We asked ChatGPT where to buy 20 Black Friday products. It got about half wrong.

Ask ChatGPT where to buy a Ninja air fryer in Australia and it may send you to JB Hi-Fi at $127. On the day we asked, JB Hi-Fi was charging $249.

On 20 October 2026 we asked ChatGPT where 20 popular Black Friday products were cheapest in Australia, twice each, and checked every answer on the recommended store's own website. 21 of the 40 answers were wrong. Every wrong price was lower than what the store was charging, by as much as 49%, and for 11 of the 20 products, the second answer named a different store or price.

This matters before the biggest sale of the year. 1 in 5 Australians already use AI tools when shopping online, and 57% of those who don't say they may start within the next year, according to research by Commerce and PayPal Australia. Below is what ChatGPT got wrong, where the wrong prices came from, and what to check and fix before Black Friday on 27 November.

How we ran the test

We asked the question shoppers ask before a sale, about 20 well-known products, and checked every answer by hand.

• The question: “Where is the cheapest place to buy the [product] in Australia right now?”, typed the same way every time.
• The products: 20 of the best-known products Australians buy over Black Friday, across gaming, audio, tech, appliances, barbecues, beauty and gifts. We fixed the list before asking anything. Every product is in the results table below.
• The set-up: a free ChatGPT account, an Australian connection and a new chat for every question. Each product was asked twice, several hours apart, on 20 October 2026. Most people use the free version: OpenAI reported more than 50 million paying consumer subscribers and 900 million weekly users in February 2026 (Digital Information World).
• The check: the same day, we went to every store ChatGPT recommended, found the exact product asked about (model, size and generation), and recorded the price on the store's own website. Every recommended store was checked.
• The rules: where an answer named two stores at one price, it counted as wrong if either store's price was wrong. Members' prices, discount codes and group-buy prices didn't count as the store's price.

About half the answers were wrong

Of the 40 answers, 19 quoted the right price at the store ChatGPT recommended. The other 21 were wrong, in four different ways.

Two answers named two stores and got one of them right. Counting those as right would make it 21 right and 19 wrong, which is why we describe the result as about half.

Every answer, product by product

Every wrong price was too low

All 16 wrong prices were below what the store was charging. Where the store had a price to compare, the gap averaged $71, and 11 of the 15 gaps were more than 10%.

The 16th wrong price, eBay at $459 for the PlayStation 5, had no new listing near it to compare against.

Why every error went the same way. The question asked for the cheapest price. When the sources disagreed, the answer went with the lowest number, and in our test the lowest number was usually out of date: an expired sale, a members' price, a group-buy floor or a different product.

Most shoppers will click through and see the real price before they pay. If you're the recommended store, that means a visit from someone expecting a price you don't charge. If you're left out, it means no visit at all.

Ask again and the answer changes

For 11 of the 20 products, asking the same question in a new chat later the same day returned a different store or a different price.

The Ninja air fryer shows how far an answer can move. The first answer recommended Amazon Australia at $199, which was the right price. The second recommended JB Hi-Fi at $127. JB Hi-Fi was charging $249.

Getting the same answer twice didn't make it right. Nine products got the same answer both times, and six of those were wrong both times: the Garmin at Kogan, the Dyson at Bunnings, the Weber at Myer, the CeraVe at Chemist Warehouse, the Estée Lauder serum at Beauty Amora and the LEGO set on eBay. The stores' prices didn't change between the two asks.

If you track how ChatGPT presents your products, ask each question several times and record every answer. A rank tracker records one result, and a shopper who asks ChatGPT gets one of several possible answers.

The stores ChatGPT sent shoppers to most

Officeworks was the store ChatGPT recommended most often, and it was usually priced right. Myer was recommended almost as often, and was usually priced wrong.

Two of Myer's three misses were the Weber at $458.15 when Myer was charging $485.10. The third sent shoppers to Myer for a Kindle Paperwhite. Myer's site lists Kindle cases and covers, but not the Kindle itself.

It flags its own doubts, then recommends the price anyway

Several answers said the price might be out of date, and still put that store first.

