Did AI send you that case? The intake question most firms get wrong

A partner asks a fair question: are we actually getting cases from ChatGPT, or is this whole thing noise? Most firms cannot tell, because the only instrument pointed at it is a free-text box on the intake form that says "How did you hear about us?" — and whoever answers the phone types "internet."

How does a law firm find out whether a client came from ChatGPT? Ask at intake, and check analytics, and accept that neither one sees the whole picture. At intake, replace the open-text "How did you hear about us?" with a named option for AI assistants, and record the answer in one structured field rather than in a notes column. In analytics, Google Analytics 4 now has a built-in AI Assistants channel that separates visits arriving from assistants such as ChatGPT and Gemini, and Google Search Console has a separate report showing how often your pages appeared inside Google's AI Overviews and AI Mode. Neither method captures an AI answer that named your firm and produced no click. The two measurements are partial, they do not reconcile with each other, and a firm that wants a real answer runs both and reads them as a pair.

The intake attribution card {#intake-attribution-card}

Print this. Tape it where intake calls get answered. Nothing in it is specific to any vendor. Or use the free AI Referral Intake Kit to copy the script, download the CRM field definitions, and start a blank monthly worksheet.

---

THE QUESTION — asked after the matter details, never first:

"One last thing so we know what's working — before you called us, did you use an AI assistant like ChatGPT or Gemini, or did you find us another way?"

THE FOLLOW-UP — only if they say yes:

"Do you remember whether it gave you our name directly, or whether you searched for us afterwards?"

THE FIELD — one required dropdown in the CRM or intake form, not free text:

Keep the free-text box as a second, optional field. Verbatims are useful. They are not data.

THE CALL-TRACKING NOTE. A client handed your name by an assistant often dials rather than clicks, so the dropdown must be filled for phone intakes too, by the person on the call. A tracking number tells you the ad channel. It cannot tell you what the client asked ChatGPT.

THE MONTHLY RECONCILIATION — fifteen minutes, one meeting:

1. Count intakes tagged to each dropdown value. Note how many are "other / unknown"; that number is your error bar.
2. Pull the AI-assistant sessions from analytics for the same month.
3. Compare the two counts. They will not match. Write down the gap.
4. Note your branded-search volume for the month — people searching your firm's name. Rising branded search alongside flat advertising is the fingerprint of recommendations you cannot see directly.
5. Record all four numbers in the same place every month. The trend is the finding. Any single month is an anecdote.

This card may be reproduced and used freely with attribution to askbriefly.ai.

---

What analytics can now see — and what it explicitly excludes

Google Analytics 4 added a built-in AI Assistants channel. The release note is dated May 13, 2026 (Analytics Help). Google's reference documentation defines the channel as traffic arriving "from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok" and states plainly that "it excludes Google's AI Overviews and AI Mode" (Analytics Help, Default channel group). The same page gives the rule: the medium is set to ai-assistant "if the referrer matches a list of AI Assistants."

Three things follow from that definition, and together they are the reason not to take the channel at face value.

The list is short, and Google's own pages do not agree on it: the reference page names five assistants, while the May release note gives its examples as "chatbots like ChatGPT, Gemini, and Claude." Perplexity appears on neither, and the rule points only to "a list of AI Assistants" that the page never publishes or names — unlike the source category it cites for organic search. Treat the channel as a partial count.

Google AI Overviews and AI Mode are excluded by design, in Google's own words. Someone who reads an AI Overview naming your firm and clicks through arrives with a Google referrer and lands in Organic Search, indistinguishable from someone who scrolled past the Overview to the blue links. No setting separates them.

And classification depends on a referrer arriving at all. The rule sets the medium "if the referrer matches" — no referrer, no match, and the session falls into Direct. That failure is silent.

Google Search Console added a generative AI performance report. Announced on the Search Central blog on June 3, 2026 and rolled out "to a subset of websites" first (Google Search Central), it covers AI Overviews and AI Mode, broken out by pages, countries, devices, and dates (Search Console Help).

