Who ChatGPT recommends when a founder asks for an accounting firm

A 29-question measurement of the startup-accounting niche. Run July 2026, fresh logged-out session per question, every named firm logged with position.

Method

Questions were drawn from a frozen 100-question panel modelling how founders actually phrase vendor research: category searches ("top startup CPA firms in the US"), situation searches ("best accounting firm for a startup that just raised a seed round"), comparisons ("Kruze vs Pilot"), and service queries (R&D credits, 409A, ASC 606). Each ran in a fresh, logged-out ChatGPT session with exact phrasing. Screenshots retained. Because the panel is frozen, re-runs produce directly comparable movement data.

Result: the answer-box leaderboard

Three findings that change how you should think about this

1. The model issues verdicts, not lists

Answers didn't stop at naming firms. They ranked them and recommended one: "Best overall" on the open category question; "If I were choosing one default CPA partner for a YC company, I'd start with…" on another. A founder reading that has a shortlist and no reason to keep looking.

2. Being known is not the same as being recommended

One firm in the panel was described accurately by ChatGPT — correct services, correct ideal client, fair strengths — and still appeared in only 1 of 29 answers. The model holds an accurate representation of that firm and recommends competitors anyway. Absence here is a ranking outcome, not an awareness gap, which is why it responds to source-layer work.

3. The engine reads vendor content and says so

Asked to compare two firms, ChatGPT volunteered that much of the comparison material online is published by one of the firms itself — flagged the bias, then used the content regardless, because that is what exists on the topic. Several answers also carried referral links tagged utm_source=chatgpt.com, meaning the firms winning this channel are measuring the traffic it sends them.

What moves the number

Roughly 85% of AI citations trace to third-party sources rather than a brand's own site. For accounting specifically, that source layer has structure worth auditing directly: the IRS federal directory of credentialed preparers, AICPA verification surfaces, Clutch and G2 provider profiles with formal "outsourced accounting" and "fractional CFO" taxonomies, mainstream review platforms, and the editorial "best firms" lists AI quotes constantly.

Two smaller firms in this run entered answers on the strength of a single well-built page each — one a published VC-partner programme page. Neither outspent the leaders. They were simply legible to the systems doing the recommending.

Limitations, stated

Single platform (ChatGPT), single day, 29 of a 100-question panel, one niche. AI answers vary between sessions, accounts, and locations — which is exactly why measurement uses a fixed panel at fixed intervals and reports month-over-month movement rather than treating any single answer as evidence. One panel question exhibited location-aware behaviour and is documented as a known variance class. Full engagements extend to 100 questions across four platforms with complete citation mapping.