The referral moved to the assistant.

High-trust buyers used to start by asking someone they knew. Increasingly they start by asking a machine — and the machine has already built a shortlist before your phone rings.

For most of the history of professional services, the opening move in a serious engagement was a conversation with a person. A general counsel called a peer. A family asked their accountant who handled estates. The referral was not a marketing channel; it was the mechanism by which trust transferred from someone who had already taken the risk to someone about to take it.

That conversation still happens. What has changed is what happens before it. A meaningful share of buyers now open an assistant first, describe their situation in their own words, and receive back an orientation to the problem — sometimes with names attached. By the time they call the peer, they are not asking who should I talk to. They are asking have you heard of these three.

This is a smaller change than the hype suggests and a larger one than most firms have priced in. The referral has not been replaced. It has been moved downstream, and something else now occupies the position it used to hold.

The first conversation is now with a machine

The clearest evidence sits on the B2B side. In October 2025, Responsive surveyed 350 B2B buyers and found that generative AI had overtaken traditional search for a quarter of them, with nearly two-thirds using generative AI as much as or more than search when researching vendors. Among U.S. buyers, 48% said they use generative AI for vendor discovery, against 14% in other regions. At companies with more than 2,000 employees, 42% used it for discovery, versus 18% at firms of 501 to 1,000 people.[]

Consumer-facing professional services show the same shape. Clio’s 2025 Legal Trends Report, published in October 2025, found that more than half of consumers have used or would consider using AI to answer a legal question, and that a growing majority say they would look for their next lawyer online.[] The underlying split is instructive: 14% of consumers had already used AI for a legal question and 43% had not but would consider it, with adoption running higher among younger cohorts — 26% of Millennials and 23% of Gen Z.[] In health-adjacent categories, KFF’s tracking poll of 1,343 U.S. adults, fielded in late February and early March 2026, found that 32% had turned to AI chatbots for health information in the previous year.[]

The distribution is not marginal. OpenAI reported 900 million weekly active users for ChatGPT in February 2026, up from 800 million four months earlier.[]

One number in the Clio data matters more than the adoption figures. Of the consumers who used AI for a legal question, 28% were directed to contact a lawyer.[] The assistant is not primarily substituting for the professional. It is functioning as triage — absorbing the low-stakes question, and, when the matter is serious enough, routing the person toward help. That routing decision is the new top of your funnel. It is happening whether or not your firm participates in it.

Your category is treated differently, and that is deliberate

If you run a law firm, an RIA, an accounting practice, or a clinical group, you are not competing in the same information environment as a software vendor. Google formalized the distinction years ago and still applies it. Its Search Quality Rater Guidelines, in the version dated September 11, 2025, define “Your Money or Your Life” topics as those where content “could significantly impact the health, financial stability, or safety of people, or the welfare or well-being of society,” and state that for pages on clear YMYL topics, “we have very high Page Quality rating standards.”[]

Google’s public guidance to publishers puts the same principle in plainer terms. It describes the E-E-A-T framework — experience, expertise, authoritativeness, trustworthiness — and is explicit that “of these aspects, trust is most important. The others contribute to trust.” It says the company gives “even more weight to content that aligns with strong E-E-A-T” on YMYL topics. The rater guidelines define the components plainly: expertise is whether the content creator has the necessary knowledge or skill for the topic; experience, whether they have first-hand experience of it; authoritativeness, whether the creator or site is known as a go-to source.[][]

The consequence shows up in behavior that varies sharply by sector. A June 2026 Search Engine Land analysis drawing on BrightEdge and Semrush tracking found that AI Overviews appeared for 93% of health condition and symptom queries as of December 2025, and that nearly 80% of the sources cited in healthcare AI Overviews were also ranking organically — a conservative posture that anchors answers to already-established sources. Finance behaved differently: 65.7% of citations came from sources outside the organic top 100. In law and government, only 12.42% of tracked keywords triggered an AI Overview as of November 2025, a figure that had declined by 4.78 points since March.[]

Read that as three different postures toward liability. In health, the system defaults to institutional consensus. In finance, it ranges much further afield to find explanatory material. In legal, it frequently declines to generate an answer at all. Each posture creates a different opening — and none of them rewards the tactics that work in unregulated categories.

