Published 2026-11-13 by destaque.ai, a Portuguese technology company that measures brand visibility in AI answers. Ler em português.
Method
We asked 40 shopper questions, in eight groups, to 11 AI assistants and surfaces. Seven were measured in the apps: ChatGPT, Claude, Copilot, Gemini, Grok, Meta AI and Perplexity. Through the API, the same questions ran on ChatGPT, Claude, Gemini and Grok, with and without web search, on DeepSeek and Mistral without search, and on Google's AI Overviews and AI Mode. The app answers were collected in Portugal on 7 November 2026; Gemini's and Perplexity's were collected again on 10 November. The API ran on 10 November, without stating the country or the date. The ten tasks were collected on 11 November. The study counts 727 answers: 276 from the apps, the first answer with text per assistant and question, excluding four Copilot forms, and 451 from the API. It excludes 19 API calls that failed and 10 questions for which Google showed no AI Overview. Answers from the two routes are counted separately, because the difference between them is a finding. The criteria were written before we read the answers, and changes made afterwards are declared as such. We read every answer by hand and kept, for each count, the sentence that supports it. Every false claim is recorded with the source that disproves it and the date. Prices are counted one by one, not per answer, because one answer can give several. Every price is measured against the shop's price: within 0.5% counts as right, within 10% as a price that has aged, and beyond that as wrong. The wrong prices in the apps were checked again in a browser on 11 November. Seven prices the first reading had marked as right, with gaps between 0.6% and 3%, became wrong when this rule was applied to all of them. When the answer does not name a colour, the nearest price among the page's colours counts. Sources are counted per answer, by the domain of each citation, and the country is read from the top-level domain. Search volumes come from Google Ads, read on 10 November.
Findings
- 208 of 254 prices checked on the shop's page matched what the AI said.
- 50 of 727 answers point to MediaMarkt as a place to shop today, in 9 of the 11 assistants.
- 101 of 240 answers given without web search are written for another country, nearly always Brazil.
Eight in ten checkable prices are right, and without search almost none can be checked
In the apps, the AI gave at least one price in euros for a specific product in 74 of 276 answers, 281 prices in all. We checked each one on the page of the shop the answer named, or on the Preço Baixo comparison site when the shop blocked automated reading. Of the 197 we could check, 168 were right.
The 29 wrong prices in the apps are not all inventions, because prices move in the weeks before Black Friday. We measured the gap against the shop's price, which is what the reader would pay. In 18 cases the price was within 10%, what we call a price that has aged, and in 10 it was more than 10% off. The other was for a model the shop no longer sold.
Through the API with search, checked prices fare worse: 39 right out of 55, checked on the shop's page without the second browser check the app prices had. Without search, the API gave 103 prices and only 2 could be checked. The rest name no shop (60), or are for 2023 and 2024 models shops no longer sell, or for shops whose page would not open (41).
The bigger problem is prices with no source. For 159 of the 513 prices in the whole study, the answer names no shop or page to back them. In the apps, Grok gave 23 such prices and Meta AI 19, against 1 from Claude and 1 from Copilot.
9 of the 11 assistants send the reader to a shop that no longer exists
MediaMarkt stopped trading under that name in Portugal on 1 October 2025, when its stores became Darty, and mediamarkt.pt redirects to darty.pt. In 50 of 727 answers, the AI points to it as a place to compare or buy this Black Friday. Only Google's AI Overviews and AI Mode never did.
Without search, the error appears in 30 of 240 answers, more than four times the rate of the API with search (6 of 211) and twice that of the apps (14 of 276). Grok does it in ten answers across both routes, Claude in nine, and ChatGPT and Mistral in eight each. In the apps, Claude is the only one of the seven that never does, but through the API without search it names MediaMarkt in seven answers.
In the apps, 18 answers name MediaMarkt, 14 of them as a shop to buy from today, and 15 name Darty, which runs this year's campaign. In one of them, Meta AI also recommends Dott, a marketplace whose address now leads to Worten.
