How to Tell If a GEO Change Worked: Count the Answers, Not the Score

Why the Score Can't Tell You

An overall visibility score averages every tracked answer. One question won on every engine is worth 5 points in a 20-question set, and one won on ChatGPT alone is worth 2.3, because ChatGPT is one weighted engine of 3. Most changes target a handful of questions, so the move they can produce is small by construction.

The noise is larger than most real changes

We measured how far scores move between consecutive runs: 14.5 points overall at the 95th percentile, more for one engine or one category. Ranqo sets its bands at 15, 20 and 27 points. Below the band, a change in the score is not evidence either way. In the example above it takes 4 questions won on every engine before the overall score clears its band.

Narrow the metric to the change

A metric that only counts the questions a change targets is less diluted. A category score in a small category clears its band with fewer wins, even though its band is wider:

The last row is the catch. In a 2-question category one question won on every engine is half the score, so a single win clears the band; that says more about the category's size than about your change. Every score is either diluted or fragile, which is why the protocol below counts answers directly.

The Protocol: Count the Answers

Counting answers works because most answers are stable. In our 102-brand study, 77.5% of brand, question and engine combinations either always or never named the brand, so a change in one of them means something.

1. Name the target answers before you ship

List the tracked questions the change is meant to win, and on which engines. A new comparison page targets the comparison questions; a fix for one engine's crawler targets that engine. Writing this down first is what stops a lucky answer elsewhere being claimed afterwards.

2. Record the baseline

For each target answer, note whether the latest run named you. Most will not, which is why you are making the change. Note the ones that already do, too: a change can lose answers as well as win them.

3. Pick control answers

Choose a few tracked answers on the same engines that the change does not touch. Engines change their models and retrieval without notice, and answers also vary on their own: a 2026 academic audit found that the products ChatGPT's web interface recommended for the same query, asked hours apart, overlapped by a mean of only 0.178, which the authors call a lower bound. If the control answers move as much as your targets, the engine moved, not your page.

4. Ship one change at a time

Two changes aimed at the same answers in the same week cannot be told apart afterwards. If you must ship several, give them separate target answers.

5. Wait in runs, not days

Your numbers only change when a tracking run happens, so count runs, and never trust one measurement. Even with sampling randomness switched off, a model can answer differently: one lab got 80 distinct completions from 1,000 identical requests to an open model. Take an early read after 2 runs: one to see a change and one to see it hold. Call "no effect" only after 3 runs for a page or technical fix and 4 for content, authority or outreach work, which take longer to be picked up.

6. Classify each answer

The two-run rule comes from how answers behave. A mention that has just appeared is still there next run only 54.3% of the time. One seen in two consecutive runs holds 96.1% of the time, and an answer that has left you out twice names you next run only 3.4% of the time.

What Analytics Can and Cannot Add

First-party data is worth checking alongside the answers, but none of it isolates one change.

Search Console

Google counts visits from AI Overviews and AI Mode inside Search Console's ordinary Web search type. Its new generative AI performance reports, available to every site since 31 August 2026, add impressions, pages, countries, devices and dates for those features. They tell you a page appeared, not which question it answered or whether you were the brand recommended.

Google Analytics

GA4 now groups recognised assistants into an AI Assistant channel, but files AI Overviews and AI Mode under Organic Search, and many AI-driven visits arrive with no referrer at all. Traffic moves for many reasons, so treat it as corroboration that arrives later, not as the test.

A Measurement Log to Copy

One log per change. Fill in the top half before you ship and tick the runs as they land.

GEO CHANGE MEASUREMENT LOG

Change:            ____________________________________________
Shipped on:        ____________   (run after this one is run 1)
Metric it targets: overall / one engine / one category: ________
Noise band:        overall 15 pts · one engine 20 · one category 27

TARGET ANSWERS (write these down BEFORE you ship)
  question                          engine      at baseline   run 1  run 2  run 3  run 4
  ________________________________  _________   absent        [ ]    [ ]    [ ]    [ ]
  ________________________________  _________   absent        [ ]    [ ]    [ ]    [ ]
  ________________________________  _________   present       [ ]    [ ]    [ ]    [ ]

CONTROL ANSWERS (same engines, questions the change does not touch)
  ________________________________  _________   ________      [ ]    [ ]    [ ]    [ ]
  ________________________________  _________   ________      [ ]    [ ]    [ ]    [ ]

VERDICT PER TARGET ANSWER
  won          absent at baseline, named in two consecutive runs after
  provisional  absent at baseline, named in the latest run only
  held         present at baseline, still named in the latest run
  lost         present at baseline, missing in two consecutive runs after

READ IT
  Early read after 2 runs. Call "no effect" only after 3 runs for a page
  or technical fix, 4 for content, authority or outreach work.
  If the control answers moved as much as the targets, the engine changed,
  not your page.
  Report a score change only when it is larger than the band.

Reporting it

Report answers won and lost, by engine. Report a score change only when it is larger than its band, and say plainly when a change produced nothing detectable. A confirmed win shows the answer changed after you acted; it does not prove your action caused it. The control answers rule out the most common other cause, an engine-wide shift. Our client report template carries the same rules for agencies.

How Ranqo Does It

The Action Center runs most of this protocol on every recommendation: it takes the baseline, counts the window in runs and classifies the answers, and you press Re-measure now when the window is up. Naming your own target answers and watching controls stay with you; the Prompts page's Tracking tab shows, for every question, whether each engine's latest answer named you.

The baseline is taken when the recommendation is written

Each recommendation is bound to one metric (overall visibility, one category, or one engine's mention rate) and records which tracked answers in that metric's scope named you at the time. The scope is the whole metric, so an overall-visibility recommendation counts every tracked answer, not only the ones your change aimed at; for a narrow change, keep your own target list as above. Its expected outcome is written in answers: how many in scope already name you, what each further one is worth, and, when enough answers are left to win, how many you would need before the change stands out from normal movement.

The window is counted in runs

Moving a card to Done starts its window. The card counts down After N more runs to the first reading and shows a run counter toward the verdict window: 3 runs for page and technical work, 4 for content, authority and outreach. Each card takes one reading and keeps it, so for a verdict rather than an early read, press Re-measure now once the run counter is full.

It reports answers won, and points only beyond the band

The reading compares the baseline answers with the two most recent runs since Done and counts an answer as won only when both name you, as described in confirmed wins. On the card and in the drawer, a score change prints in points only when it is larger than the noise band; a smaller one reads No detectable change, or the number of answers won when two runs confirmed them. Technical fixes say up front that no visibility change is expected from them: they make a page readable, and structured data on its own has not been shown to move AI visibility.