A combined AI visibility percentage can fall even when the measured rate improves in every country. The cause can be a change in the mix of observations: the report now includes more answers from a market where the brand has lower visibility. Inspect country-level counts before interpreting the headline, and use a separately labeled fixed-weight index when the business needs a consistent comparison. Neither an unweighted panel nor an editorial weighting scheme estimates global customer demand without supporting audience evidence. The number needs a defined population, event and denominator before its trend can guide spending.
The contradiction inside a perfectly correct total
An export business starts monitoring in two markets. Its home English-language market receives most of the initial attention. Later, the team expands a smaller market's question set. The next report contains more observations from the expansion market and announces a decline in overall visibility.
The arithmetic may be correct. The conclusion may still be wrong.
The report is answering a different question because its composition changed. It no longer gives the same weight to each market. This can reverse the direction of an aggregate trend, a pattern commonly discussed as Simpson's paradox in statistics. Research by Xu, Brown and Grant describes the contrast between a pooled relationship and relationships within groups. Identifying that contrast does not, by itself, identify a cause or tell a business which market to prioritise. Detecting Simpson's Paradox.
For GEO operations, the useful response is to reconstruct the counts. A more elaborate dashboard cannot repair an undefined comparison.
A two-market example you can calculate by hand
The following figures are fictional. Each observation is an eligible consumer AI answer to a recorded question. A success means a recommendation that meets a predefined entity and recommendation-reason test. It does not mean a brand string, citation or organic link merely appeared.
| Market | Period 1 recommendations / valid answers | Period 1 rate | Period 2 recommendations / valid answers | Period 2 rate |
|---|---|---|---|---|
| United States | 40 / 80 | 50% | 12 / 20 | 60% |
| Germany | 2 / 20 | 10% | 16 / 80 | 20% |
| Combined observations | 42 / 100 | 42% | 28 / 100 | 28% |
Both market rates improved by ten percentage points. The combined rate fell by fourteen percentage points.
In the first period, the higher-rate US group supplied 80% of the observations. In the second, it supplied 20%. Germany supplied the remaining 80%. More weight moved to the lower-rate group, outweighing the improvement within each group.
Nothing in this example demonstrates that real customers became more or less likely to receive a recommendation. The observations were allocated by the monitoring programme. The example proves a property of the arithmetic, not a marketing effect.
It also explains why the same brand can appear to improve in one agency report and decline in another without either report having an addition error. Their country, question or surface composition may differ.
Reconstruct the headline as a weighted average
For a combined rate, each group's weight is its share of eligible observations. Multiply its rate by that weight, then add the results.
Period 1 is:
0.80 × 50% + 0.20 × 10% = 42%.
Period 2 is:
0.20 × 60% + 0.80 × 20% = 28%.
This reconstruction is worth doing before debating algorithms, website edits or competitor activity. If the report does not expose the component counts, ask for them. Rounded percentages alone may be insufficient to reproduce the result.
The same issue can arise across platforms, industries, languages and buyer intents. Adding many low-visibility early-research questions can depress a total while existing purchase-intent questions improve. Increasing the share of one model can change a combined score even if no model-specific rate moves.
That does not make expansion undesirable. It means the programme changed. Describe the change rather than presenting it as a like-for-like performance result.
Keep the observed rate and a fixed-weight index separate
A fixed-weight comparison asks a different, useful question: what would the result look like if both periods used the same market weights?
Using Period 1's 80% US and 20% Germany weights gives:
- Period 1: 0.80 × 50% + 0.20 × 10% = 42%.
- Period 2: 0.80 × 60% + 0.20 × 20% = 52%.
The fixed-weight index rises by ten percentage points. The observed combined rate falls from 42% to 28%. Both calculations can be reported, provided their definitions remain visible.
Call the second calculation a fixed-weight monitoring index, not a global recommendation rate. The chosen 80/20 weights came from an earlier observation mix. They do not establish the distribution of buyer questions in the real market.
An equal-weight index would produce another result: 30% in Period 1 and 40% in Period 2. Equal weighting treats the two markets as equally important for this index. It does not establish that they contain equal demand, revenue or AI usage.
Choose weights for a stated purpose before reviewing the outcome. Changing them afterwards to make the trend look attractive defeats the comparison.