Without search, AI invents websites and recommends services that have closed
We counted 109 false, verifiable claims in 81 of the 727 answers. Opinions, and claims no source can disprove, are left out. The API without search accounts for 76 of them, in 50 answers. Through the API with search there are eight, and Google's AI Overviews and AI Mode have none.
Mistral, without search, makes 34 false claims in 15 of its 40 answers. It recommends seven websites whose domain does not exist, such as pricerunner.pt, acordos.pt and promocoeshoje.pt. It also says Apple has stores in Lisbon and Porto, which is not true, and that the legal guarantee in Portugal is two years, when it has been three since 2022.
Idealo.pt appears in 14 answers without search as a product comparison site. In Portugal, the address leads to a page that only compares flights. There are also products that do not exist, such as the "Xiaomi 16 Ultra", from DeepSeek without search, and the "OnePlus 14 Pro", from ChatGPT with search.
In the apps, the cases are few and unrelated. Perplexity advises comparing prices at Boulanger, a French chain with no stores in Portugal. Claude presents Zarpo, a Brazilian hotel booking site, as a price comparison site. Gemini recommends Fakespot, which Mozilla shut down in July 2025.
Without search, four in ten answers are written for another country
None of the 40 questions mentions another country, and all are written in European Portuguese. Even so, 101 of the 240 answers without search are written for another market, with prices in reais, shops that do not sell here or Brazilian law. Mistral does it in 20 of 40 answers and Claude in 19.
Search reduces the error without removing it. With search, 36 of the API's 211 answers are written for another country, and Gemini does it in 14 of 28. ChatGPT with search did it in none of its 40. In the apps, 21 of 276 answers are written for elsewhere, led by Grok, in 9 of 40.
The examples are concrete. Grok tells app users they have "7 days to change their mind", which is the Brazilian rule, and suggests trips to Gramado "from R$ 675". In Portugal, the right to withdraw from an online purchase lasts 14 days. Meta AI answers in pounds in three answers.
The three questions on consumer rights (returning a purchase, the guarantee on a discounted product and an order that never arrives) show the same pattern. Without search, 11 of 18 answers cite Brazilian law, five mix Brazilian and Portuguese law, and only two stick to Portuguese law. With search, 12 of 18 cite Portuguese law. The date also goes wrong more often without search: two of six answers get Friday 27 November right, against four of five with search.
One in three app answers cites a Brazilian website
The app answers draw on websites from other countries. Of the 276, 249 cite at least one source, and 88 cite at least one site with a Brazilian domain (.br), such as TechTudo or Canaltech. The figure is a floor, because a Brazilian site on a .com domain does not count.
The gap between assistants is wide. Grok cites a Brazilian site in 28 of its 39 answers with sources, and Perplexity in 18 of 40. ChatGPT does so in just one, and it is also the one that cites Portuguese sites most: 209 of the 280 domains it cited end in .pt, against 137 of 501 for Meta AI.
Two kinds of Portuguese source stand out. DECO Proteste, Portal da Queixa, ASAE and the government consumer portal appear in 95 answers. Portuguese banks appear in 43, nearly always through the consumer guides on their blogs: Caixa Geral de Depósitos's Saldo Positivo, Santander's Salto and Cofidis's Contas Connosco.
Fewer than half of the answers warn about fake discounts
In 307 of 727 answers, the AI warns that some discounts are fake or that prices go up before the campaign. In 449, it explains how to check, with price history, a comparison site or a price alert.
The gap between assistants is wide. In the apps, Perplexity warns in 23 of 40 answers and Claude in 22. Copilot warns in 7 of the 36 it answered (for the other four it showed a form). Through the API, Grok without search warns in 5 of 40.
AI names the brands people search for most, and nearly all Portuguese clothing brands are left out
In the apps, we compared how often each brand was named with how often its name is searched on Google in Portugal. In electronics, the two rankings are similar (rank correlation 0.77): Amazon, Worten and Fnac are the most named, in 97, 89 and 77 of the 280 answers collected. For telecoms operators the match is closer still (0.86).
In clothing, the question "Which are the best shops for clothes on Black Friday?" got seven answers in the apps. Zara and Mango appear in all seven. Parfois, which has a Black Friday campaign confirmed for 27 November, appears in one. In November 2025, "parfois black friday" had 12,100 Google searches, more than "stradivarius black friday" (9,900), and Stradivarius appears in six of the seven answers.
