Somewhere today, a prospect asked an AI a question you would have paid to answer: “Who’s a good [your trade] in [your city]?” “What should I look for when hiring a [your profession]?” “Is [your company] legit?” The machine answered — naming some businesses, citing some sources, characterizing some reputations — and you have no idea what it said. Maybe it recommended you, in which case something in your digital footprint earned it and you should know what. Maybe it recommended three competitors and had never heard of you, in which case a sales conversation you’ll never know about was lost before it began. Maybe it described you inaccurately — the wrong services, the old address, a mischaracterization — in which case you have a correction problem quietly compounding. All three states are invisible until you look.
The mention audit is the looking, done as a discipline instead of a one-off curiosity check. The distinction matters because casual checking misleads: AI answers vary by phrasing, session, region, and day; one query in one chat window is an anecdote, and businesses have made both panic decisions (“ChatGPT doesn’t know us!” from one unlucky prompt) and complacency decisions (“we’re fine” from one lucky one) on anecdotes. A structured audit — a fixed prompt set, run across the engines that matter, scored consistently, repeated quarterly — turns the question from vibes into a trend line: where you appear, how you’re described, who wins the mentions you don’t, which sources the engines lean on when talking about your category, and whether all of it is moving in your favor.
This guide is the audit kit: the prompt matrix (the five query families that map how prospects actually use these tools, from direct recommendation asks to validation checks), the engine list and the sampling rules that keep results honest, the scoring sheet — presence, position, sentiment, accuracy, and citations — the competitive read, the source-tracing step that turns findings into an action list (because every AI answer has upstream inputs you can influence), and the cadence and change-management that make it a program rather than a scare.
You can’t manage what the AI engines say about you until you systematically ask them. The audit: a fixed prompt matrix × the major engines × consistent scoring, quarterly. Prompt families (5): recommendation asks (“best [service] in [city]” and natural variants), problem-first asks (“my [problem], who should I call”), brand validation (“is [company] reputable,” “[company] reviews”), comparison asks (“[you] vs [competitor]”), and category education (“how to choose a [profession]” — where being cited as a source builds the authority layer). Engines: ChatGPT (with search), Google AI Overviews + AI Mode, Gemini, Perplexity, Copilot — sampled with variation rules (multiple phrasings, clean sessions, note the region) because single answers are anecdotes. Score each cell: mentioned? position among mentions? sentiment/framing? factual accuracy? which sources cited? Then the two analysis moves: the competitive read (who owns the mentions you don’t — and their visible footprint differences) and source tracing (the directories, review sites, articles, and your own pages the engines cite are the levers — AI answers are downstream of a fixable web). Findings route to action: absence → entity/footprint work, misdescription → the correction protocol, weak citations → the content and PR targets. Quarterly re-run, same matrix, trended.
The Prompt Matrix: Five Families That Map Real Usage
| Family | Example prompts (build 3–5 variants each) | What it reveals |
|---|---|---|
| 1. Recommendation asks | “Best [service] company in [city]”; “Who do you recommend for [service] in [city/area]?”; “I need a reliable [trade] near [neighborhood]” | The money question — direct share of AI-mediated referrals; run per service line and per key city |
| 2. Problem-first asks | “My [symptom: foundation is cracking / IRS sent a letter / AC died], what should I do and who should I call in [city]?” | How you surface when prospects describe problems, not services — often different winners than family 1, and closer to how real people talk to chatbots |
| 3. Brand validation | “Is [your company] reputable?”; “[your company] reviews”; “Tell me about [your company]” | What the machine says when your name is already on the shortlist — the accuracy and framing check, and the feed for the correction protocol |
| 4. Comparison asks | “[you] vs [competitor]”; “alternatives to [dominant competitor] in [city]” | How you’re positioned head-to-head, and whether you appear in the “alternatives” answers that steal share from category leaders |
| 5. Category education | “How to choose a [profession]”; “What questions should I ask a [trade] before hiring?”; “Red flags when hiring [service]” | Whether your content earns citations at the research stage — the authority layer that precedes recommendations, and the direct target for your question-harvest content |
Build the matrix once: families × your service lines × your priority cities — a typical single-market service business lands at 20–35 prompts; keep it versioned, because trend comparison requires a stable instrument.
