What a Contractor AI Search Audit Includes (And How to Read the Report)

First: what an AI search audit is not

Plenty of agencies now slap "AI" on a blue-link SEO report and call it an AI audit. It is not. If the document you get back is a keyword ranking table, a backlink count, and a Core Web Vitals score, you bought a classic SEO audit with a new cover. Those things matter. They do not tell you why ChatGPT skipped your name.

An AI search audit answers one question the old reports cannot: when a homeowner types "who is the best HVAC company in my town" into an answer engine, does the machine mention you, and if not, what is standing in the way. That is a different layer of the web. Google's blue links reward pages. AI answers reward entities: recognized, corroborated businesses the model is confident enough to name out loud.

So an AI audit lives next to your other work, not on top of it. Map pack ranking and "near me" proximity belong to local SEO. Page-one organic rankings and content architecture belong to SEO. This report cross-references both because they feed the AI layer, but its job is the citation layer itself: entity, schema for machine parsing, source pages worth quoting, and the outside signals a model trusts. If a report never mentions those four, it is not an AI search audit.

Here is the tell. A genuine audit shows you a screenshot of what the AI said, verbatim, with your competitor's name in it and yours absent. If nobody ran the prompts and captured the answer, nobody audited your AI visibility. They guessed.

One more distinction worth getting straight before the report lands. This is not a paid-ads report either. When an AI Overview or a chatbot recommends a contractor, that recommendation is earned through the citation layer, not bought like a Local Services Ad or a Google Ads placement. Nobody sells you a slot in ChatGPT's answer. You get named because the machine is confident and corroborated, or you do not get named at all. That is why the audit spends its pages on entity and source signals instead of bid strategy, and why the fixes it recommends compound instead of stopping the day you pause a budget.

Section one: the live prompt test (what the AI actually says)

This is the section you read first, and the one most reports skip. A real audit runs a set of homeowner-style prompts against each major engine and records the raw answer. Not a summary. The actual text, with a timestamp, because these answers change and a screenshot is the receipt.

The prompts mirror how people actually ask. For a plumber that means "who should I call for a burst pipe in Fort Myers," "best emergency plumber near me," "most reviewed drain cleaning company in [city]." For a roofer it is storm-damage and insurance-claim phrasing. The audit tests branded prompts ("tell me about [Your Company]") and unbranded prompts ("best roofer in [city]") separately, because those are two different failures. Being invisible when someone searches your own name means the model does not know you exist as an entity. Being invisible on the unbranded query means it knows you but does not trust you enough to recommend you.

You should see a table like this, one row per prompt per engine:

Read the "Named competitors" column carefully. If three of your competitors show up on the same prompt where you are absent, that is your proof the query has commercial answers and you are being left out of them. That column is the whole business case, in one glance. It also tells you the terms are not too niche to matter: the machine is willing to answer them, it just is not answering them with you.

Watch the "Partial" rows too. A partial mention means the model knows your name but hedges, gets a detail wrong, or lists you last behind competitors. That is a different fix than being absent. Absent means build the entity from the ground up. Partial means the entity exists but the corroboration or the source pages are thin, so the model is not confident enough to lead with you. The prompt table, read row by row, already starts pointing at which of the later sections deserves your attention first.

Which engines get tested, and why the answers differ

Homeowners do not all ask the same machine, so the audit does not test one. It covers the five that decide contractor answers today: ChatGPT, Google's AI Overviews, Perplexity, Gemini, and Copilot. Each pulls from different places, and reading the report means knowing why your name might show up in one and vanish in another.

The short version of how they differ:

Why this matters when you read the report: a split result is a diagnosis, not noise. If AI Overviews names you but ChatGPT does not, your local signals are probably fine and your broader entity or corroboration is thin. If Perplexity skips you, look at whether your site actually gives it something worth citing, because Perplexity will show you the exact sources it chose instead. Reading the engine-by-engine spread tells you which of the four legs is weakest without guessing.

One caution the audit should state plainly: these systems change their sourcing often, sometimes month to month. That is not a reason to chase every update. It is the reason the report is a snapshot with a date on it, and the reason any honest engagement re-runs the same prompts on a cadence. A shop that claims a permanent fix for a moving target is selling you a story.

Section two: entity clarity (does the machine know who you are)

An AI model names businesses it is confident about. Confidence comes from consistency. This section audits whether your identity reads the same everywhere a machine can find it: your website, your Google Business Profile, your directory listings, your social profiles, your license and permit records.

The report checks your NAP (name, address, phone) for exact-match consistency across every source. It checks whether your legal name, your DBA, and the name you market under are reconciled, because a plumber listed as "Borges Plumbing," "Borges Master Plumbing LLC," and "First Class Plumbing" across three sites reads as three shaky half-entities to a model, not one solid one. It checks whether your service area and your trade specialties are stated the same way, or whether one page says "drain cleaning" and another says "sewer and rooter" with nothing tying them together.

