Adding Prompts Should Not Move Your Visibility Rate

On Tuesday you add twenty prompts to a brand. On Wednesday its mention rate is six points lower. Nothing changed in how ChatGPT talks about the brand. What changed is the question you were asking it.

A mention rate is a plain ratio: runs that mention the brand, over all runs, across every prompt the brand tracks. The day a prompt joins the set, its runs join the denominator. A brand rarely adds prompts it already wins. It adds prompts for the gaps it found: the category questions where a competitor is named and it is not. So the new prompts come in at or near zero, and the rate falls on the day they arrive. The chart records a drop, and the drop is an accounting event, not a visibility event.

The reverse happens too. Add ten branded questions, the kind that contain the brand's own name, and the rate climbs for the same reason. Ripenn already excludes prompts that name the brand from the organic rate, so that particular inflation does not reach the chart. But any addition at all moves the denominator, in whichever direction the new prompts happen to lean.

Two ways to read a chart with a changing denominator

Ripenn's brand page handles this two ways, and the default is the lighter one.

Markers. Every day prompts were added gets a dashed vertical line on the trend, labelled "+N prompts added" under the chart. The lines still move, but you can see why. Most of the time this is enough: you read a dip, you see a marker on the same day, you move on. The chart also marks scan-settings changes the same way, so a day you switched from weekly to three scans a week is distinguishable from a day the answers changed.

Fixed prompt set. The switch beside the trend chart holds the set still. With it on, only prompts tracked since before the range opened count toward the mention line, the citation line and the headline figure above them. A 30-day view at the start of November counts the prompts that existed on October 2 and ignores anything added since. The headline says so in its caption: "across 84 prompts tracked since Oct 2" instead of "across 104 prompts scanned". Both numbers are computed in the database under the same denominator rule, so the headline and the chart always describe the same slice, and the tag filter composes with it.

The engine selector composes with it too. Fixed set, ChatGPT only, 90 days, tagged "comparison": that is one slice, and every number on the page is computed over it.

Why it is off by default

The fixed set hides growth, and growth is usually what a team wants to see in the first months. If you went from thirty tracked prompts to a hundred and twenty, a chart that pins the set to the original thirty tells you how those thirty did, not how the brand is doing across the questions it now cares about. Markers keep the whole set and name the discontinuities. That is the better default read.

The fixed set is the right tool for one question: did a specific change to the content move the answers? You published a comparison page on October 10 and added fifteen prompts about that comparison on the same day. Unlocked, the chart shows a dip on October 10 (the fifteen new prompts start at zero) followed by a climb as the page gets cited. Locked to the set that existed on October 1, the chart shows only whether the prompts you were already tracking changed after the page went live. The second read is the before-and-after you actually want for that question, and it is the only read where the dip is not in the way.

So the switch is off every time the page opens, including for a project where you had it on yesterday. The range, the engine and the tag selection are remembered per project; the fixed set is not. Locking the set is a decision about one reading of the chart, and it should be a deliberate one each time, not a setting that quietly stays on and makes a growing brand look flat. The hint beside the switch says what it does, so nobody has to guess from the label.

What it does not fix

A fixed prompt set removes one source of movement: the set itself. It does nothing about the others. AI engines are stochastic; the same prompt on the same day produces different answers across runs, and Ripenn measures a rate across repeated runs because that is the part of the signal that survives the variance. The web the engines read moves. The models are updated without a version bump. A flat denominator makes the remaining movement easier to attribute, not smaller.

Two mechanical limits. Deactivating a prompt cannot be dated, since there is no history row for it, so only additions get a marker. And the fixed set needs a bounded range: "all time" has no "before the window", so the headline ignores the switch there and the chart is left with little or nothing to draw.

The denominator question, more generally

Any visibility tool that reports a rate has this problem, whether or not it surfaces it. If a vendor's dashboard shows a smooth line through a month in which you doubled the prompts you track, ask what the denominator was on each day. Either the line is over a changing set and the vendor is not telling you, or the set is being held still and the vendor is not telling you that either. Both are defensible reads. Not knowing which one you are looking at is not.

The brand page lives under Visibility in Ripenn. The metric definitions behind it, including how the mention and citation rates are composed, are on the metrics explainer.