Strategy & Planning · 9 min read · July 15, 2026
What questions clients really ask the AI about tax advisors
"Which tax advisor near me handles GmbH formations?" That question increasingly goes into a chat box instead of a search bar, and what comes back is a short list of names. The scale is not marginal: ChatGPT reported 900 million weekly users in February 2026, Google says AI Overviews now reach over two billion people a month, and SparkToro measured 68% of US Google searches ending with no click to any website at all. The shortlist gets drawn before anyone opens a firm's website, and a firm the model can say nothing about is not on it. Generative Engine Optimization for tax firms is the work of making sure it can.
Why the first question goes to a chatbot, not to you
The first contact with your firm increasingly happens somewhere you cannot see it. A sole trader weighing up a switch to a GmbH no longer types "tax advisor GmbH conversion" into Google at nine in the evening and works through ten blue links. They ask ChatGPT the whole question at once: "I'm a freelancer with around 90,000 euros of profit, is a GmbH worth it for me, and which tax advisor near me handles that?" The reply arrives in one block, already written, sounding like someone neutral who has weighed it up. That reframes the entire first contact, and none of it shows up in your analytics.
The mechanics of competing change with it. Ranking third on Google still put you on the page; not being named in a generated answer puts you nowhere, because there is no second results page and no way for the client to see who was left out. Pew Research watched the real browsing of 900 US adults across nearly 69,000 Google searches: when an AI summary appeared, people clicked through to a website in 8% of visits against 15% when it did not, and clicked a link inside the summary itself in just 1%. Google disputes the methodology. The direction is still hard to argue with, and it points at the answer becoming the destination.
For tax advisors the stakes are sharper, because the sector runs on trust and demonstrable competence. If a model names three firms and none of them is yours, it reads to the asker like a quiet judgement on your standing, and you never learn you were in the running. Ranking well is not a hedge either: Ahrefs found only about 6 to 8% of the URLs ChatGPT cites also sit in Google's top ten for the same query, and roughly 80% of them do not rank in the top 100 at all. Whatever earns a mention in an AI answer, it is not simply the thing that earned you position one.
What clients actually type: situations, not keywords
Questions put to a model look nothing like search terms. They are longer, they carry the person's actual situation, and they usually arrive already framed as a decision. "I run an online shop selling into four EU countries, which tax advisor actually knows the OSS procedure?" "My advisor takes three weeks to answer anything, what do I need to watch when I switch?" "What does a small GmbH with two employees pay for monthly bookkeeping?" Those are advisory briefs, not keywords, and no keyword tool will ever show them to you.
Specialisation comes up constantly. Tax advisor for doctors. For influencers and content creators. For crypto holdings. For a hospitality business heading into an audit. A model reaches for niches because a niche is something it can match against a question. A homepage offering "comprehensive tax advice for every industry" hands it nothing to match on, and gets treated accordingly, which is to say not at all.
The other cluster is process and money: how a first meeting runs, whether annual accounts can be filed digitally, exactly which documents you need for an income tax return, how long a handover from the previous advisor takes. These are the questions you answer on the phone several times a week and almost never write down. Written down properly, they are the raw material a model quotes from, because they are the only pages on the web that answer the question the asker actually asked.
How an answer engine decides which firm to name
Two different things can happen when someone asks. Either the model answers from what it absorbed in training, or it fetches pages first. Profound looked at roughly 730,000 US English ChatGPT conversations from late 2025 and found only about 18% trigger a web search at all; the rest come out of memory. For a local, specific request like "tax advisor in Regensburg for trades businesses", retrieval is the likely path, and ChatGPT's search mode or Google Gemini will pull whatever is currently indexed: your site, your Google Business Profile, professional directories, review portals, specialist articles. Missing from all of those, there is nothing to retrieve and nothing to recommend. Answered from memory instead, you are betting on whether your name appeared often enough on the open web to be in there at all.
What decides the match is dull and mechanical: does anything you have published line up with the question. Google is unusually plain about this for its own AI features. There is no AI-specific markup, no special file, no separate channel to be in; the generative features run on the same core Search ranking and quality systems, and a page qualifies as a supporting link by being indexed and eligible to show with a snippet. So a clearly structured page on financing a start-up gets recognised when someone asks about financing a start-up. "Your partner in all tax matters" answers no question, so there is nothing for a machine to pick up.