Engines and Sampling Rules
- The engine list, weighted by your audience: ChatGPT (with its web search active — the mode that produces cited, current answers), Google’s AI surfaces (AI Overviews on your prompt-matrix queries as normal searches, plus AI Mode where available), Gemini, Perplexity, and Copilot. If your buyers skew enterprise, add Claude; the list evolves — revisit it annually.
- Variation is the method, not noise to eliminate: run each prompt family’s 3–5 phrasings rather than one canonical string — the spread across phrasings is data (a business mentioned on every variant has robust presence; one appearing on a single lucky phrasing doesn’t).
- Session hygiene: fresh chats per prompt (prior conversation contaminates answers), signed-out or clean profiles where feasible, and note that personalization and region still leak in — record the location context, and if you serve multiple metros, sample from each where possible.
- Record everything verbatim: screenshot or paste full answers into the audit sheet, including the citation lists — the sources are half the value (below), and quarter-over-quarter language comparison catches framing drift that memory won’t.
The full matrix across five engines sounds heavy; in practice a 25-prompt matrix runs in a structured afternoon — and quarterly is the right cadence for a moving-but-not-daily target. What to avoid: ad-hoc daily spot checks, which produce anecdote whiplash (AI answers vary run to run; you’ll ‘lose’ and ‘regain’ mentions that never actually moved) and no comparable trend. Between quarterly runs, the standing instruments carry the watch: the referral and brand-echo metrics from your measurement stack, and the intake question — if those move sharply, run an off-cycle audit; otherwise, trust the cadence.
The Scoring Sheet: Five Columns Per Cell
- Presence: mentioned / not mentioned / mentioned only after a follow-up push. The binary that gates everything — reported as mention coverage: the share of prompt-cells where you appear.
- Position: first-named, top-three, or afterthought — AI answers have an order, users weight it like they weight rankings, and “we’re mentioned” hides the difference between leading the answer and trailing it.
- Framing: the adjectives and specifics attached — “highly rated, specializes in X, family-owned since…” versus a bare name-drop versus hedged language (“mixed reviews”). Copy the exact phrases; framing drift across quarters is an early signal in both directions.
- Accuracy: every factual claim checked — services, service area, address, pricing characterizations, history. Errors feed the correction protocol with verbatim evidence.
- Citations: which sources the engine linked or named — your site? your Business Profile? Yelp, the BBB, a directory, a news article, a competitor’s comparison page? This column is the lever map, and it’s the one businesses skip.
The Two Analysis Moves
The competitive read
For every recommendation-family cell where you’re absent, log who is mentioned — then profile the recurring winners’ visible footprint against yours: review volume and recency on Google (the input the engines cite most for local recommendations), presence on the specific directories and platforms the answers cited, content depth on the exact questions asked (do they have the “how to choose” guide that got cited?), Wikipedia/knowledge-panel-grade entity clarity, and press or association mentions. The deltas are rarely mysterious — the businesses AI recommends are overwhelmingly the ones with strong, consistent, machine-checkable footprints — and the profile turns envy into a roadmap. Where nobody local is mentioned (the engines answer generically), you’ve found open ground: the first business in your market to build the citable footprint often becomes the default answer.
Source tracing → the action list
Aggregate the citations column across all cells and rank the sources by frequency: this is the map of what the engines trust about your category in your market. Route the findings: cited sources where you’re weak or absent → the profile/review/directory work list (claim, complete, and earn reviews on the platforms that actually feed answers — not the 40-directory spray); cited content types you don’t have → the content plan (the education-family questions, answered on your site with the expert-attribution layer that makes them citable, delivered in crawler-readable HTML); your own pages cited → protect and strengthen them (they’re earning machine trust — keep them accurate, fresh, and structured); inaccurate sources cited → the correction protocol, at the source. This routing is the audit’s whole point: every finding becomes either footprint work, content work, or correction work — the three levers of the broader AEO discipline.
