Same shopping question, three runs: how much AI answers change, and whether their prices are right

Published October 22, 2026 · Data collected October 22, 2026 (three runs, one hour apart) and October 17, 2026

Key findings

• ChatGPT changes its sources the most. Only 15% of the sites it cited appeared in all three runs, and no question got the same set of sources three times. Perplexity cited the same sites in all three runs for most questions (96% of sites in all three); Google AI Overviews sat in between (55%).
• The averages hold even when the sources move. ChatGPT cited a brand’s own site in 49–64% of answers across the three runs, against 57% on October 17. The share is a steadier number than the list of sites behind it.
• Prices were mostly right, and the errors mostly did not repeat. Across 57 “where to buy” answers, 5 of the 55 prices quoted for the brand’s own store were wrong: a sale price that did not exist, a wrong sale price, and a bottle and its refill mixed up. The same answer on the next run was usually right.
• Stock errors did repeat. For one product, ChatGPT and Perplexity said it was sold out in all three runs (6 of 6 answers). The store showed it in stock each time.
• The assistants choose the seller. In 9 answers an assistant named another seller as the cheapest or best deal, such as Target, a Walmart marketplace seller or Sam’s Club. Once, a Google AI Overview called an unverified .shop site the brand’s official store.
• Price level matters, in opposite directions. ChatGPT cited brand sites more often for expensive products (29% of answers under $100, 75% at $400 and up). Google AI Overviews did the reverse (81% and 31%). These are small groups.

Part 1: the same 60 questions, three runs

Share of answers that cited at least one site of each type, as the mean of the three runs on October 22 with the lowest and highest run in brackets, and the single run on October 17 below.

How much the cited sites moved within the day:

Two examples of how much one answer can move in an hour:

• “Best tablet for drawing for beginners” in ChatGPT: 16 different sites across the three runs and none in all three. The first run cited review sites, three tablet brands and a retailer; the second, two retailers and two other tablet brands; the third, review sites, a retailer and two tablet brands.
• “Best carry-on luggage for international travel” in Google AI Overviews: the first overview cited review sites, Reddit and social posts; the second cited one luggage brand; the third, Quora, a travel site and a different luggage brand.

By typical price of the product

Share of answers citing a brand’s own site, by the typical price of the product in the question (software questions left out), as the mean of the three runs.

Part 2: “Where can I buy it, and how much is it?”

Eight specific products sold on their brands’ own US stores, from a kitchen tool to a hard cooler, asked three times in each engine. At the start of each run we saved the store’s live price and stock; none changed during the day. Google showed an AI Overview for only 9 of the 24 questions, so its column is small.

What went wrong, run by run:

• A kitchen tool: a Google AI Overview put the brand store at a sale price about 25% below its real, undiscounted price. The overview did not appear in the next two runs. ChatGPT priced the store correctly every time but told buyers to check a marketplace seller first in two of three runs.
• A hard cooler: ChatGPT quoted a sale price about 30% above the store’s, then the right price in the next two runs.
• A hand soap sold as a bottle and a refill: ChatGPT mixed up the two prices in 2 of 3 runs; Perplexity once.
• Bamboo socks: ChatGPT and Perplexity said both sizes were sold out in every run, while the store showed them in stock. Google’s overview, when it appeared, listed other sellers and not the brand’s US store.

We cannot tell why the stock error stuck. One thing we noticed: the product page carries a hidden “Sold out” label next to the price, as many store themes do, while its structured data says in stock. Two other stores in the test had the same hidden label and were not called sold out, so this is something to check, not an explanation.

What this means for brands

• Do not judge from one run. In ChatGPT especially, a single answer says little about which sources matter. Ask the same question several times and track the share of answers that cite or recommend you.
• Ask the question closest to the sale. “Where can I buy it, and how much?” shows where buyers are sent and what they expect to pay. In our test the assistants often pointed to marketplace sellers, and sometimes to the wrong price.
• Make price and stock easy to read correctly. Keep variant prices clear (a bottle and its refill are two prices), keep sale prices current, and check what your product page says about stock to software, not only to shoppers.
• Treat Perplexity as stable and ChatGPT as a moving target. A source Perplexity cites today is likely there tomorrow; in ChatGPT, being cited is a share you earn over many runs.

Method and limitations

Part 1: the same 60 shopping questions as our Google study and assistant study, asked three times on October 22, 2026, at 9:45, 10:45 and 11:45 Pacific time, with the same settings: Google US desktop with the AI Overview loaded; ChatGPT through chatgpt.com search for a US location (DataForSEO LLM Scraper); Perplexity through the Sonar API with US web search (DataForSEO). Citations are counted as in those studies, and site types use the same scheme, assigned by one person. Percentages use answers that cited at least one source; for Google, questions where an overview appeared. Overlap compares cited sites by domain.

Part 2: eight products on the brands’ own US stores that publish their prices, asked as “Where can I buy [product], and how much is it?” in the same three runs. We saved each store’s price and stock at the start of each run and labelled every answer by hand: whether it named the brand’s store, whether the price it gave for that store was right, what it said about stock, and whether it named a cheaper seller as the best deal. Prices given for other sellers were not checked. A price counts as right if it matches any current variant, or the lowest one when the answer says “from”. We do not name the products, because several belong to brands we work with or talk to.

Limits: one day, three runs, one market; the products are a small convenience sample, not representative of all stores. The question lists are written, not sampled from real searches. API and scraper results can differ from what a logged-in user sees. The raw rows are in the CSVs: Part 1 and Part 2.

Where Hasmord fits

Hasmord is the AEO-native sales platform. Buyers ask questions across AI assistants, search, and sales conversations. Hasmord uncovers where your answers about your products and services fall short, turns evidence into better answers, finds potential customers for you, and connects every lead to your sales process.

It asks your buyers’ questions in ChatGPT, Claude, Gemini and Perplexity, reads a question’s Google AI Overview every day for a week with Question Monitors, and shows which sources and facts each answer relied on. See the AEO/GEO platform.

Frequently asked questions

Is ChatGPT accurate about product prices?

Mostly, in our test. ChatGPT gave the brand store’s price correctly in 21 of 24 answers. The misses were one wrong sale price and, twice, a bottle and its refill mixed up. It also said one in-stock product was sold out in all three runs.

Do AI answers change if you ask the same question again?

Yes, and how much depends on the engine. Only 15% of the sites ChatGPT cited appeared in all three runs on the same day, against 55% for Google AI Overviews and 96% for Perplexity. The share of answers citing each type of site moved much less.

Why would an AI assistant say a product is sold out when it is in stock?

We cannot see inside the engines. In our test the error repeated on every run for one product, so it is worth checking what your product page shows to software that reads its text: hidden stock labels, out-of-stock variants and structured data.

How often should brands check what AI says about their products?

More than once. A single run can be unrepresentative, especially in ChatGPT. Ask the same buyer question several times, track the share of answers that cite or recommend you, and check prices and stock on the questions closest to a sale.

Related: who ChatGPT and Perplexity cite · who Google’s AI Overviews cite · AI Overviews tracker · ChatGPT shopping.