whether they had elected the limited or the unlimited right to sue — the choice that can decide whether there’s a bodily-injury claim at all.

That’s the whole problem in one line. The answers weren’t wrong. The question that decides the case was never discovered first, and you have no way to know which question that is. You can’t fix that by scrolling. You fix it by changing what you bring to an agent.

Start from your story, not your zip code. What happened, when, to whom, and what has happened since. That’s the input every step downstream depends on.

Treat geography as a filter, not a ranking. Licensure follows the matter — usually where it happened or where the other side can be sued — not where you’re holding your phone.

Check the license. Every state bar publishes license status and discipline history. It’s free, it takes two minutes a name, and almost nobody does it.

Open the source. If an AI names a firm or asserts a fact about your situation, click through and confirm it says what the answer claims. All six of our runs would have failed that check.

Send more messages. Roughly a quarter go unanswered. That’s staffing, not a verdict on your case.

Arrive organized. So the first meeting is spent on your problem instead of on transcribing it.

None of that requires our agent. It requires knowing three things: preparation is real work, it’s yours to do, and it’s the one part not to hand to a public prompt.

Frankly, “near me” will still be the first thing you type. It’s the right question for finding a restaurant. The lawyer you need is the one who has handled what actually happened to you — and the only person who can put that in front of them is you. There’s help to get you there faster and more reliably.

How we tested is available upon request.

Try it for free: ccmyattorney.ai