Someone in your market opened ChatGPT this morning and asked which tool they should use. They got a confident paragraph naming two or three companies. If none of them was you, you will never see it in your analytics, because there was no click to see.
That is the whole problem. Search told you when you lost: an impression with no click, a position that slipped. Answer engines do not. You are either in the paragraph or you are not, and the not is silent.
Here is how to check, without buying anything.
Do it by hand first
Write down the ten questions a buyer actually types before they choose something in your category. Not keywords. Questions, in the words they would use. For a meeting notes app that is "best mac app for meeting notes without a bot joining", not "meeting notes software".
Ask each one in ChatGPT and in Perplexity. Use a fresh session with memory off, or the model will answer partly from what it knows about you rather than from what it knows about the market.
For each answer write down three things:
- Which brands it named, in the order it named them.
- Whether it named you at all.
- Which sources it cited, if the engine shows them.
Twenty answers takes about half an hour, and it is the single most useful half hour available to you right now. Most people discover two things. The first is that they are absent from far more answers than they expected. The second is more interesting: a large share of questions name no brand at all. When we ran this across thirty-six buyer questions for three of our own products, eighteen of them, exactly half, had no brand named in any answer from either engine. Eight more were held by a single company.
A question with no incumbent is the cheapest visibility available to you. There is nobody to displace. You need one page that answers it properly and gets retrieved.
Why the two engines disagree, and why that matters
Run the same question through both and you will often get different names. That is not noise, it is the mechanism.
Perplexity retrieves and cites live sources for nearly every answer. ChatGPT's web search does too, but leans harder on what the model already knows, which is months of training data plus whatever it decided to fetch. So a brand with strong, recent, quotable pages tends to appear in Perplexity first, and a brand with years of accumulated mentions tends to appear in ChatGPT first.
In our own measurements the split was stark: for one product, Perplexity cited it on four questions where ChatGPT never mentioned it. If we had measured only one engine we would have concluded we were invisible, and we would have been wrong by a factor of four.
Measure both. A tool that checks one engine is telling you about one engine.
The number to distrust
Every tool in this category will sell you a sentiment score. Some number saying the AI feels 72% positive about you.
Treat that with suspicion, and here is the specific reason. When researchers compare a language model's sentiment labels against human coders, agreement sits around 0.58 on Cohen's kappa. That is "moderate" on the standard scale, and it is measured on the easy cases. Worse, hosted models are not deterministic: ask the same question twice and you can get two different answers, so the same text can score differently on Tuesday and Thursday without anything changing.
A single sentiment number hides all of that. It looks like a measurement and behaves like an opinion.
What you can trust instead is evidence:
- The sentence. Which exact line in the answer carried the judgement. If a tool cannot show you the sentence, it cannot show you anything.
- The model and the date. An answer from ChatGPT on the 12th is a different fact from an answer from Perplexity today.
- The disagreement. If the judgement was made more than once and the passes disagreed, that should be visible. A borderline answer is a real finding about a borderline answer.
When we built this for ourselves, the rule we settled on was that a judgement which cannot quote text genuinely present in the answer gets thrown away rather than shown. On the first live run that rule threw away three judgements out of three on one answer. The honest output was "no reading", and a tool that reported a confident label there would have been inventing one.
What to do with a gap
You have a list of questions where you are absent. The instinct is to write a blog post per question. That is usually wrong, and expensive.
Look at what the engine cited instead. Three patterns come up:
Nobody is cited. The engine answered from general knowledge. This is the opening. One direct, specific page answering that exact question, with the question as the heading, has very little competition to beat.
A comparison page or a listicle is cited. The engine is grounding on somebody's roundup. Getting into that roundup is usually faster than outranking it.
A competitor's own page is cited. They have written the definitive answer. You need a better one, and "better" here means more concrete: real numbers, a real procedure, the caveats they left out.
In all three cases the page has to be quotable. Answer engines lift sentences. Write in complete, self-contained statements that survive being pulled out of context, put the direct answer near the top, and do not bury the specifics under three paragraphs of preamble.
Then measure again
The reason to do the manual pass first is that it tells you whether any of this matters for your market. If every question in your category already names five entrenched brands and cites their documentation, you have a hard road. If half of them name nobody, you have an open one.
Either way, the number that matters is not today's. It is whether the same thirty-six questions name you more often in eight weeks than they do now. That requires measuring on a schedule, with the answers stored, so the comparison is against a record rather than a memory.
That is what we built Yuzu to do. But do the half hour by hand first. You will know more about your market at the end of it than any dashboard can tell you on day one.