Skip to content
Published method

The numbers,
in the open.

ANSWERS READ
IN 5 STUDIES THAT MEASURE ANSWERS
193 answers · The 2026 Brazilian election141 answers · Football in Portugal2,154 answers · Politics as a reference group616 answers · A map of the companies of Portugal2,205 answers · Consumer services across seven engines

Methodology published, datasets open to download, and every figure states the number it was taken from.

Buyers no longer search. They ask. And they walk into the decision with the name the assistant gave them.

10 → 1
results in a Google list. One name in an AI answer.
39/39
answers on the best airline to fly from Portugal include TAP among the airlines they choose. A map of the companies of Portugal, August 2026.
11/540
answers pick a party when asked to choose one. Politics as a reference group, September 2026.

Published studies.
Real brands, real answers.

7
finished studies
5,309
AI answers read in the studies that measure answers
Measured with Periscopy

Radar: the brands AI recommends in Portugal

The brands AI assistants name in each sector of the Portuguese market.

THE LEADER IN EACH SECTOR% OF ANSWERS

On 12 and 13 October 2026, between the two rounds of the Brazilian presidential election, 11 AI assistants answered 15 questions: who will win, who would make the best president and who to vote for. The vote question was also asked in the voice of eight different voters. Each answer was read against a criterion written before reading, and every reading keeps the sentence that supports it.

Key findings
  • Among the assistants that search the web, 8 of 8 predictions that name a candidate say Flávio Bolsonaro wins.
  • 13 of 17 voting recommendations that name a candidate say Lula, and 63 of the 80 requests for a vote get no name.
  • Perplexity gives 8 of the 13 recommendations for Lula.
  • Google AI Overviews and AI Mode gave no name in 25 answers, and Gemini none in 26 attempts.

The methodology is not a promise. It is published.

01
The question is fixed before the data
What counts as a mention, and what counts as being chosen, is written down before a single answer is read. Deciding afterwards is how a study finds what its author was hoping for.
02
Names that are also common words are counted by reading
In Portuguese, NOS is also "in the", ERA is also a verb, and Logo is also the word for "soon". Those are counted by reading the sentence, not by matching a string, and the exceptions are listed. In one pass, this caught 13 false mentions out of 18.
03
Published numbers come from code, not from a conversation
The first study was counted by hand and got 3 of its 14 denominators wrong. Every figure now comes from a script anyone can run again, with the denominators defined in one place and the result in version control, so a later run shows the difference as a diff.
04
Human judgement is stored as data
Whether an answer chooses a brand or only lists it cannot be derived. People make that reading, it is recorded answer by answer, and it stays auditable instead of living in the head of whoever read it.
05
The denominator is always stated
A percentage without the number it was taken from is not a finding. Where a question could not be measured, it is said, and it is kept out of the ratio rather than counted as a zero. The Google barometer ran on one engine and says it is directional. The consumer-services study ran on seven.
CORRECTED IN PUBLIC
A
A 30-point gap between two assistants was ours
In the parties study, the first count put one party in 57% of one assistant’s listings and 88% of another’s. The detector was looking for prose, and the assistants answer in lists. Once the pattern was fixed, the assistants compared all read 100%, and 118 answers had been undercounted.
B
A claim about web search, withdrawn
We wrote that one assistant did not search the web in 49% of its answers. Those rows did not come from the channel we said we were measuring. Every stored answer now declares how it was collected.

Open data.
Redo the math.

One file per study, the same one destaque.ai publishes, served from this site. Each holds what its study page shows: the counts that reproduce the charts and, in the two newest studies, the reading of every answer with the sentence that supports it.

LICENCE · CC BY 4.0