The Portuguese exception is Sport Zone, in four of the seven. Tiffosi, Salsa, Sacoor and Lanidor appear in one answer each, and Quebramar, with a confirmed campaign, in none.
For supermarkets and telecoms operators, the AI also names brands with no campaign. Mercadona, which says it does not run promotions, appears in an answer about Black Friday deals at supermarkets. Digi appears in five of the seven answers about operators, and none of them credits it with a promotion.
With the product and shop in the question, 43 of 99 prices are wrong
Beyond the 40 questions, we gave the seven apps ten shopping tasks, nearly all with the product, the price and the shop, or the link to the page. One of them was "This Samsung Galaxy S26 Ultra 256 GB is €1,199.99 at Rádio Popular. Is that a good price?". We read the 70 answers against what the shop's page showed that day.
Even with that information given, 43 of the 99 prices we checked were not the shop's. In 31 the price was within 10% and in 12 it was further off. The answers completed the task in 38 cases, partly completed it in 21, and missed its core in 11, for instance by denying a discount the page showed.
Why AI gets prices wrong: memory, lag and pages with more than one price
We did not stop at the count. For each of the 72 wrong prices, in the questions and in the tasks, we recorded where the AI's number came from. In 19 cases the origin is visible: 15 point to the way the AI works and 4 to the way shops display prices.
On the AI's side, the main error is answering from memory. Without web search, the assistant answers with what it learned in training, and what it learned is the past: MediaMarkt appears in 30 of 240 answers without search and in 6 of 211 with search; the prices given without search are for 2023 and 2024 models, and only 2 of 103 could be checked; false claims drop from 76 to 8 when search is on. More than a year after the name change, MediaMarkt is still in the models' memory.
Even when searching, AI reads prices that have already changed. In 9 of the 72 wrong prices, the AI's number was the price the shop had charged on another day: on 29 August, between September and October, in an earlier promotion, on 10 November. The index the AI consults lags behind the shop, and in the weeks before Black Friday prices change from one day to the next. In three other cases the AI calculated or estimated the price, for instance applying a 7% discount the page did not show, and in another three it described a product the shop does not sell or has withdrawn.
On the shops' side, the page shows more than one price. In 4 of the 72 cases, the AI's number was on the page but was not the selling price. In three, it was the previous price or the "lowest price in the last 30 days", which EU rules require shops to show when they announce a discount; at Samsung, the AI gave that figure even though it had been given the link to the page. In the other, the AI gave the price of the other colours as "from", when the cheapest colour cost less. A page with the selling price, the reference price, the 30-day price and several colours is clear to someone looking at it, and ambiguous to something reading the code.
Language also pulls towards Brazil. The Portuguese-language web is mostly Brazilian, and AI goes there for sources. Of the 21 app answers written for another country, 16 cite a site with a .br domain. Among the answers that cite a Brazilian site, 16 of 88 address the wrong country; among those that only cite other sites, 5 of 162.
And search demand is not enough to make it into the answer. Parfois has more Black Friday searches than Stradivarius and a confirmed campaign, and appears in one answer in seven, against six. We did not measure why. The most likely hypothesis, still to be tested, is that AI names the brands it finds in the lists, guides and comparison sites it cites, not the ones people search for.
For the other 53 wrong prices, we did not find where the number came from.
What this means for shoppers and for brands
For shoppers, AI is useful for drawing up a list and remembering the rules, not for the final price. A price with no shop named should not be used. A price with a shop named should be checked on the page before paying, and an answer on consumer rights that mentions 7 days or a "nota fiscal" comes from Brazil. An assistant with search switched on makes far fewer mistakes than one without.
For brands, the lesson lies in clothing and in MediaMarkt. Having a campaign and search demand is not enough to make it into the answer: Parfois has both and appears in one of seven. A name that disappeared a year ago is still in the answers, above all without search (30 of 240).