Read the metric list before you get excited: impressions only. Google defines them as "how many times links to your site were shown to a user in a generative AI feature on Google Search." No clicks, no click-through rate, no queries. Clicks do exist elsewhere — Google's methodology page states that "clicking a link to an external page in the AI Overview counts as a click," and says the same for AI Mode (Search Console Help) — but those clicks land in the main Search performance report, mixed in with everything else, with no AI filter.

The ceiling nobody sells you

A recommendation is not a link. When an assistant tells someone that Marchetti & Doyle Family Law handles contested custody in their county, the useful thing has already happened inside the conversation. The person may click. They may also write the name down, search it later, ask a friend, or dial the number the assistant read out. Each of those paths arrives wearing a different costume — branded organic, direct, phone — and none carries a label saying an AI assistant caused this.

That is why intake and analytics never reconcile. They are partial measurements of different populations: intake sees people who converted and remembered, analytics sees sessions that clicked and kept a referrer. The overlap is small, and the union is not the truth either.

Anyone selling you a closed loop from AI answer to signed case is selling something that does not exist today. What does exist is a pair of half-measurements whose trends are informative even though their levels are not.

A worked reconciliation

The firm and the numbers below are fictional. They show the shape of the exercise, not a benchmark. Do not quote them as data.

Marchetti & Doyle Family Law (marchettidoyle.example) runs the card for one month, on 61 new intakes.

Seven intakes credit an assistant. Analytics accounts for two. The other five never appeared as an AI-referred session: they searched the firm's name, or they called. Meanwhile nine intakes are "unknown" — a bigger number than the entire AI count, and the correct thing to be nervous about.

What the downstream half cannot answer

Intake tells you what already happened, to people who already called. It says nothing about the answers being given right now to people who did not call — the ones where a competitor got named and you did not. That is the upstream half, and it means asking the assistants directly on a schedule rather than waiting for survivors: the method is in how to track when ChatGPT recommends your law firm, and the metric it produces is Case Recommendation Share.

The honest limits

• Assistant answers are probabilistic. The same question asked twice does not reliably return the same firms. SparkToro's January 2026 testing — 2,961 prompt runs across ChatGPT, Claude and Google AI Overviews — found brand and product recommendations highly inconsistent between runs, and cautioned marketers about exactly this. That testing was on consumer product categories, not legal questions. One check is a snapshot of one moment.
• A model update can reshuffle everything. Assistants change retrieval and ranking behavior without notice. A month-over-month move may be a change in your footprint or a change in the model, and from outside those look identical. Hence long, boring, consistently-collected series.

The number this article does not give you

You will not find "X% of law firm cases now come from AI assistants" here, or a signed-case figure, or a dollar value per AI referral. Nobody has published credible signed-case data from AI referrals for the legal industry — not the agencies, not the platforms, not us. The adjacent numbers that do exist are worth having with their labels intact: consumer intent is stated preference from a survey that asked 1,110 US adults which sources they would use, cross-industry LLM referral growth is observed across 166 analytics properties but not legal-specific, and the widely-quoted conversion advantage is a single unnamed client in one 2025 case study, where AI was 0.07% of that client's organic traffic. All three, caveats included, live on our statistics hub.

Estimating the missing number would be easy, and it would be fiction. The card above is the protocol that generates a real one, for your firm, in your market, over six months. Slower than a statistic, and worth more.

FAQ

What exactly should we ask clients to find out if AI sent them?

Late in the intake call, after the matter details: "Before you called us, did you use an AI assistant like ChatGPT or Gemini, or did you find us another way?" Naming two assistants makes it answerable — clients have no word for the category. Record the answer in a required dropdown, not a free-text field.

Can Google Analytics tell me how many clients came from ChatGPT?

It tells you how many sessions arrived from a list of assistants Google recognizes, via the AI Assistants channel added in May 2026. Not clients. It also excludes AI Overviews and AI Mode by design, misses assistants absent from Google's list, and cannot classify a visit that arrives without a referrer.

Does Search Console show clicks from AI Overviews?

Not as a separate figure. The generative AI performance report shows impressions only — no clicks, no CTR, no queries. Clicks on AI Overview and AI Mode links are counted in the main Search performance report, mixed together with all other Google Search clicks, with no way to isolate them.