Home Services Is the Hardest Place to Prove AI Visibility Works. That's Exactly Why I Test It Here.
Most published AI-visibility advice quietly assumes a SaaS company: a signup form, a blog with an audience, a product people write about, and a conversion that happens in a browser where analytics can see it.
Now try it on an appliance repair company operating across six metros. Four things break at once:
- The conversion is a phone call, so the moment of truth is invisible to every web analytics tool.
- The job is created by hand in a field-service CRM by a dispatcher, with no field for where the customer came from.
- The service area is a radius, not an address — often with the address deliberately hidden, which costs map-pack proximity.
- Nobody links to a repair company. There is no natural editorial reason to cite one, so domain authority stays near zero for years.
I've spent the last year running this work on sites with all four properties. It is the hardest configuration I know of for proving anything, which is precisely why the results transfer: a method that survives here isn't leaning on a friendly industry.
What does AI actually do on a home-services site?
Measured, not estimated. Server-side logs on two live sites in this industry, last week of August:
| Site A (appliance repair, multi-metro) | Site B (home services, different market) | |
|---|---|---|
| Window | 8 days | 5 days |
| Requests from AI agents | 3,498 | 3,178 |
| Distinct agents | 14 | 10 |
| Live fetches to answer a person | 425 | 104 |
| Distinct URLs touched | 1,703 | 1,290 |
| Distinct URLs pulled into answers | 207 | 53 |
So: AI engages with these sites constantly — roughly fifty live fetches a day on site A. The industry is not being skipped.
Now the same site through Google Analytics, one calendar month, 1 August to 1 September: 23 sessions from all AI platforms combined, 0.47% of traffic. Fourteen from ChatGPT, four from Perplexity, three from Gemini, two from Claude.
That contradiction is the entire problem in one line. The machines are working the site every hour; the browser almost never opens. Because in this industry the answer the customer needs is the phone number, and an assistant delivers it inside the chat. I've written up both instruments and the roughly seventy-fold gap between them in detail.
Which pages get cited when the whole site is service pages?
Here's the finding that changed how I build for this industry, and it replicated on both sites.
The URLs assistants pulled live were overwhelmingly data pages. On site A the top two content URLs were a brand-reliability page and a statistics page built from the client's own job records. On site B the top two were both from a /resources/ section — a brand failure-rate page and a reliability page — ranking above every location page on the domain.
What was fetched almost never: service pages, location pages, "why choose us". Those made up the large majority of the 1,703 crawled URLs on site A.
This is not a mystery once stated plainly. An assistant answering "how long should a refrigerator last" needs a source for a factual claim. A page whose entire purpose is a claim with a number attached is the cheapest source available. A page whose purpose is "call us today in Tampa" is not a source for anything.
The uncomfortable implication for this industry: a site made entirely of city × service pages has almost no citable surface, no matter how well written those pages are. That's why so many operators conclude AI visibility doesn't apply to them. They measured the wrong asset class.
Where does a repair company get proprietary data?
It already has it, and this is the part I find most people don't believe until they look.
A field-service business with a few years of history is sitting on the only dataset of its kind in its industry: completed jobs with what broke, which brand, which part, what it cost, how long it took, how often the customer declined the repair. Nobody has published that. Consumer magazines publish surveys — reported reliability. A repair company has observed failure.
On one client's site, statistics pages built from that record became simultaneously the most-fetched content URLs by assistants and the largest single source of referring domains on the domain — including a citation from a direct competitor using it as a source, and one from a legal-explainer site that needed a technical reference for a class action.
The rules that make it work, learned partly by getting them wrong:
Every figure carries its base. Not "most customers decline expensive repairs" but a percentage across a stated number of service visits. A number without a denominator is unquotable, and a number with one is liftable.
Frame it as what you service, not as what's reliable. "Most-serviced brand" is an observation about your job mix. "Least reliable brand" is a claim about a manufacturer, and it's a defamation risk that also isn't supported by your data — your data has a selection bias the size of your marketing.
Never publish numbers about your own quality. Repeat-visit rates, callback rates. They're the most interesting figures you have and the ones that will be used against you.
Check the denominator against your own earlier pages. I once published two pages where the same metric appeared with different numbers because I'd used different denominators. That's worse than a duplicate: an author cites one page, a reader checks the other, and both stop being a source.
Does the address problem interact with any of this?
Yes, and it's the one place where AI visibility is genuinely easier than the classic channel.
A service-area business that hides its address loses map-pack proximity, and recovering it takes many months of review and authority work — the specifics are on hidden address and map pack rankings. Meanwhile an assistant composing an answer about who repairs refrigerators in a given city has no proximity model at all. It has sources. If your data page is the source, geography is not the gate it is in the map pack.
That doesn't replace local SEO. The recovery path for a hidden address, and how long it actually takes, is on hidden address and map pack rankings. It means the two channels fail for different reasons, so a business locked out of one is not automatically locked out of the other.
What breaks if you skip the attribution part?
Everything, and this is where I've watched the most money get wasted in this industry.
The chain that has to hold: the assistant names the business → the person calls → a dispatcher creates the job by hand → a technician completes it → an invoice carries an amount. Nothing in that sequence records where the customer came from unless somebody builds it, and no analytics product ships with it because two of the five steps happen offline.
Without it you're back to counting mentions, which is how the industry ends up buying reports that show growth while nobody can name a job that came from it. With it you can answer the only question the owner asked. On the client site above the chain went live in late August and the first lead arrived carrying its own origin the same day; the build is described under full-cycle attribution. The revenue number itself I'll publish when the window is long enough to mean something — including if it turns out to be zero.
What would I do first on an HVAC or plumbing site?
In this order, because each step makes the next one measurable:
- Put server-side AI logging on, split by purpose. One afternoon, and it ends the guessing about whether AI touches your site.
- List the URLs that got live fetches. That's your citable surface today. Expect it to be short and to contain no service pages.
- Publish one data page from your own job records, every figure with its base, linked down to the service page it supports.
- Instrument the phone, so a call carries a source into the CRM. Without this the rest is a hobby.
- Only then buy a presence-tracking tool, if you still want one — and read what its share-of-voice number is actually made of first.
The reason I keep testing in this vertical rather than an easier one: if a method produces a traceable booked job in a business where the money moves over the phone and the CRM has no source field, it will survive anywhere. The vertical is a stress test, not a specialty.
I do this work end to end for home-service operators — the citable data layer, the server-side measurement, and the attribution chain that reaches the invoice. Engagement types and prices are public on pricing, and the citation side is described under AI citation visibility. If you run a multi-location service business and want to know what AI is doing on your site right now, let's talk.
Related: