Shape the Work, Never the Ruler

I changed one sentence and got a different answer about the same company.

The company was Scrolling the Heavens, a small Christian astronomy site. On questions phrased the way an app store shopper types, "is it worth paying for an astronomy app," the AI assistants we tested never named it. Not once in ninety-three answers. Change the question to "Christian astronomy website with an interactive star map," and Perplexity named it first, the same afternoon, on the same engines, with nothing on the site changing.

That is the whole game, and it is also the whole trap.

Phrasing decides the shape of the answer

Questions come two ways, and the engines treat them completely differently. A recommendation question, "who does this near me" or "best company for this," produces a list of named businesses. That is winnable by a business that shows up in the places engines draw candidates from. An explainer question, "how much does this cost" or "how does this work," produces prose that names almost nobody, and the only thing winnable there is being the page the engine quotes.

Same topic. Two wordings. A score of zero or a score of a hundred. If you control the wording, you control the number.

Where the honesty breaks

An agency can choose the questions its client already wins, run them, and report the result as market visibility. Nothing in that report has to be false. The number is still close to meaningless, because the questions were selected after the answer was known. It is a ruler bent to fit the thing it measures.

You can spot it three ways.

The set never contains the money question the client loses, the one where the top competitor appears in every answer. That question is the whole strategy, and it is the first thing removed from a flattering scorecard.

The set changes quietly when the score drops, and the change is called a refresh.

The headline is a percentage with no count behind it. A percentage from a small base can say almost anything, in either direction, and nobody can check it without the raw numbers.

What honest measurement looks like

Four rules, and they are not complicated. Every question on the scorecard carries evidence that a real buyer asks it, from Google autocomplete, Search Console query data, a community thread, or a client's own inbox. No evidence, no place on the board.

The core set stays frozen month to month. A trend line only means something if the questions underneath it hold still.

The losing questions stay in the report. The gap between the client and the competitor on the question that matters is the strategy, not an embarrassment to be edited out.

The count sits next to the percentage. "Named in 4 of 60 answers" is a fact anyone can verify. "Six percent" is a mood.

How we handle the temptation

We rephrased one client's set this month because the old questions did not match how buyers actually ask. The score moved from 19.2 to 20.5 percent without a single new mention. That is a real improvement in what the client can reasonably win, and it is not a market result, so it does not get reported as one. Phrasing experiments live in a separate ring and get reported beside the score, never blended into it.

Five questions to ask anyone selling you AI visibility

Where was a real buyer seen asking each of these questions? Which questions does my biggest competitor win, and are those in my report? How much of the set has changed since last month, and why? Can I see the count, not just the percentage? And if your measurements showed my category was hopeless, would you tell me?

If those answers are vague, the number you are being sold is a mirror, not a window.

---

E320 Consulting runs AI visibility measurement for local and regional businesses from Ruskin, Florida. We publish the questions, the counts, and the ones our clients lose. Book a 20 minute call if you want to see what your own category looks like.