What the assistant is actually reading

Firm principals tend to assume that an assistant reads their website. It does, sometimes, and that is a small fraction of the input.

Semrush’s 2026 AI Visibility Index analyzed 126 million U.S. AI search prompts from January through April 2026 across ChatGPT, Gemini, Google AI Mode, and AI Overviews. ChatGPT cited an average of 15 sources per response; Gemini averaged 3. The study also found that the overlap between the brands an engine mentions and the domains it cites can be as low as 30% on Gemini — meaning a firm can be recommended in an answer without its own site being the evidence for the recommendation. In finance, the top three brands accounted for 41.4% of total category visibility.[]

Conductor’s tracking of seven engines from September 2025 through March 2026 found Wikipedia, Reddit, and YouTube anchoring the top citation positions across most query intents, with the notable exception of Claude, which over its available data never surfaced any of the three, drawing instead on brand domains and institutional sources.[]

The translation is simple. When an assistant answers a question about your practice area, it assembles a picture from many places, most of which you do not own. Your site supplies identity, credentials, and specificity. Third parties supply corroboration. Where corroboration is thin, the model falls back on whoever the category’s incumbents are — and in finance, per the concentration figure above, the incumbents are very few.

Two things follow. First, Google’s own guidance tells publishers to make it “self-evident to your visitors who authored your content,” and to add “accurate authorship information, such as bylines to content where readers might expect it.”[] On a YMYL page, readers expect it. So does the evaluation apparatus. Anonymous firm-voice content is a missing signal, not a neutral one.

Second, borrowed authority is now explicitly policed. Google’s spam policies define site reputation abuse as “a tactic where third-party content is published on a host site mainly because of that host’s already-established ranking signals, which it has earned primarily from its first-party content,” and define scaled content abuse as generating many pages “for the primary purpose of manipulating search rankings and not helping users.”[] The placement-farm approach to authority — buying your way onto a reputable domain’s subfolder, or spinning up volume — is a named failure mode rather than a clever shortcut.

The constraint most agencies pretend away

Nothing in this article is legal, regulatory, or compliance advice. Rules vary by jurisdiction, by regulator, and by the specific facts of your practice. Anything below describes the general shape of published requirements and should be checked with your own counsel or compliance function before it informs anything you publish.

Regulated firms cannot say what unregulated firms say, and the gap is structural rather than stylistic.

The ABA’s Model Rule 7.1 states that “a lawyer shall not make a false or misleading communication about the lawyer or the lawyer’s services,” and treats a communication as false or misleading if it contains a material misstatement of fact or law, or omits a fact necessary to make the statement considered as a whole not materially misleading.[] Model rules are adopted, and varied, state by state.

On the investment side, the SEC’s marketing rule for investment advisers — Rule 206(4)-1, effective May 4, 2021 with a compliance date of November 4, 2022 — prohibits advertisements containing untrue statements of material fact or omissions that make a statement misleading, prohibits unsubstantiated material statements of fact, requires fair and balanced treatment of material risks alongside claimed benefits, and imposes specific conditions on testimonials, endorsements, third-party ratings, and performance presentation.[]

FINRA Rule 2210 sets content standards for member communications requiring that they be based on principles of fair dealing and good faith, be fair and balanced, and contain no “false, exaggerated, unwarranted, promissory or misleading statement or claim.” Testimonials discussing investment advice or performance carry prescribed disclosures, and retail communications generally require approval by an appropriately qualified registered principal before use.[]

Set the three side by side and the same instruction appears in each: no superlatives you cannot substantiate, no outcome you cannot support, no selective presentation, no omission that changes the meaning of what remains.