AI answers are non-deterministic: the same prompt can produce different mentions across runs, sessions, and days — which breaks naive interpretation in both directions. Traps: treating one absence as ‘we’re invisible’ (run the variants; presence on 3 of 5 phrasings is presence with weak robustness, not absence); treating one mention as ‘we’ve arrived’ (same logic mirrored); comparing this quarter’s single run to last quarter’s single run and narrating the noise (compare coverage rates across the matrix, where the aggregation dampens run-to-run variance); and A/B-testing your footprint changes against next week’s answers (attribution at that resolution doesn’t exist — judge the program on quarterly coverage trend plus the downstream metrics). Also honest: the engines update on their own schedules, personalization you can’t fully control colors what your prospects see versus what your clean-session audit sees, and mention coverage is a leading indicator, not revenue — pair it with the measurement stack’s lagging numbers before declaring victory or crisis. The audit’s power is the trend of a stable instrument; respect the error bars and it stays honest.
From Audit to Program: The Quarterly Loop
Each quarter: (1) run the matrix (same prompts, versioned changes only with a changelog); (2) score the five columns; (3) report four numbers and one table — mention coverage overall and by family, first-named share, accuracy issue count, competitive mention-share on recommendation prompts, and the top-cited-sources table with your presence on each; (4) route findings into the three work streams (footprint, content, corrections) with owners; (5) log the quarter’s completed work next to the trend, so the program’s narrative — footprint built, coverage moved — stays evidence-shaped. Pair the whole thing with the measurement stack: this audit is the leading indicator (are we in the answers?), those numbers are the lagging one (is it becoming demand?) — and together they’re the honest dashboard for the channel everyone’s guessing about.
5 Common Mention-Audit Mistakes
- One prompt, one engine, one conclusion. Anecdote-driven panic or complacency — the matrix and variants exist because the system is stochastic.
- Skipping the citations column. Presence data without the source map is a scoreboard with no playbook — the citations are the actionable half.
- Auditing without routing. A quarterly report nobody converts into footprint, content, and correction work is a subscription to watching competitors win.
- Chasing daily fluctuations. Run-to-run variance narrated as trend — quarterly cadence, coverage rates, and error-bar humility.
- Ignoring the education family. Recommendation mentions are downstream of authority; the “how to choose” citations are where that authority gets built first.
Frequently Asked Questions
The AI engines don't mention my business at all. What actually gets a business into the answers?
Work the evidence chain the engines themselves use — your audit’s citation column shows it for your specific market, and the pattern is consistent: AI recommendations are assembled from machine-checkable public footprint, so the program is making yours strong, consistent, and citable. The usual priority order for a local service business: Google Business Profile excellence (complete, accurate, category-correct, actively earning reviews with substance — the single most-cited input for local recommendation answers); review mass and recency on the platforms your audit saw cited (Google first, then the vertical’s platforms — and responses that show an operating business); entity consistency everywhere (same name, services, and facts across your site, profiles, and the directories that matter — contradictions read as unreliability to systems cross-checking sources); a website that answers the education-family questions in crawlable HTML with real authorship (the content that earns citations at the research stage, which precedes recommendation presence); and third-party corroboration (association memberships, local press, the ‘best of’ lists the engines cited for your category — earned, not bought, where legitimate). What doesn’t work: prompt-stuffing your site with ‘best in [city]’ claims (self-assertion isn’t corroboration), directory spam, or any scheme faster than deserving the mention — the engines’ whole design goal is synthesizing what credible sources agree on, which makes the honest footprint the only durable exploit. Timeline honesty: quarters, not weeks — which is why the quarterly trend, not next month’s answers, is the success metric.
Which AI engines should a local service business actually care about?
Weight by usage and by influence-per-use, and the 2026 answer for most local services is a short list. Google’s AI surfaces first — AI Overviews sit on top of the search behavior your prospects already have, so their reach into your funnel exceeds every standalone chatbot combined for most local categories; audit your money queries’ Overviews (and AI Mode where rolled out) as seriously as you ever audited rankings. ChatGPT second — the largest standalone assistant by usage, increasingly consulted for exactly the recommendation and validation questions in your matrix, with web-search-backed answers that cite and link. Perplexity and Copilot next — smaller but citation-forward (Perplexity’s users skew research-heavy; Copilot rides Windows and Bing distribution), cheap to include since the same matrix runs everywhere. Gemini’s standalone app matters more as Android integration deepens; Claude indexes toward professional and B2B research. The honest calibration: your intake question’s ‘AI’ answers and your referral data tell you which engines your prospects actually use — let your own funnel’s evidence promote or demote engines in the matrix annually, and don’t spend audit effort proportional to tech-press coverage rather than to your customers’ behavior.
