First, know exactly what you are testing for
Before you type a single prompt, get clear on what a pass looks like, because AI search does not grade the way Google does. On Google you scroll a list and count where you land. Ranking #6 is a real, findable position. An AI answer does not hand back a list. It names one, two, maybe three businesses and stops. There is no page two and no #6. You are either in the answer or you are not in the room.
So the test has one primary question: when a homeowner asks an AI who to call, does it say your name? Everything else is secondary detail that helps you understand the result. You are checking four things on every prompt, in this order:
- Named or not. Does your business appear by name in the answer? This is the pass/fail line.
- Who got named instead. If it is not you, which competitors did the model reach for? Those are the shops the AI understands better than yours right now.
- The sources cited. Perplexity and AI Overviews show little citation links. Those are the pages that decided the answer. If none point at you, you were never a candidate.
- Facts correct or wrong. If you are named, did it get your service area, trade, and phone right, or does it think you only do one thing or work in the wrong town?
One more setup note that changes results more than people expect: location. AI engines lean on where they think you are asking from. Run the test on a device and network in your actual service area, not on a VPN or a phone parked at your kid's out-of-state college. If you want to see what a homeowner two towns over gets, say the town by name in the prompt ("in [that town]") so location is stated, not guessed. Test both your home base and the edges of your service area. The answer often changes at the edges, and the edges are where you are quietly losing work you assumed you had.
The five tools to run, and how they differ
"AI search" is not one box. It is at least five surfaces a homeowner might land on, and they behave differently enough that skipping any one leaves a blind spot. Here is what each is and where to run it.
| Engine | Where to run it | What makes it different |
|---|---|---|
| ChatGPT | chatgpt.com or the app | The one homeowners name out loud. Turn on search/browse so it pulls live results, not just training memory. |
| Google AI Overviews | A normal Google search | The AI box above the blue links. Reaches the most people because it shows up mid-search, unasked. |
| Perplexity | perplexity.ai | Shows its citations openly, so it is the best tool for seeing which pages decided the answer. |
| Google Gemini | gemini.google.com | Google's standalone assistant. Pulls heavily on Google's own understanding of your business. |
| Microsoft Copilot | copilot.microsoft.com | Built on the Bing index, so a different source pool than the Google-fed engines. Catches gaps the others hide. |
Why run all five instead of just ChatGPT. The engines pull from different places. ChatGPT and Perplexity lean on live web search plus their own model; Gemini and AI Overviews lean on Google's index and Google's read of your business; Copilot leans on Bing. You can be named cleanly in one and completely absent from another, and that split is itself useful information. If Google-fed engines name you but ChatGPT does not, your Google presence is solid and your broader entity and citation footprint is thin. If it is the reverse, something in your Google understanding is muddy.
A practical setup tip. Open all five in separate tabs and run the same prompt across them back to back before moving to the next prompt. Comparing five answers to the same question is far more revealing than comparing one engine's answers to five different questions. And use a fresh chat or incognito window for each new prompt. AI chats remember what you said earlier in the conversation, and that memory will quietly contaminate the next answer if you keep going in the same thread.
One more thing worth checking on each engine before you start: make sure live search is on where the setting exists. ChatGPT in particular can answer from training memory alone, which is stale and can name businesses that closed or miss ones that opened last year. When the model is actually searching the web as it answers, you are testing the real, current answer a homeowner gets today, not a memory from a year ago.
The prompts to run (steal these)
The single biggest mistake in a DIY test is asking about yourself. "Tell me about [My Company]" proves nothing. Of course the AI can describe you when you hand it your own name. The homeowner does not know your name yet. That is the entire point. Your prompts must sound like a person who has a problem and no idea who to call.
Run these, swapping in your trade, city, and the jobs you actually want. Do each one across all five engines.
- The core "who" question. "Who is a good [trade] in [your city]?" This is the money prompt. It is the one a homeowner types verbatim.
- The urgent job. "Who should I call for [your most urgent, highest-value job] near me?" For a plumber, a burst pipe. For an electrician, a panel that is sparking. For HVAC, a dead AC in July. Urgency is where AI answers get short and decisive, so being named here matters most.
- The comparison ask. "What are the best [trade] companies in [your city]?" This one often returns a list of two or three, which shows you the exact set you are competing against inside the answer.
- The specific-service ask. "Who does [a specific service you want more of] in [your city]?" This tests whether the AI understands your specialties or has you filed under one generic label.
- The edge-of-area ask. "[Trade] in [a town at the edge of your service area]?" This finds the towns where you think you show up but do not.
- The trust ask. "Who is a licensed and reputable [trade] near [your city]?" This surfaces whether your license and reputation signals are legible to the model.
As you run each one, resist the urge to argue with the AI or feed it hints. If it does not name you on the plain homeowner prompt, that is the result. Nudging it into saying your name proves nothing except that it can read a name you gave it. Log the clean, unassisted answer, because that is the one a real homeowner sees.
