How Do I Find My Ghost Citations?

Run 15 to 20 real buyer questions across the AI platforms your customers use, ChatGPT, Google AI Overviews, Gemini, and Perplexity, and log whether your brand got cited, mentioned, both, or neither for each one. Most people get stuck on two things: finding real questions, and reading every answer the same way. Here's how to handle both.

Where do you get 15 to 20 real buyer questions?

Pull them from places where buyers already ask in their own words. Four sources cover almost everything you need: sales calls and support tickets, Google's "People also ask" box, forum threads in your category, and the five question types every buyer moves through before they purchase.

Sales calls and support tickets show what prospects ask before they buy and what customers ask when comparing you to something else. That's real buyer language straight from the conversation, no guessing involved. Google's "People also ask" box on a search for your category shows what Google already expects buyers want to know next. Reddit and forum threads carry the doubts and comparisons buyers write unfiltered, without a sales conversation shaping the wording.

The last source is a checklist you work through: the five question types buyers move through on the way to a purchase. A definition question ("what is X"), a comparison question ("X vs Y"), an alternatives question ("X vs doing it myself"), a use case question ("how do I do X in my situation"), and a buying question ("is X worth it," "what does X cost"). Two or three real questions per type, pulled from your own category, gets you most of the way to 15 or 20.

How should you phrase each question?

Write each question twice: once the way you'd type it quickly into a chat box, and once the way you'd write a longer, more specific message. Short and long versions of the same question can pull different answers from the same AI platform, because the model reads more context into a longer question and narrows its answer.

Take "best CRM for a small agency" as the short version. The long version might be "what CRM would you recommend for a 12-person marketing agency that needs client reporting and email automation." Run both, and log them separately. A brand mentioned in the short answer sometimes disappears entirely once the question gets specific, because the longer prompt gives the model more to work with when it decides who's relevant to mention.

What do you log for each answer?

Log one of four outcomes for every answer, no partial credit: Both (cited and mentioned), Mentioned (mentioned, no citation), Cited only (a ghost citation), or Absent. Note the platform, and paste the exact answer text somewhere you can reread it, since it's easy to miss the split on a first skim.

Fifteen to 20 questions across four platforms, checked by hand and read closely, runs about 15 to 20 minutes per question. That's roughly four to five hours for a single pass, and it's worth repeating periodically since AI answers shift as models update.

What mistakes trip people up when running this by hand?

Two mistakes show up constantly. The first is only searching for your exact legal or trademarked name and missing the informal version buyers read as the same brand: a shortened name, a founder's name, or a parent company used interchangeably with the product. Check for every version buyers use, including the one on your letterhead.

The second is stopping after ChatGPT, since a brand's split on one platform often looks nothing like its split on another. Google's AI Overviews source differently than Perplexity does, and a site invisible on one can be doing fine on the next.

What counts as a good or bad split?

There's no single benchmark, because results vary by category and by how much AI answer volume touches your space. Three patterns show up across almost every category:

• Mostly citations, almost no mentions across ten or more prompts: AI engines trust your content but don't connect it to your brand. That's a recognition gap, and the mention lever is the one to work first.
• Almost nothing at all, neither citations nor mentions: the questions you picked probably aren't ones AI platforms answer from your site. Try different phrasing before you conclude you have no presence.
• A roughly even split between mentioned and cited only: common, and it usually means one lever needs tightening.

What do you do with the results?

Count your split first. If citations outnumber mentions, or the reverse, that split points at which lever to fix first. If you want the full mechanical difference between the two signals first, Mentions vs. Citations breaks it down. CitedScore's AI Visibility Report runs the same method at scale, without the four to five hours of reading.

It asks 20 buyer questions on four platforms, reads and sorts all 80 answers, and asks a set of questions both the quick way and the long way so you can see where the answer changes.