• Kindle Paperwhite: the answer labelled Bing Lee's $247 and Officeworks' $249 an “older listing from August”, then told shoppers to check those two stores first. They were charging $369 and $367.
• Dyson V8 Advanced: it said it “couldn't confirm that this price is still available today” and made Bunnings at $319 its top pick. Bunnings was charging $382.
• Estée Lauder Advanced Night Repair: it said it “couldn't independently verify” Beauty Amora's checkout price and told shoppers to “treat it as a lead”, then recommended it first. Beauty Amora was charging $97.65, not $81.05.
• PlayStation 5: it called eBay's $459 “a price-comparison listing rather than a confirmed current checkout price” and recommended it first. We couldn't find a new Digital Edition listed anywhere near that price.

Where the wrong numbers came from

We can't see inside ChatGPT, but the answers show their sources, and the wrong prices follow them. They came from sites that copy retailers' prices, and those copies were often weeks old.

Third-party mentions help a store get recommended, but the price in the answer should come from the store's own site and feeds. The stores ChatGPT recommended were the ones comparison sites and deal posts cover, and the prices it quoted often came from those same sites. Asked where something is cheapest, it picked an out-of-date copy of a price when the current one was on the store's own site.

The ads under the answers

ChatGPT now shows ads under shopping answers, and almost every answer in our test had one.

Twice, a Bing Lee ad appeared directly below an answer that misquoted Bing Lee's own price: a Kindle Paperwhite at $247 when Bing Lee was charging $369, and a Ninja air fryer at $142 when it was charging $199. Some ads showed the same product as the answer: a MECCA ad showed the ghd Platinum+ at A$439 under an answer recommending it at $299 from Shaver Shop. Others matched only a word. Under the LEGO Orchid answer, the ad was for artificial potted orchids.

The ad sits below the answer and doesn't change it. OpenAI says product results “are not ads” and that “Ads are separate from product results” (OpenAI). Retailers who advertise early will get seen, but the price in the answer above the ad still comes from whatever source ChatGPT picks.

For how to approach the ads themselves, read ChatGPT Ads and eCommerce: What Merchants Should Actually Do Now.

What this means for your Black Friday marketing

Shoppers use the answer to decide where to go, then buy on the retailer's own site. Only 4% want AI involved at checkout, according to the same Commerce and PayPal Australia research, so your site is where they find out whether the answer was right.

Your sale price may not be the price in the answer

Sale prices reach your own site first. Comparison sites, deal posts and marketplace listings catch up later, and some hold onto old prices for weeks: the Garmin price in our test came from comparison data 40 days old. In our September study of 51 retailers' product pages, four of the six wrong prices were for products on sale, and the comparison-site prices we could check were from before the sale started. Black Friday puts most of a range on promotion, changes prices daily and ends offers at short notice, so the gap between your price and the answer's price is likely to be widest in the week that matters most.

A wrong price costs you in two ways

If ChatGPT names you at a price you don't charge, the shopper arrives expecting it. That turns into a price-match request, a live chat complaint or an abandoned cart, and the shopper may blame you as much as the chatbot. If it leaves you out, the shopper never visits, and the loss doesn't show up in your analytics at all.

Conditional prices get presented as the price

Group-buy floors, marketplace listings and members' offers can all end up in an answer as the price everyone pays. In our test, a group-buy floor was quoted as Crowdshop's listed price, and a third-party eBay listing was presented as a LEGO set it wasn't. If you run member pricing or marketplace listings over Black Friday, put the condition next to the price everywhere it's published.

You can pay for the ad and still be misquoted above it

Twice in our test, a Bing Lee ad sat directly under an answer quoting a Bing Lee price that Bing Lee wasn't charging. The shopper sees your brand in the ad and a price you don't charge directly above it.

What to do before Black Friday

You can't buy or write what ChatGPT says about your prices. You can measure it, and you can fix many of the sources it draws on. Start with the measuring, because it shows which sources need fixing.

