You win work by referral, but ChatGPT never names your firm. Does it matter, and what should you fix first?

Does it cost a referral-led firm clients when ChatGPT never names it?

Yes, some, because buyers now ask an AI assistant for firms like yours and check the names they were given. The loss is smaller than most agency pitches claim and larger than zero. It shows up in two places: shortlists you never heard about, and referrals that cooled after a buyer checked them.

The commercial share of AI conversations is now large enough to matter. Profound, which sells AI visibility tools, analysed 7.5 million en-US ChatGPT conversations between 23 June 2025 and 22 June 2026 and found that the share with commercial intent rose from 13.9% to 19.2%.¹ In professional services the share was 18.9%, close to legal services at 19.3% and B2B software at 21.8%.¹

Roughly one conversation in five in your category is a buyer deciding something, in a channel where your firm has no presence.¹ One disclosure before we go on: Sigzen AI sells AI visibility audits and programmes at published prices, so we have an interest in the answer.² We have tried to show where the honest answer is "not much".

How often do AI answers name or link firms like yours?

AI answers name firms far more often than they link to them, and in professional services links are rare. Similarweb, which sells web and AI traffic data, found that 6.8% of US ChatGPT answers carried a citation in May 2026.³ For professional services the rate was under 4%.³

That changes what you should measure. A consulting firm will see little AI referral traffic even when it is doing everything right, because the engine rarely links anyone in the category. The question is whether the engine names you, describes you accurately, and places you beside the right competitors. Our note on Similarweb's citation rates by sector explains why the right KPI differs from one category to the next.

Our view: a referral-led firm should stop looking for AI traffic in its analytics and start looking for its own name in the answers. Being absent from a shortlist leaves no trace in any dashboard, which is why the problem goes unnoticed for so long.

What do clients expect, and where do they ask?

Clients expect AI-enabled work from their advisers, and they ask outside tools rather than the firm's own. The Thomson Reuters Institute surveyed 1,816 professionals in 62 countries for its Future of Professionals 2026 report (March to April 2026). In it, 78% of corporate clients called AI-enabled quality improvements very important or essential, and 6% said most or all of their providers deliver them.⁴

Where they ask is the part that touches visibility. Gartner surveyed 3,566 B2B and B2C customers between February and March 2026 (geography not stated). People were about three times as likely to use a third-party GenAI tool such as ChatGPT as a company's own chatbot.⁵ Your website assistant, if you have one, is not where the question gets asked.

What B2B buyers report

B2B buyers say the same about research. In G2's survey of 1,076 B2B decision makers in North America, EMEA and APAC (March 2026), 51% said they start software research in an AI chatbot more often than in Google.⁶ In the same survey, 69% said a chatbot had surfaced information that led them to a different vendor than expected.⁷ Software buyers are not consulting buyers, but they are often the same people. What a firm should publish about its AI-enabled work is a separate question, covered in an earlier post.

The referral now gets checked before the call

A referral used to arrive as a name and a phone number. It now arrives with a question typed into an assistant on the way to the call. That is our reading, and the buyer surveys support it from three sides.

Buyers check what AI tells them. In TrustRadius's survey of 1,862 technology buyers in January 2026, 63% used AI somewhere in the purchase journey, 94% of those fact-checked what it told them, and 83% shortlisted three or fewer products.⁸ They also meet errors. In G2's March 2026 survey of 1,076 B2B decision makers, 64% said they encounter inaccurate AI recommendations often or very often.⁹

Trust in the human adviser stays high, which is why the referral survives the check. Gallup and Edward Jones surveyed 2,117 Canadian adults between 25 March and 3 April 2026. Of them, 76% had at least some confidence in the expertise of financial advisers, while 4% had a great deal of confidence in AI tools for financial guidance.¹⁰ The referral is being verified, not replaced, and a verification that returns nothing is a quiet loss.

How an engine builds its candidate list

An engine answers "which firms do X for Y in Canada" by retrieving pages, matching the entities it finds, and keeping the ones that more than one source corroborates. The sequence below is simplified, but it shows how a well-regarded firm goes missing.

Entity

A thing the engine can identify and keep facts about: your firm, its leaders, its services, its registrations. An entity exists for the engine when the same facts appear consistently in more than one place it reads.

Corroboration

Agreement between independent sources about an entity. A professional-body listing, a partner directory and a press mention that all describe the same firm the same way corroborate each other; a lone website does not.

Where the engine looks matters. In Profound's study of 11.84 billion citations across eight engines (16 April to 16 July 2026), about 57% went to company-owned sites, with ChatGPT at 47% and earned media taking 30% of its citations.¹¹ Semrush, which also sells AI visibility tools, found ChatGPT cited about 15 sources per response against about 3 for Gemini (US, January to April 2026).¹²

Read those two together. ChatGPT reads widely and mixes in press and directories, so a firm that exists on its own site alone is out-corroborated by a firm that exists in five places. Gemini reads few pages and favours the brand's own, so a site without a plain service page gives it nothing to match. Our AI visibility work starts at this stage: what the engines retrieve for your prompts, and whether your firm can be matched in it.

Why is a referral-led firm invisible?

