Your AI-Visibility Self-Audit Is Probably Testing the Wrong Question

A placement runs. Someone on the team opens ChatGPT, types the category the article covered, reads the answer, sees no mention of the company, and reports the placement did nothing for AI visibility. That verdict is usually wrong, and the test that produced it is the reason.

The question you ask decides the answer before the placement gets a fair read

Machine Relations measured this directly: 57 live earned-media placements, three buyer questions written per article from what the article actually said, asked once each across the six surfaces that expose a checkable source URL — ChatGPT, Perplexity, Microsoft Copilot, Google's AI Mode (read through DataForSEO's AI Mode SERP endpoint), Google's classic AI Overview, and Gemini — 995 answers, each checked for an exact-URL citation back to the placement. The three question types were the article's own angle with no company name, the same angle plus the company name, and a question about the specific story the article told.

The hit rate moved by a factor of more than two depending only on which of those three questions was asked: questions naming the company or a term the article coined were cited in 17% of answers; the article's own angle with the company unnamed, 11%; the single narrowest version of the story, 7%. A smaller first pass of 12 placements showed the same pattern at the extremes — 6 of 7 placements with a company-naming question were cited at least once, against 3 of 29 generic category questions across the full set. The pair that made it concrete: two placements for the same client, same month, comparable outlet tier. One named the client in the piece; three engines cited it. The other covered the client's category without naming the client; nothing cited it. The variable that moved the outcome was not the outlet. It was whether the question — and the article — named the company.

Full methodology, both runs, and the complete numbers.

What this means for how you check your own placements

A generic-category self-audit — "what's the best [category]" typed into ChatGPT — is close to the worst-performing question shape in this data. It asks the question buyers ask least often about a specific vendor and most often about a category, and a single article competing against everything else written about that category is unlikely to win it regardless of whether the placement is actually being read and cited elsewhere. Judging a placement's AI visibility on that one question is judging it on the test case most likely to return nothing.

This is also the question most self-audit playbooks reach for first. Ahrefs' guide to monitoring ChatGPT brand mentions opens with "ask ChatGPT 'what's the best [your product category]?' right now." Sona's manual-audit walkthrough and MindStudio's brand-visibility checklist both build their prompt sets around category and comparison questions first, with a branded prompt added as one line among twenty or thirty. Rankability's tracking guide and Pepper's brand-visibility audit both distinguish mention rate from citation rate, which is the right instinct, and Search Engine Land's report on Google AI Overviews citing self-promotional listicles while recommending competitors makes the adjacent point that a citation and a favorable outcome are not the same measurement either — but none of the four flags that the category-only prompt most of them lead with is the shape this data shows returns a citation least often. None of that makes the standard playbook wrong to include a category question. It makes a single category question, run alone, the wrong basis for a verdict on one specific placement.

The fix is not a different tool. It's adding the question a buyer who already knows about the company would ask: the company name plus the topic, or the specific claim or story the article made. Those are the questions this test shows placements actually get cited on. A team that only ever runs the generic version will conclude AI visibility isn't working when the placement is being cited — just not on the question they happened to check.

Before spending more on placements, check the test, not just the coverage

None of this says every placement gets cited, or that outlet tier doesn't matter for other reasons. The same test found four in ten placements got no citation at all even on their best-shot question, and one placement that explained a category concept by name was cited on the company-free version of the question too — the coined term did the identifying work a company name usually does. Other audit guides land on a version of the same caution from a different angle: Wellows' LLM-visibility audit notes that roughly two-thirds of the sources an engine cites on a given prompt can churn within two weeks, which is one more reason a single run of any question, generic or branded, is a snapshot rather than a verdict.

What the data does say is that a single generic question is not a fair test of whether a placement worked. A founder deciding whether to renew, expand or cut an earned-media program should run the same three-question shape — company-named, angle-only, story-specific — across the AI engines that answer their buyers' questions before drawing a conclusion from one prompt, and weigh the result against where the rest of the AI-visibility budget is actually going.

In practice that means writing down, per placement, the exact claim the article made and the term (if any) the client or the article coined for it, before running a single prompt. A question built from the article's actual content is a different test than a question built from memory of "the category we're in." The team that skips this step and reruns the same five category prompts every month is measuring category visibility, which is a real and useful number, but a different number from placement-level citation, and the two should not be reported as if they answer the same question to a CFO deciding whether next quarter's earned-media line item survives.

The same discipline applies across the three question types this test tracked, not just the branded one. A story-specific question — the exact claim the article made, not the general topic — cited placements at a lower rate than the branded version but still well above the generic category prompt, which means it is worth keeping in a monthly check even though it is not the highest-yield shape on its own. Running all three, logged separately rather than averaged into one score, is what turns a placement audit into something a budget decision can actually rest on.

FAQ

Why doesn't ChatGPT mention my company when I ask about my category? In this test, generic category questions were the least likely of three question shapes to return a citation to a specific placement — 3 of 29 in the smaller run, 11% of answers in the larger one for the unbranded version of an article's own angle. That's a property of the question, not necessarily proof the placement failed.

What question should I ask to check if a placement got cited? The company name plus the topic the article covered, and a version specific to the exact claim or story in the article. Both outperformed the generic category question in this measurement, with company-naming questions cited in 17% of answers against 7% for the narrowest single-story version.

Does the outlet a placement runs in matter more than the question you ask? Not on its own, in this sample. Two placements for the same client, in comparably sized outlets, in the same month, produced opposite citation outcomes — the one that named the company was cited, the one that didn't was not.