Mapping Portugal's companies: across 616 AI answers, being named is not being chosen

Two counts, not one score

Counting mentions does not distinguish a name buried in a list from a name the answer actually chooses, and that difference is what this study measures. So there is no composite score: there are two counts. Named is in how many answers the brand appears anywhere in the text. Chosen is in how many of those the answer distinguishes it above the others it itself names, through a place in an ordering, a verdict against a stated need, or the closing sentence where the model commits. Conversion is one divided by the other.

A superlative inside a description is not a choice, however strong. "World leader in cork" describes the brand on a shelf; it does not say it is the best Portuguese brand, which was the question asked. The full criterion is published on the Portuguese page and was written before the re-read, so that it could not be shaped by what appeared.

This has a name: share of recommendation

The metric the market uses for AI visibility is share of voice, and it counts mentions: of everything the category says, how much is about you. It was inherited from advertising, where it makes sense, because an ad seen is an ad seen. In an AI answer it stops making sense: the name that closes the answer and the twentieth name in a catalogue earn the same point.

We call the other half share of recommendation: of every choice in the category, how many are yours. Same arithmetic, different numerator, and it is that numerator this study went to read, answer by answer. A third number follows from the two: the recommendation rate, your choices divided by your mentions, which is the conversion the tables publish.

The three mean different things and ask for different work. Losing mentions is a presence problem: you are not in the sources the model reads. Losing choices is an argument problem: you are there, and nothing written about you gives the model a reason to put you above the rest. Both fall on the same chart and neither is fixed the same way.

The distance between them is wider than you would expect. In supermarkets 96% of mentions are choices: whoever is named is recommended, because the question has few possible answers and the model commits. In the open question about the best Portuguese brands it is 20%: the model returns a catalogue of thirty names per sector, and four out of five mentions choose nothing. The same brand, in those two questions, would have the same share of voice and two different businesses.

The clearest case is in the Portuguese brands table. Salsa is the most named brand in that question, in 31 of 38 answers, and is chosen in 7. EDP is named in 26, fewer, and chosen in 15, more than double. A dashboard counting mentions puts Salsa ahead. The answers put EDP. And then the zeros, which a mention metric never shows: thirteen brands in this study are named dozens of times and chosen never. Compal in 18 answers, Sonae in 14, Feedzai in 12. They are not absent from the market of answers. They are absent from the decision, and their share of voice looks healthy.

Two things share of recommendation is not. It is not approval: a brand counts as chosen even when the next sentence points at its price, and there are categories where the models choose and criticise in the same breath. And it is not market share: it is recommendation share, which is what the buyer sees before deciding, and the two diverge precisely where a large brand is badly explained.

This is not a concept for one study. It is what a brand needs to see every week, which is why Periscopy now measures it, with this same criterion, next to share of voice.

The same question, four ways, two rankings

Changing the wording of the banking question does not add noise to one ranking; it produces two. Asking for a top 5 returns the institutional hierarchy: Novo Banco appears in 87% of answers and Revolut in 8%. Asking what to recommend to someone living in Portugal inverts almost everything: Revolut rises to 90% and Novo Banco falls to 36%.

The reading extends to choices, and there it is harsher. From the most formal wording to the most personal, mentions rise from 245 to 341 while choices fall from 201 to 140. The question that most resembles a real customer produces the longest list and the smallest commitment. The practical consequence for anyone measuring AI visibility with a single prompt formulation: they are measuring one of the two rankings, and generally do not know which.

Who writes the answers

328 of the 616 observations show their sources, across 2,993 citations, and they are not the sites a brand usually monitors. The most cited domain in the whole study is the Portuguese consumer association DECO Proteste, with 105 citations concentrated in 6 of 14 question blocks. Reddit follows with 88 and appears in 13 of 14, making it the most transversal source in the study. Two foreign fintechs writing guides about Portuguese banks, Wise and Qonto, together account for 135 citations. Wikipedia appears 34 times in 2,993.

Per category, the dominant source changes completely. In telecommunications a single site accounts for 56 of the category's 220 citations. In the open question about the best company in the country, the dominant source is a workplace ranking. In private healthcare, the most cited source is the hospital group's own website, so the brands are themselves the source of the answer that ranks them.