Measure four things

Pick 10 to 20 hero products for the sale. Ask ChatGPT and Google where each one is cheapest, three times each in new chats, on a free account, and record the four measures. Do it weekly until the sale and daily in sale week. Our 15 copy-and-paste prompts are a ready-made starting set, and How do you measure AI visibility? covers tracking over time.

Track price accuracy alongside the recommended rate. In our test, Myer was recommended five times and priced correctly in two.

Your Black Friday timeline

What marketing controls

• Member, code and group-buy prices. On the page, in feeds and in any deal you brief to publishers, the condition goes next to the price, so a copy of it can't be read as the price everyone pays.
• End dates on every deal you brief out. Deal posts and trackers keep quoting a sale after it finishes. The $319 Dyson price ChatGPT recommended came from a deal reported 11 days before our test. A dated offer is easier for a reader, and for anything summarising the page, to recognise as expired.
• Marketplace listings. If you sell through a marketplace, make the seller name, the model and any conditions clear in the listing and the feed. ChatGPT presented a third-party eBay listing as a LEGO set it wasn't.
• Your front line. Expect “ChatGPT said it was cheaper” in chat, in store and in price-match requests, and agree the policy before the sale starts.

What marketing can influence, with the eCommerce team

• Comparison sites. If you supply a feed to a comparison or shopping site, find out how often it pulls the feed, and make sure sale prices reach it the moment they go live. If you don't supply one, check what those sites show for your hero products on launch day, because they're quoting you anyway.
• Feed timing. Every sale start and end date in the promotion calendar should include the moment your feeds and comparison listings change. Product Feed Optimization for AI Shopping covers the data work behind it, and Do I need a separate product feed for ChatGPT? answers the most common question about it.
• Direct product feeds. OpenAI says ChatGPT's shopping information comes from “first-party and third-party providers” and that price changes can take time to show. It also says Shopify stores are already connected through Shopify Catalog, and that other merchants can apply for direct feed access (OpenAI). If you send a feed, update prices whenever they change, including when each sale starts and ends, as OpenAI's feed specification asks. Google Merchant Center can update Shopping prices from your product pages' structured data, but Google warns this may not work for prices that change more than once a day (Google), so for Black Friday, update the feed itself.
• Product naming. Model, generation, size and variant belong in the title and the product data. The LEGO answer recommended a 547-piece set from another maker to a shopper who asked for LEGO 10311.
• Access to your pages. In our September study, one reason ChatGPT turned to third-party sources was that it couldn't use the retailer's own page. If bot protection is tightened for peak, check it still lets verified AI agents through to product pages. Why can't ChatGPT find my products? walks through where access usually breaks.

What nobody controls

Which source an answer picks, and whether the answer changes when someone asks again. That's why the measuring has to continue through sale week.

For why ChatGPT quotes the wrong price in the first place, and what our September study found about product pages, schema and stock, read We checked what ChatGPT tells shoppers about 51 retailers.

All of this depends on the price being the same everywhere it's published: the product page, the feeds, the comparison listings and the emails. That's an operational question as much as a marketing one, and it's how Core dna's eCommerce platform is built, with one source that the site, the storefronts and the feeds all read from. For groups running several brands, it's a platform decision.

About this test

• Two asks per product, one day, one account, one assistant. 40 answers on 20 October 2026. Prices change, and the figures describe what we saw that day.
• Every recommended store was checked on its own website, the same day, for the exact product asked about.
• Answers that named two stores at one price counted as wrong if either store's price was wrong. Two answers had one store right and one wrong.
• Members' prices, discount codes and group-buy prices weren't counted as the store's price. The Kogan gap is measured against Kogan's members' price, the lower of its two prices.
• One question was asked again in a new chat after it was entered with the wrong formatting. The first answer wasn't counted.
• Where ChatGPT offered two responses to choose from, we used the one kept in the chat.

About this study. The AI Retail Readiness Study is run by Core dna, the operations platform for multi-property operators. Our eCommerce platform runs product pages, storefronts and product feeds for retailers, and we run tests like this to see how our clients' products and prices are picked up by AI shopping tools, and where the gaps are between what a retailer publishes and what a shopper is told.