Can I just pay a tool to track AI mentions instead of doing this manually?
Tools now exist that automate the mechanical layer — scheduled prompt-sampling across engines, mention detection, share-of-voice dashboards, change alerts — and they’re worth evaluating once the manual audit has proven the stakes; but understand what automation does and doesn’t replace. What it does well: frequency and scale (daily sampling across hundreds of prompt variants dampens the variance problem better than your quarterly afternoon), competitive breadth (tracking mention share across a whole competitor set), and alerting (a framing shift or accuracy problem surfacing between your quarterly runs). What it doesn’t replace: the matrix design (the tool samples the prompts someone chose — garbage prompts, garbage share-of-voice), the citation analysis and routing (dashboards report presence; converting the source map into footprint, content, and correction work is still judgment), and the accuracy check (whether claims about your business are true is something only you can score). Evaluation criteria if you shop: disclosed sampling methodology (how many runs, which modes, what session hygiene), region control (answers for your metro, not a data center’s), verbatim answer capture (you need the language, not just a mention count), and honest variance handling (rates with sample sizes, not single-run snapshots presented as state). Sensible sequence: two manual quarterly cycles first — they cost an afternoon, teach you the terrain, and make you a competent buyer — then automate if the category’s competitive stakes justify a subscription.
We showed up in the audit but described as 'having mixed reviews.' Where does that framing come from and can it change?
Hedged framing is almost always traceable — and therefore workable. Where it comes from: the engines synthesize characterizations from the sources they retrieve, so ‘mixed reviews’ typically reflects something checkable — a rating gap across platforms (4.8 on Google, 3.2 on Yelp reads as ‘mixed’ to a system averaging sources), a cluster of negative reviews (recent, or old-but-prominent, or concentrated on one platform), a low review count that makes any negatives loom large, unanswered complaints on BBB-type sites, or a negative press item or forum thread ranking for your brand queries. Your audit’s citation column plus a brand-query search usually identifies the culprit sources in minutes. The workstream, in effect order: fix the real problem if reviews reflect one (framing follows reality with a lag); build review velocity where you’re strong and especially where you’re thin (the platform dragging the average needs volume of genuine recent experiences — solicited legitimately); respond publicly and professionally to the negatives (response presence measurably softens how humans and synthesizing systems read a profile); resolve the resolvable complaints on complaint platforms (many allow status updates); and where a specific characterization is factually false, run the correction protocol at the source. Then re-audit on the normal cadence: framing shifts as source reality shifts — typically visible within a couple of quarters of genuine review-mass improvement — and the verbatim-language column of your audit is exactly the instrument that will show the ‘mixed’ hedge fading.
How do I audit Google AI Overviews specifically — they seem different from the chatbots?
They are different, in ways that change the method. Mechanics: AI Overviews trigger on searches rather than conversations, so your instrument is your keyword set, not a prompt matrix — run your money queries (recommendation, problem, validation, and education families translated into search phrasing) as normal Google searches, per priority city, and record: does an Overview appear at all (coverage varies enormously by query type — heavily present on informational and education-family queries, more sporadic on straight local-commercial ones where the map pack still dominates), are you cited within it (the link cards and referenced sources), and how does its summary characterize the category and any named businesses. Region matters more than in chatbots — use location settings or sample from within your metro, since local results shape the answers. What to score: Overview presence per query, your citation rate within present Overviews, whether the local pack/your Business Profile appears alongside (the Overview and the pack often share the visual field — your total presence is the combination), and the cited-source map (frequently your category’s ranking articles and directories — the same lever list as the chatbot audits, which is why the workstreams converge). Two Overview-specific notes: volatility is high (features appear and disappear across quarters; hold your reads loosely and trend them), and the measurement tie-in runs through Search Console inference — the impression/CTR patterns on Overview-affected queries — since no labeled Overview data exists; the audit’s hand-collected presence data is what makes that inference readable.
Do you know what the machines say when prospects ask about your category?
We’ll run the full mention audit — the matrix across every engine that matters, scored and source-traced — and hand you the coverage report plus the routed action plan: the footprint, content, and correction work that moves the answers.
Get an AI Visibility Audit Strengthen the Local Footprint