How to Run an AI Visibility Audit for an Accounting Firm

How to Run an AI Visibility Audit for an Accounting Firm

An AI visibility audit tells you whether ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews recommend your firm when a prospect asks for one — and whether what they say about you is accurate. This is the Q3 2026 version of the playbook: the prompts, the recording method, the page-level checks, and a monthly template you can run without any tooling beyond a spreadsheet and 90 minutes a month.

Traditional SEO audits check rankings on a results page. An AI visibility audit checks something different: whether your firm gets *named* inside a generated answer, whether the facts attached to your name are correct, and whether a competitor gets named instead. Those are three separate failure modes, and most firms have never measured any of them.

Why this matters now

Direct answer: because prospects increasingly ask an AI engine for a shortlist before they ever open Google, and if your firm isn't in that shortlist — or is described inaccurately — you lose the deal before you know it existed.

A partner search, a startup founder picking an accountant, a family office comparing tax advisers — a growing share of these journeys start with a prompt like "best small business accountant in Leeds" or "accounting firm for e-commerce sellers near me" typed into ChatGPT or Perplexity rather than Google. Google itself now surfaces AI Overviews on a large share of commercial queries, which means even "traditional" search traffic is being intercepted by a generated summary before a user reaches your website. If that summary doesn't mention you, your organic ranking is irrelevant.

None of this shows up in Google Search Console. None of it shows up in your usual rank tracker. You have to go and look, manually, engine by engine. That's the audit.

What questions should you actually test?

Direct answer: test three categories of prompts — direct discovery queries, comparison queries, and fact-check queries — across each engine, using the same wording every month so results are comparable over time.

1. Direct discovery prompts. These mimic how a prospect finds a firm cold:

• "best accounting firm for [your niche] in [your city]"
• "who should I use for [service, e.g. R&D tax credits] in [region]"
• "recommend a small business accountant near [postcode/zip area]"
• "top CPA firms for [industry] in [state]" (US) or "top chartered accountants for [industry] in [county]" (UK)

2. Comparison prompts. These surface who the engine considers your peer set — and whether you're in it:

• "compare [Your Firm] vs [Competitor A] vs [Competitor B]"
• "what's the difference between [Your Firm] and other accountants in [city]"
• "which accounting firm is best for a startup raising seed funding"

3. Fact-check prompts. These test accuracy, not just presence:

• "what services does [Your Firm] offer"
• "how much does [Your Firm] charge for [service]"
• "is [Your Firm] a good accountant for [niche]"
• "what do reviews say about [Your Firm]"

Run all three categories through ChatGPT (with and without browsing/search mode, since answers differ), Perplexity, Gemini, Claude, and a logged-out Google search that triggers an AI Overview. That's a minimum of 5 engines × 3 categories × 3–4 prompts = roughly 60 data points. Keep the exact wording identical every audit cycle — paraphrasing invalidates month-to-month comparison.

A note on Claude: it doesn't browse live by default, so its answers reflect training data and whatever you paste in-session. It's still worth testing because a growing number of tools now pipe Claude output into client-facing apps and agent workflows — if Claude has outdated or wrong information about your firm, that error propagates into places you can't see.

How do you record citations, non-citations, and hallucinations?

Direct answer: use a single tracking sheet with one row per prompt-engine pair, and score four outcomes — cited, not cited, competitor cited instead, and fact error — because each requires a different fix.

Build a spreadsheet with these columns:

| Column | What to enter | |---|---| | Prompt | Exact wording used | | Engine | ChatGPT / Perplexity / Gemini / Claude / Google AIO | | Date | Audit date | | Cited? | Yes / No | | Position in list | 1st, 2nd, 3rd... or "not ranked" | | Competitors named | List them, in order shown | | Firm facts stated | Address, services, fees, partner names, awards — whatever the engine claimed | | Fact accuracy | Correct / Outdated / Fabricated | | Source shown (if any) | URL the engine cited, if visible | | Notes | Anything unusual — tone, phrasing, missing niche |

Four outcomes to watch for specifically:

• Citation: your firm is named, with or without a source link. Record position — being third in a list of three is different from being first.
• Non-citation: the engine answers the query but never mentions you, even though you're a legitimate local option. This is the most common outcome and the hardest to diagnose, because there's no error message — you simply don't exist in the answer.
• Competitor mention instead: the engine names a rival firm for a query you should win — same city, same niche, comparable size. Note *why* it might have picked them: better schema, a recent press mention, a directory listing you don't have.
• Hallucinated firm facts: the engine states something false — a wrong address, a service you don't offer, a defunct partner's name, an inflated or invented fee range, a review score that doesn't exist. This is urgent to catch because it actively misinforms prospects and you have no notification system for it. Screenshot every hallucination with a timestamp; you'll want the evidence if you ever need to request a correction from a platform or update source content that's feeding the error.

Run this exact prompt set every month. A single snapshot tells you where you stand; three consecutive months tell you whether your fixes are working.

What page-level checks explain the results?

Direct answer: AI engines pull from pages that answer a question in the first 200 words, carry structured data confirming who you are, and show independently verifiable proof of expertise — audit your site against those six factors before touching anything else.

1. Answer blocks. Open any service page and check whether the first paragraph directly answers the question a prospect would ask ("What does a management accountant cost for a 10-person business?"). If the page opens with a mission statement or a stock photo caption instead of a direct answer, an AI engine has nothing clean to lift. Rewrite the first 40–60 words of key pages as a standalone answer — the kind of sentence that would make sense pasted into a chat window with no other context.