How to log it so the test is worth something
A test you do not record is a test you cannot repeat, and one loop of prompts is a snapshot, not a signal. AI answers shift, and they differ by engine, so the value comes from a simple grid you can rerun in ninety days to see whether anything moved. A basic spreadsheet does the whole job. No tool to buy.
Build one row per prompt-and-engine pair. For six prompts across five engines that is thirty rows, which sounds like a lot and takes about twenty minutes to fill. Columns:
| Column | What to write |
|---|---|
| Date | The day you ran it, so you can compare later |
| Engine | ChatGPT, Gemini, Perplexity, Copilot, or AI Overviews |
| Prompt | The exact question, word for word |
| Named? | Yes / No: were you in the answer |
| Who else | The competitors it named instead of or alongside you |
| Sources cited | The citation links shown (Perplexity and AI Overviews show these plainly) |
| Facts right? | If named, did it get your area, trade, and details correct |
Read the finished grid for patterns, not single rows. A few honest reads: if the same two competitors keep showing up, those are the shops the answer layer currently belongs to, and their sites and citations are worth a look. If Google-fed engines (Gemini, AI Overviews) name you but ChatGPT and Perplexity do not, your Google footprint is doing its job while the wider web has too little to corroborate you. If you are named but the facts are wrong, that is often worse than being absent, because a homeowner gets pointed at you and then told you do not serve their town.
Take a screenshot of any answer that names a competitor or gets your facts wrong. Answers change week to week, and a dated screenshot is the receipt that proves what the machine actually said on the day you looked. Keep the grid. In ninety days you rerun the same prompts, and the before-and-after tells you the truth about whether the channel is moving for you.
Reading your results honestly (a scorecard)
Once the grid is full, resist the two bad reactions: panic at every "No," or relief at a single "Yes." One prompt on one engine is noise. The pattern across the grid is the signal. Here is a plain way to score where you actually stand.
| What the grid shows | What it means |
|---|---|
| Named on most core prompts across most engines, facts correct | You are in good shape. Keep the entity and citations current and rerun quarterly so you do not slide. |
| Named on Google-fed engines only (Gemini, AI Overviews) | Your Google presence is solid; your broader entity and off-site corroboration are thin. That gap is the work. |
| Named only when you say your own name | The model can read you but cannot recommend you. That is an entity and citation problem, not a content problem. |
| Named but facts wrong (wrong area, wrong trade) | Your entity is muddy. The AI has you filed incorrectly, and it is actively misdirecting homeowners. |
| Not named anywhere on any homeowner prompt | You are invisible in the answer layer. Not rare, and fixable, but it is real work, not a setting. |
Notice what the test does not tell you. It shows you the score, not the reason. Two roofers can both come back "not named" for completely different underlying causes: one has a name the model cannot pin down, the other has a clear name but nothing off-site vouching for it. The prompt test cannot see the difference. It is a thermometer, not an X-ray. It tells you that you have a fever, not what is causing it.
Keep this test in its lane, too. It measures whether an AI names you inside an answer. It does not measure your Google ranking, your Maps 3-pack position, or your paid ads, and a strong or weak result here does not translate to any of those. Those are separate layers with separate mechanics. If your real worry after this test is "why am I not #1 on the map" or "why am I not on page one of Google," those are honest questions, but they belong to a different diagnosis. This test answers one thing: does the AI say your name.
Where the free test ends and a real audit begins
The DIY test is genuinely useful and we tell contractors to run it. It costs nothing, it takes half an hour, and it turns a nagging "I heard ChatGPT never mentions us" into a documented grid you can act on. Run it. But be honest about its ceiling, because that ceiling is where owners waste months guessing.
Here is what the free test cannot do. It cannot tell you why you were left out. It cannot separate an entity problem from a schema problem from a thin-citation problem, and those three call for completely different fixes. It cannot check the plumbing under your site, whether your structured data is built so a model can parse and cite it, whether your name, address, and phone are consistent across the sources AI trusts, whether the pages an answer engine would want to cite even exist on your site. And it cannot tell you which fix moves the needle first, which is the difference between results and busywork.
A real audit does the diagnostic the thermometer cannot. It runs a wider, structured battery of homeowner prompts across every engine, then goes under the hood to read your entity clarity, your schema, and your off-site corroboration, and maps the specific gaps in the order they need fixing. That is the X-ray behind the fever. It is the part that turns "we are not named" into "here is exactly why, and here is what to do about it, first."
On timeline, keep expectations straight. This is not a switch. Entity and schema fixes can register in weeks. Getting named for competitive terms, where a well-known local shop currently owns the answer, runs closer to the 4-to-9-month arc that competitive organic terms take, and AI models refresh on their own cadence on top of that. Anyone promising a ChatGPT mention next week is selling you something. The contractors getting an edge are the ones running the test now, reading it honestly, and doing the work before the channel gets crowded.