How to tell whether AI engines are naming your business, and what to do about it

You find out by asking. No report will tell you whether AI engines are naming your business, so the only reliable method is to write down the questions your customers would ask, put each one to each engine, and record what comes back. It takes about an hour a month, and it's the hour that tells us most.

Your rankings won't answer it for you. Semrush found that when ChatGPT search cites a webpage, that page ranks in traditional organic positions 21 or beyond for related queries almost 90% of the time (Semrush, July 2025). Ranking well still helps, and Semrush is careful to say so, but it plainly doesn't decide the matter, so you have to go and look at the answers themselves.

The short version.

1. Write ten to twenty prompts a customer would actually type.
2. Run every prompt through every engine separately, in a clean session.
3. Record three states: named, linked, or absent. Note who got cited instead.
4. Repeat monthly, same prompts, same week.
5. Turn the absences into pages.

Step 1: build the prompt panel

A panel is just a fixed list of questions you'll ask every time. Fixed is the important part, because the value comes from comparing this month against last.

Write the prompts as a customer would type them, not as you'd describe your business. Cover four shapes:

• Category and place. "Plumber in Paisley", "accountant for a small limited company in Glasgow".
• The problem, not the service. People describe a symptom before they know what to buy. "My boiler keeps losing pressure", "nobody finds my shop on Google Maps".
• A straight comparison. "Best marketing agencies in Scotland", "cheapest way to get a website built".
• Your own name. This one is a control. If the engine can't describe your business accurately when handed the name, that's a different problem from not being recommended, and it's usually easier to fix.

Ten prompts is enough to learn something. Twenty is the most you'll keep re-running by hand.

Step 2: run each AI engine on its own

Run the panel through ChatGPT, Gemini, Google's AI Mode, Perplexity and Copilot as five separate exercises, and resist the urge to average them into one score.

The reason is bigger than most people assume. Across a random sample of 20,000 prompts, Kevin Indig found that 91.07% of cited URLs turned up in a single engine only. Just 2.37% were shared by all three of ChatGPT, Perplexity and Google's AI Overviews on the same prompt (Growth Memo, May 2026). Two caveats worth carrying: that comparison covers three engines, so it says nothing directly about Gemini or Copilot, and the underlying panel is weighted towards European markets. The direction still holds. Being named by one engine tells you almost nothing about the others, and a single blended visibility score hides exactly the thing you set out to learn.

If you only have time for one, pick the one your customers actually use. Two practical things while you're in there. Use a clean session with no history, because an engine that already knows you will happily hand you back your own company. And set your location honestly, because a local prompt answered from the wrong region is a wasted run.

Step 3: how do you record whether an AI engine named you?

Most people log this as a yes or a no, and lose the information that matters. Five columns, and the whole thing fits in a spreadsheet:

That last column does more work than the other four. It's your competitive set as an engine understands it, and it's often not the competitors you'd have named. Ours turned out to include directories and listicles we'd never thought about, alongside the agencies we expected.

Keep named and linked apart. They move independently, and they call for different work. Being named without a link means the association between your business and the question exists somewhere, but no page of yours was worth quoting. Being linked without much naming usually means one page is carrying you.

Step 4: repeat monthly

Same prompts, same week, same clean session. The picture moves faster than you'd expect, which is the argument for a fixed cadence rather than checking when you happen to remember.

We put our own site through a concerted push and then measured it. Seven prompts, run in the last days of June 2026, came back a straight zero. Thirty days later the same seven came back three, and the useful detail is that they were three different parts of the business rather than three versions of one:

Worth being straight about the method behind those numbers. That run used a search-synthesis proxy rather than us sitting in front of each engine, which is coarser than the procedure above and part of why we're describing the better version here. The direction was not ambiguous either way.

Step 5: read the panel as a content instruction

Our own log split cleanly along one line. All three of the prompts where our name showed up were ones we'd built a page for. The four blanks were topics we sell and had never written down. The workshop result pointed at our strategy page. The PPC result pointed at an article. Nothing pointed anywhere we hadn't deliberately gone.

That turns the panel into a to-do list. Work through it this way.

1. Absent, and a competitor is cited. Read the page that won. If you've nothing equivalent, that's your next page, and it has to be about that question rather than a paragraph inside a service page.
2. Absent, and only directories are cited. On-site content won't win this one. These answers get assembled from listings and third-party mentions, so the work is off-site.
3. Named but not linked. You exist to the engine, but nothing of yours was quotable. The fix is usually a page answering that question with a number and a source up top.
4. Linked, then gone next month. Check whether your page changed or something better arrived. Changing the date isn't an update.

The part we didn't expect: the engine does some of the qualifying

Citations are the easy thing to count, so they're what everyone counts. The thing that actually changed how I think about this arrived through the door.

At the end of July, someone got in touch who had put roughly that second prompt to Gemini a few days earlier. The tidiness of that is satisfying and it's the smaller half of the story. What mattered is that they weren't a stray enquiry. They wanted a workshop. Ours is a workshop. The answer they'd been handed described who it's for and roughly what it does, and the person who turned up had read that description and decided it fitted them.

A ranking doesn't do that. Someone clicking a blue link has seen a title and a line of text and is still working out whether you're relevant. Someone arriving from an answer has already been told what you do, in the engine's own words, in the context of a question they asked in their own. Whatever qualifying normally happens on a first call has partly happened before the call.

One case is one case, and I'm not going to dress it up as a trend. It's still why this sits in the work plan and not in a folder of experiments.

What should you do if an AI engine is describing your business?

Two consequences follow, and they're both slightly uncomfortable.

The engine's description of you is doing sales work, so it has to be right. Ours was accurate because the page it was reading from was specific about who the thing is for. Where your pages are vague about who you serve, the answer will be vague too, and vague descriptions attract people you then have to turn away.

Volume is the wrong measure here. The numbers involved are small. For most businesses this is still a thin slice of visits, and we'd rather say so than sell you an emergency. What it seems to lack in volume it makes up in fit, and there's some evidence beyond our own experience for that: Semrush's July 2025 study found the average visitor from a non-Google AI source was 4.4 times as valuable as the average visit from traditional organic search, measured on conversion rate (Semrush, July 2025).

What this method can't tell you

Two honest limits.

It's a sample, not a measurement. Ten prompts on one day is a spot check. Answers vary between runs of the same prompt, and any single result can be luck in either direction. Trends across months are worth something; a single month isn't.

Your analytics will under-count the result. Some AI referrals arrive labelled and plenty don't, because assistants hand over links without a referrer, people copy addresses into a fresh tab, and a good number of readers see your name in an answer and then search for it, which shows up as brand search. Whatever your analytics says about AI traffic, treat it as a floor.

Find out where you stand this week

Block an hour. Write ten prompts, open the two or three AI engines your customers actually use, and fill in the five-column spreadsheet. You'll finish it knowing more about where you stand than any report has told you, and the absences will hand you a content plan you didn't have that morning.

If you want the background on why any of this works the way it does, we've set out what generative engine optimisation actually is and what the evidence supports. If you'd rather we ran the panel for you and told you what it found, get in touch, and our SEO service covers the foundations that decide whether an engine can read you at all.