2. Schema markup. Check for Organization, LocalBusiness (or AccountingService where supported), Person (for named partners/advisers), FAQPage, and Review/AggregateRating schema. Use a validator to confirm the markup parses without errors — broken schema is often worse than none, because it signals inconsistency. If you publish FAQs, mark them up with FAQPage schema; this is one of the most directly "liftable" formats for generative answers.

3. Author and entity signals. Does each technical article or tax-guidance page show a named author with credentials (ACCA, ACA, CPA, EA, CGMA) and a link to a bio page? Engines weigh demonstrated expertise, and an anonymous "Team" byline gives them nothing to verify. Add short author bios with credential bodies named explicitly — "CPA licensed in California" reads very differently to a model than "our expert."

4. Niche proof. If you claim a specialty — R&D tax credits, dental practices, crypto accounting, nonprofit audits — check whether that claim is backed by a dedicated page with specific detail (typical client size, sample outcomes, relevant HMRC or IRS guidance cited by name) rather than a single line on your services page. Generic "we serve various industries" language reads as unverifiable and gets skipped in favour of a competitor with a dedicated, detailed niche page.

5. Reviews. Check Google Business Profile, Trustpilot, and any accounting-specific directories (in the UK: ICAEW "Find a Chartered Accountant," in the US: state CPA society directories) for review volume, recency, and whether star ratings are marked up with schema so they're machine-readable. A firm with 40 recent, detailed reviews is a far stronger citation candidate than one with 3 reviews from 2021.

6. Local consistency. Pull your firm's name, address, and phone number (NAP) from your website footer, Google Business Profile, Companies House filing (UK) or state registration (US), and any directory listings. Inconsistencies — an old suite number, a dissolved trading name, a phone number that only appears on one listing — create the exact ambiguity that produces hallucinated facts, because the model is reconciling conflicting sources and picking whichever appeared most often or most recently in training or retrieval.

Do this page-level pass on your homepage, your top 3 service pages, your about/team page, and your niche pages if you have them. That's usually 6–8 pages — enough to find the pattern without auditing the whole site.

How do you prioritise the fixes?

Direct answer: score every fix on a 2×2 of citation impact versus implementation difficulty, and do the high-impact/low-difficulty fixes inside the same week you find them — don't let them sit in a backlog next to the hard ones.

Use a simple matrix:

High impact, low difficulty — fix this week: - Rewrite the first paragraph of a service page into a direct answer block - Add FAQPage schema to an existing FAQ section - Correct a wrong address or phone number on one directory listing - Add named author bylines with credentials to existing articles

High impact, high difficulty — plan for this quarter: - Build a dedicated niche page with genuine specificity (case examples, typical engagement size, relevant regulatory detail) - Run a review-generation campaign to lift volume on Google Business Profile and Trustpilot - Get a partner quoted or profiled in trade press or a professional body publication — this earns the kind of independent third-party mention that AI engines weight heavily - Consolidate NAP data across every listing and directory (a genuine cleanup project, not a five-minute fix)

Low impact, low difficulty — batch these, don't prioritise them: - Minor formatting or metadata tweaks - Adding schema to already-low-traffic pages

Low impact, high difficulty — deprioritise: - Full site redesigns undertaken purely for AI visibility reasons without a clear citation hypothesis behind them

Score each item you found during the audit against this matrix before you fix anything. It's tempting to start with whatever's easiest to fix regardless of whether it moves a citation — resist that. A perfectly marked-up FAQ page nobody asks about doesn't move your visibility; a hallucinated address on your single highest-traffic page does.

What does a monthly reporting template look like?

Direct answer: one page, four sections — citation rate, fact accuracy, competitor movement, and fixes shipped — reviewed at a fixed time each month so trend lines are meaningful rather than anecdotal.

Keep the report to a single page. A longer report doesn't get read consistently, and consistency is what makes the trend meaningful.

Section 1 — Citation rate. Out of the ~15–20 prompts tested this cycle, what percentage cited your firm across all five engines combined, and per engine? Track this as a simple percentage over time: "Cited in 6 of 20 prompts (30%), up from 4 of 20 (20%) last month."

Section 2 — Fact accuracy. How many hallucinations or outdated facts appeared this cycle, and were the ones flagged last cycle corrected? List each one with the engine, the specific error, and status (open / source updated, awaiting re-index / confirmed fixed).

Section 3 — Competitor movement. Which competitors appeared this cycle that didn't last cycle, and vice versa? A new competitor appearing consistently across engines is worth investigating — check what changed on their site (new schema, new press mention, new reviews) between cycles.

Section 4 — Fixes shipped and next up. A short list: what you shipped since the last report, what's queued from the priority matrix, and one clear owner and date for the next highest-impact item.

Run the audit and file this report on the same day each month — the first Monday works well for most firms — so the interval between measurements stays constant. Three months of consistent measurement is enough to tell the difference between a real trend and noise from an engine's model update.

What to do with the first audit

Direct answer: expect a low citation rate on the first pass — a typical benchmark for a firm that has never optimised for AI visibility is single-digit to low-teens citation rates across the full prompt set — and treat that number as the baseline you're moving, not a verdict on the firm.

Most firms running this audit for the first time find they are simply invisible on the majority of prompts, rather than actively misrepresented. That's the easier problem to have. The harder finding — a hallucinated fact or a consistent competitor substitution — tells you exactly where to spend the next month's effort: fix the specific inaccuracy, strengthen the specific page, and re-test the exact same prompts next cycle.

The audit itself costs nothing but time. What separates firms that improve their AI visibility from those that don't isn't access to better tools — it's whether they run this same 90-minute check every month, log it in the same format, and act on the priority matrix before the next cycle starts.