Then there is consistency, which sounds like housekeeping and is not. If your firm name, address, opening hours and service list read the same on your website, your Google profile, the chamber directory and LinkedIn, a model has one version of you to state. If they contradict each other it picks one, and it will state the wrong one with complete confidence: Columbia's Tow Center ran 1,600 queries across eight AI search tools and found more than 60% of the answers wrong on basic source details, with ChatGPT expressing any uncertainty in only 15 of 200 cases. Schema.org markup helps here, not as a ranking trick, since Google has said for years that structured data is not a direct ranking factor, but because it turns facts like hours and fee basis into something machine-readable instead of guessable.
From a services page to answer pages
Generative Engine Optimization comes down to writing so that a paragraph of yours can be lifted out and used as the answer. In practice: the client's question becomes the heading, the answer lands in the first two or three sentences, and the detail, the exception and the worked example follow after. It is why a page called "What does monthly bookkeeping cost a small GmbH?" earns citations that a page called "Services" never will. One of them is an answer. The other is a menu.
Fees are where most firms go quiet and where the questions are loudest. The StBVV sets the framework for what statutory work is billed at, and § 57a StBerG confines you to factual information about your practice rather than touting for a particular engagement. Neither stops you explaining how a fee is actually built: which basis applies to which service, what pushes the effort up, whether a first meeting is chargeable and what it covers. Say nothing and the model fills the gap from somewhere else, which is the version you least want quoted.
Then add the depth that proves the positioning. One properly worked example of VAT on intra-community supplies, in language a founder can follow, does more than a page listing every service you offer, because it gives a machine something specific to hang a recommendation on. The one controlled study in this area, a 2024 SIGKDD paper by Aggarwal and colleagues, points the same way: adding statistics, quotable lines and cited sources moved visibility most, while keyword stuffing did nothing. Read it with care, though. It measured a simulated engine assembling answers from Google's top five results, scored on its authors' own metrics, not clicks in a live product.
Reviews and mentions: the best-evidenced lever you have
None of this rests on your own website alone, and the strongest evidence anyone has points off it. Ahrefs compared about 75,000 brands and found that how often a brand is mentioned across the web correlates with how often AI cites it at roughly 0.664, about three times the correlation for backlinks at 0.218. Put plainly: being talked about elsewhere outweighs anything you can do to your own pages. For a firm that means Google reviews, tax-advisor directories, the regional business press and professional portals, all feeding the picture a model assembles when your name comes up.
So ask for reviews that say something. "Took our GmbH formation from the notary appointment through to the first payroll run" is useful in a way that "very good, highly recommended" never is, because it names a service that matches a question somebody will ask. Reply in public too, including to the awkward ones. Responsiveness is one of the few things about a firm that a stranger, or a machine, can verify from outside.
The rest of the mentions you have to go and earn. A piece in the regional chamber magazine, a quote in a trade newsletter, a guest article on a tax portal: each attaches your name to a specific competence somewhere you do not control, which is exactly why it counts for more than your own claim to the same thing. It is worth knowing which channels get cited. In Ahrefs' analysis of 1.4 million ChatGPT prompts, pages retrieved through the general search channel were cited 88% of the time, Reddit 1.9% and YouTube 0.5%. Go after the trade press and the directories before the social feeds.
A routine a small firm can actually keep
Start with a baseline you can actually look at. Write down the ten questions clients genuinely bring you, then put each one to ChatGPT, Gemini and Perplexity. Save every answer. Are you named? Who is? Run each question twice, once with web search on and once without, because those are two different systems answering: one is reading pages today, the other is reciting what it absorbed months ago. Then read the websites of the firms that did get named and note what they publish that you do not.
Then build the answers. Five to start, one page each. The question as the heading, the answer in the first two sentences, then the detail, a worked example and a plain statement of who you do this for. Write it the way you would say it out loud to a client who does not have the vocabulary; the clarity that helps them is the same clarity a model needs to lift a sentence out with confidence. Publish under a named author with a date, so reader and machine can both see who stands behind it.