You will see an entity confidence read, usually scored, with the specific mismatches listed. Read it as a checklist, not a grade. Every inconsistency is a fixable thing:

• Phone number formatted or transposed differently across listings.
• Business name with and without "LLC," "Inc," or the trade suffix.
• Address abbreviations that do not match ("Ste" vs "Suite," old suite numbers).
• A different primary category on the Google Business Profile than the site claims.
• Missing or contradictory hours, service areas, or founding year.

This is the least glamorous section and often the one that moves the needle most. AI models resolve who you are before they decide whether to recommend you. If the entity is muddy, nothing downstream works, and no amount of good content fixes a business the machine cannot pin down.

Section three: schema and structured data for LLM parsing

Schema is the machine-readable label on your pages. It tells a crawler, in a format it cannot misread, "this is a LocalBusiness, here is the name, the phone, the service area, the services, the reviews." Blue-link SEO uses schema for rich results (star ratings, FAQ dropdowns). The AI layer uses it to parse you cleanly into an entity it can quote. Same tool, different job, and this audit checks it for the second job.

The report inventories the structured data already on your site and flags what is missing, malformed, or contradicting the visible page. Common findings on a contractor site built by a generic shop or an old WordPress theme:

• No LocalBusiness or trade-specific schema at all (a plumber should carry Plumber, a roofer RoofingContractor, an electrician Electrician).
• Schema present but the name, phone, or address inside it does not match the page or the Google Business Profile.
• Service pages with no Service schema, so the machine cannot tell you do slab leaks or standing-seam metal.
• Review markup that is invalid, self-serving, or fabricated, which models increasingly discount or penalize.
• Broken JSON that silently fails to parse, meaning the schema you paid for does nothing.

Read this section as a build-quality report on your site's plumbing. It will list each issue with the page it lives on and the corrected markup that should replace it. One honest note: schema is necessary, not magic. Clean structured data makes you legible to the machine. It does not, by itself, make you the answer. It is one of the four legs, and this audit tells you if it is broken.

Section four: citation-worthy source pages and corroboration

Models name businesses they can source. This section asks whether your website gives an answer engine something quotable, and whether the rest of the web backs up what your site claims. Both halves matter, because a model weighs your own pages against outside corroboration and trusts the ones that agree.

On your own site, the audit looks for pages built to be cited: clear service pages that state what you do and where in plain sentences, an about page that establishes founding year and credentials, direct answers to the questions homeowners actually ask. Thin pages, marketing fluff, and walls of adjectives are not citation fuel. A model will not quote "we are your trusted local experts." It will quote "Family-owned HVAC company serving Lee County since 2004, licensed CAC1234567." Specificity is the whole game.

Off your site, the report inventories third-party corroboration a model trusts: your Google Business Profile, established directories, review platforms, licensing and permit records, local news or association mentions. This is not the backlink-count exercise from old SEO reports. The question is not how many links point at you. It is whether independent sources confirm you are a real, established, specific business, because that corroboration is what tips a model from knowing you exist to being willing to recommend you.

You will see a gap list: the source pages you are missing, the claims on your site that nothing outside corroborates, and the trusted platforms where you are absent or inconsistent. Read it as the roadmap for the content and listing work that follows. This section usually sets the timeline, because building corroboration is the slow part. Competitive terms typically move in 4-9 months, and this is why: you cannot fabricate a track record the web will vouch for. You build it.

How to read the scoring and the priority list

By this point the report has shown you what the AI says, whether it knows you, whether your schema is clean, and whether the web backs you up. The last job is turning findings into an order of operations. A good audit does not hand you forty problems flat. It sequences them.

Expect a prioritized fix list, roughly in this order, because that is the order that actually works:

1. Entity first. Reconcile name, address, phone, and category everywhere. Nothing downstream compounds until the machine agrees on who you are.
2. Schema second. Fix and add structured data so your clean entity is machine-legible on every page.
3. Source pages third. Rebuild thin service pages into quotable, specific answers.
4. Corroboration fourth. Close the gaps on trusted third-party platforms and build the outside signals over time.
5. Re-test. Run the same prompts again on a schedule to confirm movement.

On scoring, treat any single number with suspicion. A "37/100 AI visibility score" is a headline, not a diagnosis. What you want under it is the component reads: entity confidence, schema completeness, source-page quality, corroboration depth, and current mention rate. Those tell you where the real work sits. A shop scoring high on entity but low on corroboration needs a very different plan than one that is muddy at the identity layer.

Last thing to check: does the report tell you how it will be measured again. AI answers move. An audit is a snapshot. Any honest engagement re-runs the prompt test on a cadence so you can watch the "Named you?" column flip from No to Yes on the queries that pay your bills. If there is no re-test plan, you got a photograph when you needed a dashboard.