Most marketing advice fails immediately against that standard, because most marketing advice is built on superlatives, outcome claims, and curated proof. What survives is narrower and, conveniently, better suited to the machine: verifiable specificity. Which matters, in which venues, under which body of law, decided when, by whom, with what disclosed. That is a compliance-safe sentence, and it is also the highest-signal sentence you can give a retrieval system.

The assistants are not reliable narrators

Everything above assumes the assistant reads you accurately. It often does not.

The Tow Center for Digital Journalism tested eight generative search tools across 1,600 queries in research published by the Columbia Journalism Review in March 2025, asking each to identify the source of a supplied article excerpt. The tools answered incorrectly on more than 60% of queries. ChatGPT misidentified 134 articles while signaling uncertainty only 15 times across 200 responses, and never declined to answer. More than half of the responses from Gemini and Grok-3 cited fabricated or broken URLs.[]

The pattern holds in tools purpose-built for a regulated field. Stanford RegLab researchers ran the first preregistered evaluation of AI legal research products and found that the leading commercial tools hallucinated between 17% and 33% of the time — better than general-purpose models, and far from the marketing claims made for them.[]

Buyers half-know this. In KFF’s 2024 polling of 2,428 U.S. adults, 56% of AI users said they were not confident they could tell what was true from what was false in a chatbot’s answers.[]

So the risk to a firm is not only omission. It is confident misdescription: the wrong jurisdiction, a lapsed practice area, a partner who left three years ago, an outdated fee posture, all delivered fluently and without hedging. The only durable defense is unambiguous, current, machine-legible fact about who you are, said in enough places that no single stale source can dominate the reconstruction.

What actually holds up

The signals that survive all of the above are the ones the profession already respects.

Named people, not a firm voice. Real authorship with credentials, jurisdictions, admissions, and the disclosures your regulator expects, attached to the person and consistent across your site, your bios, your bar or licensing listings, and your professional profiles. If the assistant cannot resolve who wrote something and what standing they have, the content contributes nothing to the trust signal Google itself calls the most important.[]

A real editorial presence, published continuously. Hinge Research Institute’s 2025 High Growth Study, covering 770 professional services firms, found that the highest-growth firms ranked producing high-value educational content as their top marketing priority, alongside developing individual thought leaders through speaking, writing, and podcast placement. Those firms spent roughly twice as much on marketing as their lower-growth peers and grew more than four times as fast — an association, not a proven cause, but a consistent one.[]

Independent corroboration. Because a large share of what an engine assembles is third-party, presence outside your own domain is not vanity. It is the evidence layer. Regulators generally permit factual, substantiated, properly disclosed material about your firm to appear elsewhere; what they constrain is what that material may claim and how testimonials, ratings, and performance may be handled.[][][] Build the corroboration inside the constraint rather than treating the constraint as a reason not to build it.

Structured, current firm data. Practice areas, locations, admissions, languages, engagement models, and the mechanics of contact — expressed cleanly enough that a retrieval system can lift them without inference. Most of the misdescription risk lives in inference.

An assumption that the answer will be assembled without a click. Write so that the first paragraph of any page states the substantive answer, with the qualifying detail immediately after. The reader you are optimizing for may never load the page.

The stance

Here is the position we will defend.

The referral was never really about the phone call. It was about a trusted third party vouching for expertise that the buyer could not evaluate on their own. That function is now partly performed by software that reads what has been published about you and reconstructs a recommendation from it — imperfectly, confidently, and increasingly early in the process.

Firms that treat this as a search problem will buy visibility tactics and be disappointed, because the tactics that move unregulated categories are the tactics your regulator has already prohibited. Firms that treat it as a publishing and evidence problem — named experts, substantiated specificity, independent corroboration, current structured fact — end up satisfying the compliance requirement and the machine requirement with the same work.

The referral moved. The thing it was carrying did not. Expertise still has to be demonstrable to someone who does not already know you. The only change is that the someone is now, very often, not a person.