To compare a jewelry brand's presence in AI answers, save the text and keep the counting rule consistent. A detected name tells you where the brand appeared. Decisions about inventory, advertising or customer service also require the business's own inquiry and sales records.
This distinction is one lesson from BRGEO-1, Alexandre Caramaschi's audit of entity-mention measurement in generative engines. The corrected empirical version, dated September 11, 2026, is available as a Zenodo preprint. It has not completed peer review.
The amount of text examined changes the result
Saving only the opening of an answer can hide names that appear later. BRGEO-1 compared the same 7,436 canonical Perplexity responses collected between April 23 and August 31, 2026. The name extractor remained fixed; only the text window changed.
The longer retained text produced detections in 5,631 responses, or 75.73% of that set. The first 200 Unicode characters produced detections in 3,863 responses, or 51.95%. The shorter window missed 1,768 positive responses, a difference of 23.78 percentage points. These figures come from the paired comparison in the September 11 preprint.
A retailer receiving two reports should ask how much of each answer was examined. If one provider stores the opening and another retains later passages, the measurement procedure can account for a difference. That difference alone does not establish a gain or loss in the brand's presence.
The longer text retained by the study is not necessarily the complete original response. Recording what was available preserves this limitation. Asking for the answer behind a metric often provides more useful information than a percentage on its own.
Variations of a name can increase counts
Define which names refer to the same brand before adding occurrences. In BRGEO-1's retail subset, 316 responses contained overlapping Amazon and Amazon Brasil labels in the same passage. Merging those labels reduced incidences from 6,037 to 5,721 while leaving all 3,838 positive responses unchanged.
These figures, also from the September 11, 2026 version, illustrate the difference between counting labels and counting responses. They do not measure market share. They describe how a counting system handles two forms of a name.
A jewelry catalog may include a manufacturer's name, a shop name and a collection name. Record how they relate: a collection belongs to a brand, while a reseller may carry several manufacturers. Merging every name hides useful distinctions; treating every variation as a separate business can inflate the list.
Mentions and sales require separate records
Read the sentence containing the name before interpreting the metric. An answer can give an example, make a comparison or state that information is unavailable. A string search can detect a name in any of those settings.
BRGEO-1 audited measurement rules. It did not test Herreira sales, reseller performance or advertising conversions. The application to jewelry retail is a proposed way to examine reports; it does not attribute participation in the research to the brand.
If mentions increase after a communication initiative, retain the questions and consultation dates as well. Changes in questions, services or reading rules may explain part of the variation. Commercial evaluation requires separate records of inquiries, their sources and completed sales.
Keep a record that allows comparison
Use a short spreadsheet to retain each consultation and its counting rule. Check the following items before using a reported number to decide on inventory or advertising:
- Copy the question, service name and consultation date.
- Save the available answer and record whether any text is missing.
- Mark the detected name and the sentence containing it.
- Record which variants refer to the same brand, shop or collection.
- Compare periods after confirming that the measurement rules stayed consistent.
Alexandre Caramaschi, CEO of Brasil GEO, former CMO of Semantix (Nasdaq) and cofounder of AI Brasil, plans to submit the study for journal evaluation. The public preprint makes the methods and results available while that editorial step remains pending; it does not represent journal acceptance.
Open your latest brand-presence report and inspect one answer that entered the count. Save that answer and the counting rule alongside the number so the next comparison has a traceable basis.