Location is more than a country written in the prompt
Before comparing countries, verify what the country field means. It might describe the prompt's target location, the collector's network location, a database classification, an account setting or an interface language. These are related pieces of context, not interchangeable evidence.
OpenAI explains that ChatGPT search may use approximate location from an IP address and can use saved context when forming search queries. A question mentioning Germany does not prove that the session originated in Germany. OpenAI's search and location guidance.
Vendor documentation also uses different regional structures. Ahrefs describes locale parameterisation in its Brand Radar methodology as reflecting the country and language distribution of its keyword data, and explicitly distinguishes that from estimating actual AI usage by country. That is a documented modeling choice, not a universal definition of regional monitoring. Ahrefs Brand Radar methodology.
Keep the actual context fields in the export. If a field is unavailable, say so. Do not turn a market label into proof of a verified local consumer environment.
A six-step review before reacting to a regional trend
- Name the event. Decide whether the numerator counts mentions, citations, recommendations or another reviewed outcome. Keep the definition consistent across periods.
- Recover the counts. Obtain successes and eligible observations for every country, surface and question cohort used in the headline.
- Check composition. Compare each group's share of observations. Identify newly added questions, removed markets, model changes and collection gaps.
- Compare matching groups. Read the direction and size of changes inside stable country and surface groups before using the combined number.
- Calculate a declared fixed-weight view. Use unchanged weights and label the resulting index separately from the observed combined rate.
- Record the operational decision. Explain whether the finding calls for a content correction, more complete sampling, a market-specific project or no action yet.
These steps are diagnostic. They do not turn a small convenience panel into a representative survey or establish that a content change caused an improvement.
Collection failures can change the mix accidentally
A team may intend to keep its regional allocation fixed and still observe a changing mix. If one country loses sessions to an access limitation, its completed answers contribute less to the combined result.
Suppose a difficult market has many failed attempts. Dropping those attempts from the numerator and denominator may be necessary when calculating a valid-answer rate, but it also changes the composition of the completed sample. The headline can improve simply because the difficult market is less represented.
Show the planned and completed counts by group. Keep unavailable results separate from answers in which the brand is absent. Report whether a fixed-weight view can actually be calculated: if a required group's rate is entirely missing, a complete index may be unavailable. Do not silently redistribute its weight to the groups that succeeded.
An operational failure is a reason to repair collection within the platform's permitted access conditions. It is not evidence of a visibility gain or a reason to refresh repeatedly until a preferred brand appears.
Decide what the business wants to learn
An exporter can reasonably prioritise a small, high-value country over a much larger market. The monitoring panel should then be described as a business-priority panel. A different programme might weight markets using qualified pipeline or an approved expansion plan.
Those are management choices. Document the source, owner and effective date of the weights. Revenue weights should not be relabeled as search-volume weights. A question suggested by an AI tool is not measured demand merely because it sounds plausible.
When assessing Semvi by Xindar, ask how country, language, collection surface and model appear in the underlying evidence. Its product owner's confirmation of consumer ChatGPT and Gemini sampling does not, by itself, confirm every country, city or reporting-weight configuration. Test the markets relevant to your brief and retain the actual outputs.
For manufacturing exporters, this can prevent a misleading budget decision. A weaker combined score following expansion may reveal that the company has finally started measuring a poorly served market. The appropriate response may be a local product-evidence project, not abandoning a strategy that improved in the original market.
Questions teams ask
Is a combined visibility rate always misleading?
No. It accurately summarises the observations included under its definition. The risk is treating a changing mixture as a stable comparison or as a measurement of all real users.
Should we always use equal country weights?
No. Equal weights are one declared policy. Business-priority or baseline weights can also be useful. Explain the purpose and keep the policy fixed during the comparison.
Can a fixed-weight index prove GEO caused an increase?
No. It addresses composition, not every source of bias or change. Model updates, question selection, session context and other interventions still require attention.
What if a country has no valid answers in one period?
Mark the group as missing and disclose the collection gap. Do not assign it zero brand visibility or quietly remove its weight from an index that was supposed to be fixed.
Does English-language monitoring represent the United States?
Language alone does not establish location, account context or demand. Record what was actually verified and avoid geographic conclusions broader than that evidence.
Method and further reading
All numbers in the worked example are invented for arithmetic demonstration. The proposed reporting procedure separates composition from within-group movement; it is not a population estimate or a causal test. For the broader setup, see designing a stable prompt panel and regional AI search conditions.