Articles · 17 November 2026 · 6 min read
Black Friday: your sales exist, just not for ChatGPT
We asked four AIs about a highly rated French brand with a technically clean website. They know its reputation. They ignore its sales, and one of its pages makes them believe products are sold out.
Ten days before Black Friday, brands are preparing their discounts, banners and e-mails. One question rarely gets asked: what does ChatGPT tell a customer who asks for the price of one of your products?
To find out, we analysed the website of a French brand of bags and travel accessories, sold online and rated 4.7 out of 5 by several thousand customers. We do not name it: its case is not unusual, and that is exactly what makes it instructive.
The method
We read 20 product pages, then asked shopper questions about some twenty products (price, availability, delivery time, return window) to four AI models: OpenAI's (ChatGPT), Google's (Gemini), Anthropic's (Claude) and Perplexity's. 77 questions in total. Every answer was compared with what the website showed on the same day.
The contrast is the first lesson. AIs know this brand is trustworthy. They do not know what it sells today, or at what price.
1. Sales are invisible
Six products in the sample were on sale, from −30% to −58%. For each one, two to four AIs out of four quoted the full price, the one crossed out on the website.
A real example: a backpack shown at €62.30 instead of €89. Three AIs out of four answer “€89”. A pouch sold at €5 instead of €12: three AIs out of four answer “€12”.
The website did nothing wrong. The AIs remembered the usual price, seen more often and for longer, and have not picked up the discount yet. For the customer, the effect is the same as an error: they think the product costs more than it does, and may give up before reaching the page where the sale is shown.
On Black Friday, discounts last a few days. AIs often take longer than that to see them.
2. One hidden word is enough to “sell out” a product
For two products that were in stock, two AIs out of four said they were unavailable. The structured data said “in stock” and the page showed the buy button.
The cause was in the page code: the word “indisponible” (unavailable), invisible to the customer (the store theme uses it for another size or state) but read by the AI. The customer sees an available product. The assistant they ask reads a stock-out.
You cannot spot this kind of defect by browsing your own website. You only see it by reading the page the way a machine reads it.
3. A technically clean website is not enough
This website does better than average on the technical points usually recommended: price and availability in the structured data, brand, product identifiers, return policy. It lacks little (shipping cost and time in that data, a clear stock line in the text).
And yet the AIs get it wrong. Many of the errors come from elsewhere: what they memorised, old pages, resellers, or a detail of the code nobody looks at. Optimising your pages is necessary, but only checking the answers tells you whether it worked.
What it costs
We estimated the lost revenue from the website's public traffic ranking, a cautious conversion rate, the average price of the products read, the share of customers who ask an AI before buying (30% according to Ipsos for Joko, May 2026) and the share of wrong answers measured on this website.
Central estimate: about €7,000 of lost sales per month, within a range of €3,500 to €21,000 depending on actual traffic. It is an estimate, not a measurement: no public study measures how many misinformed customers give up, and we set it at 25%. But the order of magnitude is enough to show the issue is not cosmetic.
What to do before Black Friday
- Write the sale where machines read it. The discounted price and its end date (“priceValidUntil”) in the structured data, the crossed-out price visible on the page.
- Publish a dedicated sales page early, one AIs can find and cite, rather than relying on banners that appear on the day.
- Hunt down hidden text that contradicts the page: “sold out” or “unavailable” left by the theme, old prices in the code.
- Check resellers and comparison sites still showing old prices: AIs read them too.
- Check the answers, not just the pages. Ask AIs your customers' questions, before and after your fixes.
Limits of this test
This is a case study of one website, not a statistic about French e-commerce. The questions were asked once, to the AI models available through Perplexity's interface; the consumer apps of these AIs may answer differently. Sales and stock change from day to day: these results apply to the date of the test, 17 November 2026.