Then keep it alive, which is what separates this from a project. Make the details match everywhere, keep the Google profile current, ask for a review after every engagement that went well, and revise pages when the law under them moves; a page on this year's thresholds still quoting figures from two years ago is worse than no page. One note on speed: in Ahrefs' prompt study the pages ChatGPT cited had a median age of around 500 days, so this compounds slowly and rewards whoever started earliest. Put a recurring slot in the calendar instead of waiting for a free afternoon, because the free afternoon is not coming.
What to skip, and what actively hurts
Volume is the standard mistake. Twenty thin posts written to feed the machine lose to five pages that answer something, and Google says as much itself: its guidance warns against producing separate content "for AI" and files that behaviour under scaled content abuse. The AI-specific plumbing sold alongside it is no better a bet. John Mueller confirmed in 2025 that no Google Search system reads llms.txt, and across roughly 137,000 sites publishing one, an Ahrefs analysis found about 97% saw no measurable referral traffic from it. Add the file if you like, it costs an hour. Do not mistake it for the work.
Two things actively cost you. The first is publishing model-generated tax copy without reading it, because a wrong threshold or a misstated deadline stops being a visibility problem and becomes a liability one with your name on it. The second is guarantees. "We will save you thousands in tax" runs into the profession's advertising rules before it runs into anything else, and it reads to a model the way it reads to a sceptical client, as a claim with nothing behind it. Sober and checkable travels further with both.
The last one is quieter. Referrals still work, and it is tempting to treat them as a reason to skip all of this. They are not, because the referred client now checks you before calling, and the check happens in a chat window. If the machine has nothing on you, or has two sets of opening hours and an address you left three years ago, a good referral can still die on the way to the phone.
Conclusion: findable is the new reachable
The way a client finds a tax advisor changed quietly, with no announcement. The question stopped being whether your website looks the part and became whether a model can say something true and specific about you at the moment someone asks. This is not a fad sitting on top of search. It is search carrying on in the direction it was already heading, faster, and with far less room for whoever came second.
The good news is that this profession is unusually well set up for it. Your working week is made of exact questions with defensible answers, which is precisely the material a model needs before it will name anybody. And the upside is measurable: Ahrefs found that pages cited inside an AI Overview pick up around 35% more organic clicks than pages that rank without being cited. The job is not to get louder. It is to stop hiding the substance behind phrasing that says nothing.
Starting now is worth something in itself, because most firms in your postcode have not looked at this yet and the effect builds with age. Every question you answer properly, every detail you make consistent, every review that names a real piece of work adds to your AI visibility. Begin with three pages rather than thirty, and check your progress by putting the same ten questions to the models each quarter. Keep that up and your firm stops being the one left out of the answer and starts being the answer.
Common questions
How do I check whether ChatGPT knows my firm at all?
Two tests, five minutes each. First ask the client question without naming yourself, along the lines of "which tax advisor in my town helps with a GmbH formation?", and see who gets named. Then ask directly what the model knows about your firm, and read the answer for errors, because a confidently wrong address or service list is a problem in its own right. Run both in ChatGPT with search on, in Gemini and in Perplexity, since they retrieve differently and will disagree. Repeat each a couple of times, because answers vary between runs, and keep the screenshots so you have a baseline to compare against later.
Am I even allowed to publish fees so the models can pick them up?
Not as a flat price list for statutory work: the StBVV sets the fee framework, and § 57a StBerG limits you to factual information about your practice rather than soliciting a specific engagement. What is allowed, and what people are actually asking about, is the mechanics. Which basis applies to which service, what pushes the effort up or down, whether a first meeting is chargeable and what it covers. That is more useful to a model than a bare number anyway, because it can be explained rather than just repeated. If you are unsure where the line falls for a particular wording, check it with your chamber before publishing.
Is a well-kept Google Business Profile enough on its own?
No, though it is the cheapest item on the list and worth doing first. It is often the source of the plain facts a model repeats about you, address, hours, what you do, which is exactly the material that goes wrong when it is stale. What it cannot do is answer a question. That takes pages of your own addressing real client situations, plus the reviews and third-party mentions that turned out to be the strongest correlate anyone has measured. The profile keeps the facts straight; the answer pages and the mentions are what get you named.
Read on