Right now a buyer is asking ChatGPT a question your company answers. The model names three companies. You have no idea whether one of them is you.
You can find out. It takes an hour, a spreadsheet, and a list of the questions your customers actually ask. Here is how to run that audit yourself, and what to do with the result.
Do not run this audit because search engines are dying
Search engines are not dying. The click died, and that is a far more interesting problem.
In February 2024, Gartner predicted that traditional search volume would fall 25 percent by 2026 because of chatbots. It is 2026. That did not happen. Google still holds roughly 90 percent of the search market and shows no sign of giving it up.
Something else happened instead. SparkToro found that in 2026, fewer than one in three Google searches ends in a click on any result. People still search where they always did. They just get the answer on the page and stop.
That changes the question you should be asking. Not "are people still searching", but "what does the answer say when nobody clicks". If your company is not in that answer, you are invisible at the exact moment a buyer decides who to call. That is what the audit measures.
The audit starts with your customer's questions, not your company name
Here is the mistake almost everyone makes: they type their own company name into ChatGPT, the model describes the company, and everyone feels good. That test measures nothing.
The model knows your name because your site is in its training data, or because it can find you in two seconds. That is not an achievement. A buyer who already knows your name will reach you anyway. The one you care about is the buyer who does not know your name and asks about a problem.
So the audit starts with a list of questions someone would ask who has never heard of you. That is the whole secret, and it is also the hard part. Do not invent them at your desk. Take them from sales calls, from emails, from the first question people ask on a discovery call.
Five prompt types that cover the whole funnel
One prompt gives you one piece of noise. You need five types, because each one hits a different moment in the buying decision.
1. Recommendation. The buyer wants a list of vendors. "Who runs company blogs for SaaS companies?"
2. Comparison. The buyer has narrowed the field and wants the difference. "Content agency or freelancer for a B2B blog, which is actually worth it?"
3. Problem. The buyer does not yet know a service exists that fixes this. This is the most valuable type, because you win here before competitors show up. "Our company blog generates no inquiries, what are we doing wrong?"
4. Verification. The buyer already has your name and is checking whether you are credible. "Is Lumi Zone a decent agency for running a blog?"
5. Local. For any service with a geographic component. "Good employment law firm in Krakow."
Build three variants of each type, phrased in your customer's language, not yours. Your customer does not say "content strategy implementation". They say "nobody reads our blog".
Ask the same prompt in four engines, because each one pulls from different places
Four engines, four different answers to the same question. Testing only ChatGPT gives you a quarter of the picture.
The minimum set is ChatGPT, Perplexity, Gemini, and AI Overviews inside Google itself. Perplexity shows its sources most honestly, so start there if you are short on time. Gemini leans hard on the Google ecosystem, including business profiles and reviews. AI Overviews is the only one on the list that reaches a buyer who has no idea they are using AI at all.
One technical detail that matters: do this logged out, or in a private window. A model that knows your chat history will helpfully suggest your own company and ruin the result.
Record the result in a spreadsheet with four columns
Four columns: prompt, engine, are you in the answer, who is there instead and from what source.
That last column matters more than the first three combined. The first three tell you that you are losing. The fourth tells you why.
| Prompt | Engine | You? | Who instead, and from where |
|---|---|---|---|
| Who runs blogs for SaaS? | Perplexity | no | Agency X, cited from their own case study |
| Blog gets no leads, why? | ChatGPT | no | Agency Y, cited from a listicle on portal Z |
| Is Lumi Zone any good? | Gemini | yes | no competitor in the answer |
Thirty rows is enough. Fifteen prompts times four engines is sixty, but half will repeat and the pattern shows up fast.
Your gap is the list of sources you are missing
Look at that last column. Every source the model cited a competitor from is a specific hole in your presence. That is your backlog.
The pattern is usually the same. The competitor being cited is not winning because they are better. They are winning because they have a piece of content aimed exactly at that question, and you have one aimed near it. Or because they are in an industry listicle you did not know existed. Or because they have thirty Google reviews and you have four.
That column normally produces one of three jobs:
- You are missing an article on that specific question. The most common case and the easiest to fix. You write a piece that answers the prompt directly, structured so a model can lift it. We covered the mechanics in our guide on how to get cited by ChatGPT.
- You have the article, but the model picks a competitor. Usually that means your piece is general and theirs is specific. Or that you have one article and they have twelve around the topic, which is topical authority.
- You are absent from the external sources the model pulls from. Listicles, industry directories, reviews. You will not fix that one with content on your own blog.
We run this audit for every client at the start, and it looks different every time, because the questions a SaaS buyer asks and the questions a law firm's client asks have nothing in common. The framework above is the common denominator we start from. If you would rather see a finished result for your industry than do it by hand, we will run it for free along with a one-month content plan.
Repeat it monthly, because a single run is noise
Models are non-deterministic. The same prompt asked twice gives two different answers, and that is expected behaviour, not a bug.
So a single audit tells you where you are, but only a series tells you whether you are moving. Once a month, same spreadsheet, same prompts. After three months you have a trend, and you know whether your content is actually entering answers or just sitting on your blog.
This is also the only honest way to report on AEO right now. A Google position is measurable with a tool. Presence inside a model's answer has no good cheap tool yet, so you measure it by hand and with patience.
Lumi Zone runs company blogs for SEO and AEO for SaaS companies, niche e-commerce brands, and B2B services. We write, edit, and publish on your CMS, and it costs you 15 minutes a month. See how we work.