Published October 27, 2026
A shopper asks ChatGPT about your best-selling jacket and is told it costs $40 less than it does, comes in a color you dropped last season, and ships free when it does not. The shopper either arrives expecting a price you do not offer or never arrives at all.
This explainer covers why AI assistants get product facts wrong, where the wrong facts come from, how to find them, and what a brand can and cannot fix, as of 22 October 2026.
ChatGPT and other assistants give wrong prices or outdated details mainly because they build answers from sources they retrieve, and some of those sources are stale, conflicting or unreadable to their crawlers.
Microsoft described the problem plainly in a May 2026 post on how Bing grounds AI answers. "In grounding, a stale fact produces a misleading response," it wrote, and a system that silently chooses between contradictory sources "may confidently assert the wrong thing."
Product facts in ChatGPT come from both crawled pages and merchant feeds, and the balance moved in 2026. Profound looked at 812,190 product cards ChatGPT showed in June 2026 and found that 87.3% came from web crawls, then reported that from 10 July most Shopping picks came from feeds instead. Either way, a page a crawler read weeks ago or a feed that was not updated can still shape today's answer.
No independent study measures product errors across assistants, but several vendor datasets point the same way, and pricing comes out worst in each.
| Source | What was checked | Finding on prices |
|---|---|---|
| Searchable, data to July 2026 | 32,556 AI-stated facts graded by 1,704 brands | Pricing had the highest false rate, 13.5% |
| Profound, data to August 2026 | More than 158,000 claims from its customers | Pricing was 12% of claims but 24% of inaccurate ones |
| Infatica, September 2026 | 826 claims about 45 products against live pages | 62.5% of price disagreements understated the price |
These are vendors that sell monitoring tools, and their samples are their own customers or small product sets, so read them as direction rather than rates. They also measure different things and cannot be combined into one figure.
Stock status is the other weak point. In Infatica's test, an "out of stock" claim was wrong 43% of the time, against 1 in 20 for "in stock". Products with many variants were the least accurate, with price agreement of 68.5% against 97% for products with a single price.
Wrong facts come from several kinds of source, and a brand controls only some of them.
A brand's own pages are a frequent cause. Profound found that 54% of brands had at least one inaccurate AI claim citing their own site. Likely culprits include old product pages left live after a relaunch, sale prices still shown in copy, discontinued variants and policy pages that were never updated.
Pages that only work with JavaScript add a second problem. In a 2024 log study by Vercel and MERJ, none of the major AI crawlers, including OpenAI's and Perplexity's, rendered JavaScript. If a price, size chart or review score loads after the page, an AI crawler may see an older value or none. Neither company documents its rendering, so treat this as the best evidence available rather than a rule.
Google Merchant Center requires that price and availability match the landing page, structured data and checkout, and its help pages name prices added by JavaScript as a cause of mismatches. When a feed, the markup and the visible page say different things, every system reading them has to guess.
Third-party product pages go stale on their own schedule, and some brands see them as a major source. RoC's CMO told Modern Retail that retailer product pages are "a big source" of generative AI search traffic for the brand. Profound found that errors citing earned media spread across a long tail of sites, with the top ten domains behind only 13% of them.
Ask the assistants directly about your products, the way a shopper would, and record what they say against what is true.
Start with hero products and the questions that affect a purchase.
Run each question in ChatGPT, Google AI Mode, Gemini, Perplexity and Copilot. Log the assistant, the question, the stated fact, the true fact and the cited URL. The cited URL is the most useful field, because it shows which source to fix. Repeat monthly, since answers change between runs.
Fix what you own first, then work on what others publish.
Some brands are rebuilding product pages for AI visibility. Borghese told Modern Retail it was "scrubbing all of our formulas to get really factual information" and rewriting product copy, after finding its top product hard to surface in AI search. When the issue is products missing altogether rather than wrong, how to get your products into ChatGPT and Google AI Mode covers the setup.
A brand cannot edit an AI answer directly, choose which sources an assistant cites, or set when a corrected page is read again. We found no brand correction channel at OpenAI, Google accepts Knowledge Panel suggestions without promising a change, and answers vary from one run to the next.
What a brand can do is make the correct fact the easiest one to find and the same across every source it controls or can influence.
We are a performance marketing and AI search agency, and for DTC brands a large part of our AI search work is tracing a wrong product fact back to the page or feed it came from.
Get a clear picture of where your store is losing revenue, what to stop now and which actions to take first with our DTC Growth Audit.
DTC Growth AuditNot directly. ChatGPT users can flag a response, but there is no brand correction channel. Fixing the source pages and feeds is the most dependable route.
No platform gives a timeline. The answer changes once the assistant retrieves the corrected source, which depends on when the page or feed is next read.
It helps when it matches the page. Markup that disagrees with the visible price or the feed adds another conflicting source.
Mike Le
Co-Founder, CB/I Digital and Conative AI
Mike Le co-founded CB/I Digital in New York in 2007 and has helped brands grow through search, paid media and data ever since. He started the agency's AI team in 2019 and later started Conative AI. He writes about how customers find and choose brands now that AI answers sit between them and the search results.
Follow Mike on LinkedInCopyright © 2026 CB/I DIGITAL INC. All rights reserved | Privacy Policy