Most teams discover how AI systems describe them by accident: a prospect repeats something odd on a call, or a competitor turns up in an answer where they expected their own name. The check itself takes an afternoon. What makes it useful is knowing what to compare the answers against.
Start from the three realities
The reason a single ChatGPT query tells you almost nothing is that an answer on its own has no reference point. It is only meaningful next to two other things: what your company actually is, and what your company says about itself.
Those are the three realities the Digital Reality Gap framework separates:
- Business Reality. The products, the customers, the results. Everything true about the company today.
- Stated Reality. Your website, your positioning, your messaging. What you have chosen to publish.
- Internet Reality. What the systems and the market actually find.
An AI answer is a sample of Internet Reality. Read on its own it is trivia. Read against the other two it tells you exactly where the distance sits, and the distance is what you can act on.
Write the questions your buyers ask
Not your brand name. Asking a model about a company you already named tells you how well it can retrieve a page it has been handed. That is the easy case.
The useful questions are the ones a buyer asks before they know you exist:
- The category question. “Best X for Y.”
- The problem question, phrased the way someone with the problem would phrase it, not the way your category does.
- The comparison question. “X vs Y” for the two competitors you actually lose to.
- The qualifying question. “Who should I use if I need Z.”
- The disqualifying question. “What are the downsides of X.”
Ten to fifteen is enough. Write them before you open anything, or you will drift into questions you already know the answer to.
Ask across systems, and ask more than once
Run the set against ChatGPT, Gemini, Perplexity, Copilot and Google’s AI Overviews. They retrieve differently and they will disagree, and the disagreement is informative: a company that is well described in one system and absent from another usually has a corroboration problem rather than a content problem.
Two practical notes. Use a logged-out or temporary session, or you are measuring your own history. And ask each question more than once, on different days. Answers vary between runs, so a single response is an anecdote.
Record four things, not one
For each question, note:
- Whether you appear at all. Absence is the finding people most often skip past, and it is usually the largest one.
- How you are described. In the system’s words, not yours. This is where old positioning surfaces, years after the website changed.
- Who appears instead. Named competitors, in order.
- Which sources are cited. This is the part almost nobody records, and it is the most actionable thing on the page.
That fourth column is the map. It tells you which pages, directories, reviews and third-party write-ups the systems are actually reading about your category. Those sources are the surface you can influence, and most of them are not your website.
Read the answers against Business Reality
Now do the comparison the exercise exists for. Take each description and ask whether it is true, current, and complete.
You are sorting findings into four kinds, because each has a different fix:
- Absence. You are not in the answer. Usually a corroboration or coverage problem: nothing findable connects your company to the question.
- Staleness. The description was accurate two years ago. Your Stated Reality moved and Internet Reality has not caught up.
- Distortion. The description is assembled from sources you did not choose, and it frames you in a competitor’s terms.
- Vagueness. You appear, and the summary could describe any of six companies. This one reads as harmless and is not: an answer engine has to state what you are in a sentence, and if your material never does, it will guess.
What to do with it
The temptation is to fix whatever looks worst. The better move is to fix whatever the most sources depend on. If four of your fifteen questions cite the same industry directory and it describes you wrongly, that single entry is worth more than a month of publishing.
Beyond that, the fixes divide along familiar lines. Absence and distortion are usually evidence problems, solved by getting corroborating material into the places the systems already read. Staleness and vagueness are usually clarity problems, solved on your own site by stating plainly what you do, for whom, and what makes you different, in a form a machine can extract without inference.
Run the set again in a quarter. Individual answers move around; the pattern does not, and the pattern is what you are measuring.