AI assistants recommend professional services firms by matching a client’s question to what they can read about each firm: its own website, directories and rankings, reviews and regulator records. Each assistant reads different sources, so each names different firms. The firms that appear most are the ones described most specifically, which is not always the largest.

This guide brings together six 2026 studies of how ChatGPT, Gemini, Claude, Perplexity and Google’s AI recommend accountants and lawyers. It is written for partners, practice managers and marketing leads who want to know what the evidence says before spending money.

Do clients really ask AI who to hire?

Yes, and the numbers moved fast in 2026.

Finding Study
41.9% of US consumers would use ChatGPT to research a lawyer, up from 28.1% in 2025 and 9% in 2023 iLawyer Marketing, 1,110 consumers
50.1% would use at least one AI answer engine. 9.5% would use AI and nothing else iLawyer Marketing
Google’s share for lawyer research fell from 86.7% to 71.9% in a year iLawyer Marketing
45% of US consumers used AI tools for local business recommendations in the past year, up from 6% BrightLocal, 1,002 consumers
70% of UK SMEs always or often act on AI advice about tax, finance or business without checking with their accountant Ravical, 500 SMEs

Two cautions. These are surveys, and some were run by marketing agencies, so treat them as direction more than precision. And most people still verify. In the iLawyer study, more than 70% of those who would use ChatGPT said they would also check Google. AI is an extra step in the decision, often the first one, where the shortlist is formed.

How does an assistant build its shortlist?

The assistant reads the request for three things: the problem, the place and the kind of client. “Accountant for a limited company contractor near Reading” has all three.

With web search on, it then runs several searches of its own. Google calls this query fan-out: one question becomes a set of related searches across subtopics. The assistant reads the pages it finds and writes an answer from them. Without search, it answers from what the model remembers, which favours firms that were widely written about before its training ended.

A firm gets named when the pages the assistant reads connect it to that problem, place and client. That is the whole mechanism. There is no submission form and no ranking to buy.

Which sources does AI use?

It depends on the profession and the assistant. Here is what the studies measured.

For lawyers in the US, 5W asked 24 questions three times each on ChatGPT, Claude, Gemini and Perplexity in October 2026, for 288 answers:

  • State bar associations were named in 57.3% of answers.
  • Martindale-Hubbell was named in 47.6% and Avvo in 46.2%.
  • FindLaw, Super Lawyers and Justia each appeared in more than a third.
  • Directories and rankings supplied 30.5% of all citations. Firm and brand websites supplied 27.4%. Review and comparison publishers supplied 11.4%.

For solicitors in the UK, Whito found assistants citing Chambers, The Legal 500, ReviewSolicitors and Google reviews. Where review data was quoted, the median firm had 113 reviews.

For accountants in the UK, the same researchers found ChatGPT citing ThreeBestRated in five of six cities, and Perplexity leaning on Clutch and local directory roundups. Every firm whose reviews were quoted was rated 4.5 or higher, with a median of 37 reviews.

For CPA firms in the US, AIWorthy’s study of audit work found the opposite emphasis: the firms’ own websites were the most cited source. The heavily cited sites named their region, described the buyer’s kind of work in depth and had been updated recently. Reddit, Quora and Wikipedia were largely ignored. Social media mostly meant LinkedIn.

The studies disagree on how much your own website matters, and that is the useful finding. The mix changes with the question. A broad “best accountant in Leeds” pulls in directories and reviews. A specific “auditor for a charter school in the Triangle” pulls in the pages that describe exactly that work.

Why do the assistants name different firms?

Because they read different sources and weigh them differently.

In Whito’s test, three assistants named 109 different accountancy firms across six UK cities. 93 of them, 85%, were named by only one assistant. One firm was named by all three. For law firms the overlap was a little higher: 74% named by one assistant, and three firms named by all three.

5W saw the same thing in its citation data. Gemini named Martindale-Hubbell in 62.5% of its answers. Perplexity named it in 20.8%.

For a firm, this has two consequences. A single search on one assistant tells you almost nothing. And there is more than one way in: a firm missing from ChatGPT may already be named by Gemini, for reasons you can find and repeat.

Only for broad questions.

Expert Press ran 100 buyer questions about tax, audit and advisory services through seven assistants. 25 firms took 87.2% of all mentions, and visibility fell steeply with revenue. That is the national, no-location picture.

Add a place or a niche and it changes. No Big Four firm was recommended in Whito’s six-city accountant test. No Magic Circle firm was recommended in its solicitor test. The one accountancy firm named by all three assistants was incorporated in 2022.

AIWorthy’s result is the sharpest. It took ten firms that actually perform single audits in North Carolina’s Research Triangle, according to federal records, and asked five assistants who to hire. The four busiest firms were recommended 24 times. The six least busy were recommended 184 times. The busiest firm’s website had 1,821 pages and never said where the firm operates.

Experience that AI cannot read does not count. A smaller firm that states its work plainly is the easier match.

What makes a firm easy for AI to recommend?

Across the studies, five things recur.

  1. One page per service, written for one kind of client. “Accountants for dental practices in Leeds” can be matched. “Our services” cannot.
  2. Location in words. Name the city, the area and the nearby towns you serve on the pages that matter, not only in the footer.
  3. Facts a client would ask for. Fees or a fee basis, who does the work, their qualifications, what happens first and how long it takes.
  4. The same details everywhere. Firm name, address, phone number and services should match across your website, Google Business Profile, directories and regulator records.
  5. Recent, honest reviews. The firms whose reviews AI quoted were rated 4.5 or higher. Ask every satisfied client, without incentives.

None of this is a trick. It is the information a careful client wants, written so that a machine can find it too. For regulated firms it also has to be true. Our guides for accounting firms and law firms cover the professional rules.

What should a firm do first?

Run a baseline before changing anything.

  1. Write ten questions a new client would ask, each with a problem, a place and a type of client. Leave your firm’s name out.
  2. Ask each one on ChatGPT, Gemini and Google’s AI Mode, with search on.
  3. Record which firms are named, in what order, and which sources are cited.
  4. Find the firm named most often and read the pages cited for it. That is your gap, in plain view.
  5. Fix one page, then ask the same questions again the following week.

The checklist for professional services firms turns this into a half-hour exercise.

Akkrev does this for you every week. It writes your client questions, checks them across four AI model families, shows the firms and sources behind each answer, audits your website and prepares the fixes and articles that close the gaps. You approve what gets published. No one can guarantee a recommendation, but you can see exactly where you stand and what to do next.

Questions

Does AI prefer large firms?

For broad national questions, yes. In Expert Press's 2026 study, 25 firms took 87.2% of mentions across 3,480 answers. For local and niche questions the pattern reverses. No Big Four firm and no Magic Circle firm was recommended in Whito's UK city tests, and in AIWorthy's audit study the least busy firms were named far more often than the busiest. What they had in common was a specific, readable description of their work.

Do I need special markup or an llms.txt file to be recommended?

No. Google says its AI features need no special markup or AI text files. A page must be indexed and eligible for a snippet. Accurate structured data that matches your visible content is still good practice, and it helps any system read your firm's name, address and services correctly.

How often should a firm check its AI visibility?

Weekly is a sensible rhythm. Answers vary by day, by assistant and by wording, so a single check can mislead in either direction. Use the same questions and settings each time so that one week can be compared with the next.

Can a firm pay to appear in AI answers?

Adverts exist. ChatGPT's free tier showed a labelled advert under some answers in Whito's UK test. They are separate from the recommendations, and the researchers saw no evidence that adverts changed which firms were named. The unpaid answer is earned through what AI can read about you.

Sources

Related