Because a referral-led firm has spent twenty years being known by people and almost no time being known by machines. Four gaps follow, and they usually arrive together.

• Thin entity facts. No single page states the firm's legal name, founding year, leaders, services, sectors, countries served and registrations, so the engine has little to match.
• No third-party corroboration. The firm is missing or outdated on its professional body's directory, absent from the partner directories of the software it implements, and never quoted in the trade press.
• No service pages written as answers. A page headed "Advisory" does not answer "which firms handle transfer pricing for manufacturers", so the engine matches a competitor's page that does.
• Category-shaped prompts. A buyer asks for "a firm that does X for Y companies in Canada", so the engine builds a named candidate list from directories and lists. A firm on no list never enters.

Our view: the fourth gap is the one most firms miss. The buyer's prompt describes a category and the engine answers with a list, so a firm that has never been listed anywhere cannot be on it.

Where the corroboration comes from

The second gap has its own evidence, and most of it points to a short list of sources. For an advisory firm those are the professional body's member directory and the partner directories of the accounting and ERP software it implements. Add the business-profile and review sites its clients already use, LinkedIn, and the local and trade press. A listing on each that states the same name, services and sectors as the entity page is what the engine reads as agreement.

Two earlier posts cover the sources in more depth: which review sites and directories the engines actually cite, and what a Canadian accounting firm should publish when clients bring AI answers to meetings. Neither replaces the check in month three, which shows the sources your own category draws on.

What should you fix first?

Fix the facts an engine needs before writing anything new. That means entity facts and professional-body listings in week one, three answer-shaped service pages in month one, a twenty-prompt check from month three, and content after that. This is the order the professional-services playbook uses, because each step feeds the next.

Week one: entity facts and body listings

Write one entity page that states the firm's legal name, founding year, leadership, services, sectors, countries served and registrations, and mark it up with Organization schema.¹³ Then make the professional-body listings say the same thing. For an advisory firm in Canada that means the provincial CPA body and, for consultants, CMC-Canada. In the United States it is the state CPA society or IMC USA; in Australia, CPA Australia, Chartered Accountants ANZ or IMC Australia. Add the partner directories of the software you implement.

Month one: three service pages written as answers

Pick the three services a referral most often arrives for and write one page for each. The page answers the category question in its opening sixty words: what the service is, who it is for, how pricing works, and who delivers it. A page titled "Transfer pricing for mid-sized manufacturers in Canada" can be matched; a page titled "Advisory" cannot. Name the sector and the country plainly, because that is how the buyer phrases the prompt.

Month three: a twenty-prompt monthly check, then content

Write twenty prompts the way a buyer would, run them on ChatGPT, Gemini, Microsoft Copilot, Google AI Overviews and Google AI Mode, and repeat each on three separate days every month. Record whether the firm is named, which firms are named instead, and which sources each answer cites. Content comes after the check, because the check tells you which sources to earn and which questions to answer. Our view: content written before the entity work is the most common wasted spend in this category.

Example prompts for Canada, the United States and Australia

A good check prompt names the service, the sector and the country, because that is how a buyer phrases it and how the engine builds its list. The three below are the shape we use; swap in your own services and sectors.

Illustrative prompts; the right-hand column is our reading of the fix order, not a measurement.

Write the rest of the twenty the same way, then add a handful of branded prompts. Use your firm's name on its own, your firm's name with "reviews", and your firm's name beside the competitor a referral most often mentions. The branded prompts show what the engine believes about you today, which is the error register the week-one entity page has to correct.

Keep the prompts stable from month to month. A check that changes its questions cannot show movement, and movement is the point. Our view: twenty stable prompts run on three days beat a much larger set run once, because the engines vary enough between runs that a single pass is noise.

Referral strength against AI visibility

A strong referral network and strong AI visibility are different assets, and the matrix shows what each combination costs you. Most forty-person firms that ask this question sit in one cell, and it is the cheapest cell to leave. The axes are deliberately plain: how much of your work arrives by introduction, and how often the engines name you for category prompts.

Our view: the top-left cell is where most referral-led firms sit, and leaving it costs entity work and listings, not advertising. The firm already has the proof, the clients and the specialisms. It has not written them down where an engine can read them.

The bottom-right cell is the warning. A firm that becomes visible without a referral base is relying on an engine that changes its answers from one run to the next. Visibility supports a referral; it does not replace one.

What this desk research cannot tell you yet

It cannot tell you whether your firm is named, because nothing in it comes from our own captures. Every figure above comes from a published study, so this post is labelled desk research. A captured version follows once our AI answer monitor has run long enough to report intervals rather than single runs.

One informal observation, offered as colour rather than data. In a check we ran this month, the sources the engines drew on for this question were LinkedIn posts and YouTube videos. A single run per engine is not a measurement, so we give no figures from it. It does suggest the question has no sourced answer in the engines' pool yet, the gap this post tries to fill.

When the captured version arrives, the fix order above will not change. The entity facts, the listings and the answer-shaped pages are the inputs; the monitor shows whether the engines picked them up. If they have not after three months, the sources the engines cite for your category will tell you where to go next.