Compare runs over time
MarqIQ compares the latest visibility check with the one immediately before it, and shows the change only when the two asked the same questions on the same engines and both got every answer back. This article covers how to run a check that stays comparable, where the change is shown, and what to do when it is withheld. If you have only one run so far, start with See whether AI assistants recommend you.
Before you start
Comparison is between two runs, not across a chart. MarqIQ does not draw a history graph of your visibility, and no screen lists past runs: the Overview tab shows the latest finished run and its change against the run before it.
Tip
Decide your prompt set before you take a baseline. Changing a question is what breaks a trend, and a set you are happy with for a quarter is worth more than a perfect one you keep editing.
The four conditions for a change to be shown
All four must hold. When any one fails, MarqIQ names the one that failed rather than printing a change that would measure something else.
| Condition | What breaks it |
|---|---|
| Both runs recorded which prompts they asked | A run from before prompt sets were recorded. Nothing can be done about it; the next two runs compare. |
| The same scored prompts | A question added, removed or switched on or off between the runs. Aided prompts are left out of the comparison, as they are left out of every score. |
| The same engines | A key connected or disconnected between the runs. The line names which engine was added or removed. |
| Both runs got every answer back | Either run is partial. A rate over the answers that came back is not the same measurement as a rate over all of them. |
A run too damaged to be quoted at all fails the fourth condition as "one of the two runs lost too many answers to quote".
Note
The change is in whole points of the visibility rate: "visibility up 7 points", not 7%. Both runs' rates are the ones frozen when each finished, so nothing about a past run is recomputed when you change your prompts today.
Run a check that stays comparable
Open AI Answer Visibility and stay on the Overview tab.
Click Run visibility check.
Read the panel headed "Before this run: which prompts should it ask?"
It states what this run will ask — active prompts, how many are aided, how many engines and the total calls — and then whether it can be compared with the last one.
Read the comparability sentence in that panel before you choose.
- "The same scored prompts as the run of …, so the two can be compared if both get every answer back" — the set is unchanged.
- "The scored prompts differ from the run of … (2 added, 1 removed), so this run will not be compared with it" — somebody has edited the set since.
- "The run of … did not record which prompts it asked, so this run will not be compared with it" — an older run, and nothing can change that.
- "This is the first run, so there is nothing to compare it with yet."
Click Keep the same prompts and run.
When the set already differs from the last run's, this button reads Run with the current prompts (not compared with the last run) instead, so the click never claims something the panel has denied.
Choosing Update the prompts first opens the Prompts tab and you start the run from there with Run the check with these prompts. That run is recorded as one taken after the prompts were updated.
Leave the tab open until every call is made.
A run that stops part-way is unfinished, and an unfinished run is never compared with anything — no screen quotes it at all.
Where the comparison appears
Under the four tiles, below the provenance line that gives the run's date, prompt count and health.
- Comparable: "Compared with the run of 12 September 2026, on the same 32 scored prompts and engines: visibility up 7 points."
- Not comparable: "Not compared with the run of 12 September 2026:" followed by every reason, including the counts added and removed, and the engines by name.
- The provenance line also says which choice was made before the run — kept as they stood, run on the same scored prompts as the run before, or run after the prompts were updated.
Take care
A run-over-run change is withheld unless both runs got every answer back, so a partial run shows no change even when it shows a score. That is deliberate: with one engine quota-stopped after six prompts, a move in the rate is partly a move in which engine was missing.
Reading a trend one prompt at a time
The Prompts tab carries the per-question view, which is often where a change is legible before the headline moves.
Click Prompts.
Read the Answers from its last run column.
Each engine gets a chip for the prompt: a tick for an answer naming you, a dash for an answer that did not, and a stated reason where the call failed, was not asked or was not measured.
Check the line under the chips.
It reads Latest run or From an earlier run, with the date. A prompt switched off for a while, or added since, carries answers from an older run, and the label says so rather than passing them off as current.
Hover a chip to see the model that answered.
Where the call failed, was not asked or was not measured, the hover carries the recorded reason beside the model. A tick or a dash has no reason beyond the answer itself.
What else can move a number without your visibility moving
- A model change. The By engine card flags a model that differs from the previous run — "was gpt-4o-mini last run" — because the trend then compares two different models. A run that timed out and fell back mid-way is marked "models changed during this run".
- Removing a prompt. Past runs keep the score they froze, but the breakdowns are rebuilt from the answers still stored, so a removed prompt takes its answers with it. The overview then says how many of that run's answers are still counted.
- Age. Once the latest run is more than 30 days old, a banner gives its date and age, says AI answers move quickly, and offers Run again.
Tip
Read By engine alongside the headline before you report a trend. One engine at zero and another at half average out to a figure neither engine produced, and the two need different work.
If it did not work
No change is shown although two checks have finished
Why: One of the four conditions failed. The line under the tiles names which.
Fix: Read that line. A changed prompt set is the commonest cause, and it is fixed by choosing Keep the same prompts and run on the panel before each check.
The panel says the prompts differ and nobody changed them
Why: A switch in the Tracked column, a suggestion added on the Prompts tab, or a prompt removed during the last run all change the scored set. A teammate's edit counts as much as yours.
Fix: Open Prompts and compare against what you expect to track. The panel gives the counts added and removed, so a difference of one is usually a single toggle.
The change is withheld because not every call answered
Why: One of the two runs is partial. The line gives how much of each came back — "this run 84 of 96, that run 96 of 96".
Fix: Fix the engine named on the amber banner and run again. Why one AI engine returned nothing, or half your prompts were not asked covers each reason and says whether to run again now or fix billing first.
The score has not changed although an engine clearly failed
Why: A visibility figure is frozen when its run finishes and is never recomputed. You are reading the last finished run, not the one that failed.
Fix: Check the date on the provenance line. An unfinished or discarded run never replaces the figure above it.
You want last quarter's numbers and cannot find them
Why: The Overview tab shows the latest finished run and its comparison with the run before it. There is no list of past runs and no history chart.
Fix: Record the headline figure, the date and the health line each time you run a check, which is what the provenance line is for. Every figure it quotes states the count it rests on, so a note taken from it stays meaningful months later.
Next steps
Related articles
Was this article helpful?