An AI answer recommended an accountancy firm in Leeds. Beside a claim about its services, the answer linked to another practice’s website. The source label even carried the other practice’s name.
The recommended firm might offer those services. But a page describing a different business cannot establish that it does.
That was one finding from a small Answari pilot: eight fixed questions about accountants and financial advisers, put to ChatGPT, Perplexity and Gemini on one date. The 24 answers included useful, supported details as well as discrepancies. Three examples show why being recommended by AI is only part of the picture. The description a buyer reads matters too.
The source belonged to another practice
The question asked which accountants a small limited company in Leeds should consider. One recommendation described a firm’s broader commercial advice, including salary and dividend planning for directors and tax planning. The source beside that claim described another accountancy practice’s services.
Two other source markers in the same recommendation did lead to the recommended firm’s own website. A correct link elsewhere did not resolve this mismatch. Separately, the answer placed the firm on a road that did not match the Leeds address on its current contact page.
The firm’s own site lists year-end accounts, corporation tax and tax planning. The limited pages inspected did not settle whether it offers the specific salary and dividend planning described. We do not conclude that it lacks that service, and we do not know why the citation or address differed.
A published fee held up
Another answer, comparing accountants for a small limited company in Leeds, gave a firm’s typical monthly fee range and what the package normally includes. It noted that more complex or VAT-registered companies pay more, supplied contact details and linked to the firm’s own service and contact pages.
Those pages supported the fee, scope, phone number and email address. The answer also named an individual, while advising the buyer to confirm who would handle day-to-day work. The service page described that person’s role in reviewing accounts and as the usual named accountant for smaller clients; it did not promise who would perform every task for a new client.
Here the answer was useful. It carried the published price with its conditions, preserved uncertainty about the working relationship and did not present advertised fees as a guaranteed quote.
The description differed from the firm’s disclosure
Asked which financial advisers to consider for pension options in Manchester, one answer introduced its recommendations as “three well-regarded independent firms”.
One of the three states on its own website that it represents a single group and advises only on that group’s wealth management products and services.
The collective description therefore differed from the firm’s own disclosure. This finding concerns the AI answer’s wording. It is not evidence of poor advice or a regulatory breach, and it says nothing about the other two firms. The opening paragraph’s grouped source marker was not fully recoverable, so we do not infer where the wording came from.
We withhold business names and the product behind each example. These cases illustrate checks, not judgements on firms or a ranking of AI products. Distinctive details can still make a business traceable.
What the comparison answers contained
Twelve answers responded to questions asking explicitly about cost, people and contact details. Across the first three firms in each answer:
- 21 of 36 included a numerical fee or fee range.
- 23 of 36 named an individual in a professional role.
- 30 of 36 gave a phone number or email address.
A numerical fee appeared in 11 of 18 accountancy appearances and 10 of 18 financial-adviser appearances. Fee practices differ, so these are separate descriptions of the sample.
These are counts of information present, not accuracy scores. A repeated firm counts each time it appears. Across all 24 answers there were 72 business appearances, not 72 different firms. The three examples were selected to show different findings; they cannot establish how often any finding occurs.
Read the evidence extracts and counting rules.
Check what AI tells a prospective client
- Keep the question and answer. Ask a question a buyer would ask, naming your service and town. Record the date, product, exact question and full response. One answer is a snapshot; repeat on another date or product before drawing wider conclusions.
- Check the business and the claim. Confirm the identity, location and service. A similarly named firm or an old office can produce a plausible-looking mismatch.
- Check fees and people in context. Distinguish a starting price from a quote. Check inclusions, conditions and VAT. A person’s name does not establish that they would handle a new client’s work.
- Follow the sources. Open each link. Does it concern your firm and support the sentence beside it? Check phone numbers, email addresses and postal addresses against your current contact page. Use the relevant official register for regulated status or permissions.
- Record the result and improve what you control. Mark a claim supported, contradicted, unclear or not verifiable. An inaccessible page does not prove a claim false. Correct outdated pages and profiles you control; this pilot did not test whether an edit changes an AI answer.
Download the AI answer check worksheet.
Our reviews of accountancy practice pages and financial advice firms’ pages examine what businesses publish for buyers. This pilot examines what AI answers say about them.
Read the full method and limits
Eight questions were fixed before collection: two sectors, two cities (Leeds and Manchester), and two question types. One asked which three firms to consider; the other asked for comparisons, source links and practical details. Adviser questions concerned pensions. One successful answer per question and product was retained on 5 October 2026, with the first three recommendations counted.
Signed-in sessions used ChatGPT Temporary chat with Instant and web search, Perplexity Incognito Search, and Gemini Temporary chat with Flash. Four Gemini first attempts returned errors; one permitted retry each supplied the retained answers. Personalisation was not audited. Location was specified in the questions but not otherwise controlled; Gemini displayed a saved UK location outside both cities.
Fees required a number attributed to a paid service, including a range, percentage, starting price or qualified estimate. Free meetings, asset minimums and general market estimates were excluded. Names required an explicit professional role; a first name appearing only through testimonials was excluded. Clarifying that rule before publication reduced the original named-person count from 24 to 23. Contact details required phone or email, not a website alone.
The original captures are retained. The example sources were rechecked on 6 October; page-reading tools may use cached content, and not every grouped citation was inspectable. A firm’s own statement does not independently verify its credentials. One date, two cities and two sectors cannot establish an error rate, national picture, product ranking or effect on enquiries. No firm was contacted.
This is Answari’s own research, not independent validation.
Have your own business checked
Answari’s £1,500 Visibility Audit tests at least ten agreed buyer questions across four AI products and provides captured evidence and a prioritised action plan. The fee covers an agreed scope, no VAT is currently charged, and any implementation or ongoing support is a separate choice. It cannot promise a recommendation, ranking or enquiry.