Your buyers are asking ChatGPT which supplier to use. You don’t get to see the answer.
That’s the whole problem. Search used to leave a trail. Impressions, positions, queries, a report you could argue with. An AI answer leaves nothing. Someone asked, got three names, picked one. If you weren’t among the three, nothing in your analytics will ever tell you.
You can check by hand. It takes an afternoon and it’s worth doing before you spend money on anyone’s opinion, including ours.
If you want the vocabulary underneath it first, answer visibility sets out what’s being measured here and why two honest readings of the same brand can disagree completely.
Ask the buying question, not your own name
Most people test this wrong. They type their company name into ChatGPT, read the paragraph that comes back, then decide they’re visible.
That proves nothing. You handed the model the answer. Your buyer won’t.
Write down the five questions someone asks when they have a problem and no shortlist. They fall into four shapes:
- The category question. “Best project management software for a small agency.”
- The comparison question. “X versus Y, which is better for Z.”
- The alternatives question. “Alternatives to [the biggest name in your market].”
- The situation question. A sentence describing their circumstances, ending in “what should I use”.
Add one local variant if geography matters to your sale. “Who should I use for commercial fit-outs in Manchester.”
If your name doesn’t appear in the prompt, the test is honest.
Run each prompt more than once
These models are probabilistic. Ask the same question twice and you can get two different shortlists. One run tells you almost nothing.
Run each prompt three times, in a fresh chat, logged out or in a temporary session. If you’re signed into an account that has been discussing your own company for a year, the model has learned who you are and the result is flattering fiction.
Then repeat across the surfaces your buyers actually use. ChatGPT, Perplexity, Google’s AI Overviews, Gemini, Claude. They retrieve differently and they disagree with each other constantly. Being named in one is not being named.
Record four things
Not a transcript. Four columns:
| Column | What goes in it |
|---|---|
| Named | Yes or no. No partial credit. |
| Position | First, middle, last, or a passing mention at the end. |
| Who else | Every competitor named, in order. |
| Cited | The sources the answer leaned on. |
Fifteen prompts across five surfaces is seventy-five rows. It’s dull. Do it anyway, because the pattern only shows up at that size. One lucky mention will otherwise convince you everything’s fine.
Read the citations, not the answer
This is the step everyone skips. It’s the only one that tells you what to do next.
The answer is the output. The citations are the mechanism.
Perplexity and AI Overviews show their sources. ChatGPT shows them when it searches the web rather than answering from training. Open them. Across seventy-five rows you’ll find the same small set of pages doing most of the work: two or three directories, a comparison post someone wrote in 2024, a trade publication, a Reddit thread, one competitor’s own page.
That set is the battleground. Not the ranking. Not your domain rating. Those pages.
An audit tells you everything. The citation list tells you what to fix first.
Three reasons you’re missing, each needing a different fix
Once you have the pattern, you’re in one of three situations.
Nothing answers the question. Nobody has written a good answer to the prompt your buyers are asking, including you. The model is assembling something from fragments. This is the cheapest position to be in and the one people are slowest to notice, because it looks like everyone is losing rather than you specifically.
Something exists and it isn’t citable. You’ve covered the topic. The page is a brochure. It asserts rather than answers. It carries no figure, no method and no date a model can anchor to. Models cite text that reads like evidence. Marketing copy about how passionate your team is doesn’t qualify.
The sources exist and you’re not in them. The directory that gets cited doesn’t list you. The comparison post covers four competitors and stops. The Reddit thread names everyone else. Nothing you publish on your own site fixes this one, which is why it’s usually the expensive answer. It’s also usually the true one.
Three problems, three different responses. Guessing which you have costs more than checking.
What this method won’t tell you
It’s a sample, taken once. Run it again in six weeks and it will have moved, because the underlying index moved and the model changed.
It also won’t tell you which of those prompts anyone actually types. Fifteen questions you invented at your desk are a hypothesis about your market, not a measurement of it. Some of them are worth thousands a month. Some are worth nothing. This method treats them all the same. Every other way of measuring this has its own blind spot: the five methods in commercial use and what each one misses.
And it says nothing about what happens after the click. You can be named first in every answer and still lose the sale on a page that doesn’t convert.
It tells you where you stand. Not what it’s costing you.
That gap is the job we do: the same read, run properly across every surface, weighted by what each question is worth, with the SEO and paid problems underneath it in one ranked list. What that costs, against every published UK price we could verify, is set out in what an AI search audit should cost. But do the afternoon version first. You’ll learn more from seventy-five rows of your own data than from anyone’s sales pitch, ours included.
Start today
Pick your five prompts. Open a private window. Run the first one three times.
You’ll know inside ten minutes whether you have a problem.