The guide: five steps, and ten requests tested in seven apps
1. Use AI to decide between products, not to fix the price. Comparing two products at a given price was the only request all seven apps completed. Asking whether it is worth waiting for Cyber Monday was completed by five of seven.
2. Give it the link to the page, not just the product name. With the link, the three-month price history and the all-time lowest price got a useful answer in five of seven apps. Without a source, 159 of the study's 513 prices did not say where they came from.
3. Do not ask it to find a cheaper price or to validate a discount. "Find me this product cheaper at another shop" was completed by two of seven, and three named as cheaper a shop that was not. "Is this 25% discount real?" was completed by one of seven, and two denied a discount the page showed. A price comparison site with price history is the tool for that.
4. Distrust prices with no shop and lists built to a budget. The 500 euro shopping list was completed by one of seven. In one answer, two of the three prices did not exist at the shop it named and, at the shop's prices, the list went over budget. Another gave caps per category and no product.
5. Check at the shop, and against Portuguese law, before paying. Even with the product and the shop in the question, 43 of the 99 prices checked were not the shop's. In Portugal, the period to withdraw from an online purchase is 14 days and the legal guarantee is three years. An answer that mentions 7 days or a "nota fiscal" comes from Brazil. If the app allows it, switch on web search.
The ten requests, ordered by how much they are worth asking:
"Compare these two products and tell me which is worth it at this price: [product A, price] and [product B, price]": 7 completed, 0 partly, 0 failed. Verdict: Use.
"Show me this product's price history over the last 3 months. [link]": 5 completed, 2 partly, 0 failed. Verdict: Use.
"Should I buy this [product] now or wait for Cyber Monday?": 5 completed, 2 partly, 0 failed. Verdict: Use.
"This [product] is [price] euros at [shop]. Is that a good price?": 5 completed, 2 partly, 0 failed. Verdict: Use, and check the prices it quotes: 7 of 19 were wrong.
"What is the lowest price this product has ever had? [link]": 5 completed, 1 partly, 1 failed. Verdict: Use with care: 6 of 13 prices wrong.
"Did this product's price go up before Black Friday? [link]": 4 completed, 1 partly, 2 failed. Verdict: Use with care: 5 of 7 prices wrong.
"Is this online shop trustworthy? [shop link]": 3 completed, 3 partly, 1 failed. Verdict: Use with care.
"Find me this product cheaper at another shop in Portugal. [link]": 2 completed, 2 partly, 3 failed. Verdict: Do not rely on it.
"Is this [x]% discount on [product] at [shop] real?": 1 completed, 4 partly, 2 failed. Verdict: Do not rely on it.
"Make me a Black Friday shopping list with a 500 euro budget: [list of needs]": 1 completed, 4 partly, 2 failed. Verdict: Do not rely on it.
Each app answered each request once, with one product per request, so these numbers read as a test, not a ranking.
Dataset
Counts per assistant and per route (app, API with search, API without search) for the 727 answers: prices right, wrong, without a source and not checked; answers pointing to MediaMarkt as a current shop; answers written for another country; false claims; warnings about fake discounts; sources cited in the apps; brand mentions per segment with Google search volume; and the outcome of the ten shopping tasks. Column labels are in Portuguese (metrica, segmento, valor, unidade, nota).
Download the dataset (CSV). Licensed CC BY 4.0: reuse with attribution to destaque.ai.
Limitations
Each assistant answered each question once. A second run may give a different answer, above all on prices and shops. Prices are checked against the shop on the day of the check, and around Black Friday prices change from one day to the next. That is why we separate a price that has aged from one more than 10% off. 100 prices could not be checked, because the page would not open, did not list the product, or the model is no longer sold. The API ran without a date. An answer without search that gives the 2024 Black Friday date is answering from what the model learned, so the date is counted separately and not as a false claim. Sources and brands are counted in the apps only. Clothing, supermarkets and telecoms had one question each, so seven answers. Those figures are given as counts and are too small for percentages.
destaque.ai is a Portuguese technology company, based in Lisbon, that measures brand visibility in AI answers. This study is published in full in Portuguese at https://www.destaque.ai/estudo/black-friday-portugal-2026. Methodology available on request: info@destaque.ai.
