A buyer asks ChatGPT what your business account costs. The answer comes with a link, but not to your pricing page. It points to a PDF of your fee list that someone uploaded to a document-sharing site, and that fee list is from January 2023.
That's a real pattern in our August 2026 data, and it grew fast.
The short version: On 13–14 August 2026, ChatGPT started citing Scribd, a site where people upload and share documents, about four times as often. Many of the cited files were old company fee lists, manuals and offers. Almost every brand we track was affected. Check which documents ChatGPT cites for your buyers' questions, and make sure the current version lives on your own site.
What changed in August?
In internal Genezio research, we tracked ChatGPT's answers to 3,179 prompts that got an answer on every day of August 2026. These were simulated, persona-based conversations, and the brands skew Romanian.
On 12 August, 1.7% of answers cited Scribd. Two days later, it was 7.0%.
The jump stuck, though we don't know what caused it. On 31 August the share was still 6.7%.
It wasn't a handful of brands either. Among the 103 brands with enough answers to compare (more than 200 each), the Scribd share rose for 93. After 14 August, Scribd showed up in at least one answer for 105 of the 114 brands we track.
What this means: The sources ChatGPT uses for your category can change for almost everyone at once, in a couple of days. An audit you ran in July may already be out of date.
What kind of documents is ChatGPT citing?
Mostly PDFs that users uploaded: 98% of the Scribd links went to uploaded documents. And the most-cited ones were company paperwork about money and features.
At the top of the list after 14 August were a bank's fee lists for business customers. One was dated June 2025, another January 2023. Next to them sat copies of the same bank's mobile app manual and a telecom operator's phone and instalment offers from early 2025.
These are exactly the documents buyers ask about: "What's the monthly fee?" or "Can I pay in instalments?" When the cited source is a 2023 fee list, the answer may lean on 2023 prices. (That's our reading; we measured citations, not the prices in the answers.)
Is it your copy, or someone else's?
It can be either, and they're different problems. We looked closely at the cited files for 12 heavily affected brands, and they split evenly into two groups.
For six brands, ChatGPT mostly cited copies of their own documents: fee lists, manuals, offers and data sheets, many of them outdated. Here, an old version of you is speaking for you.
For the other six, ChatGPT mostly cited other people's files: competitors' fee lists, industry reports or regulations. Here, a third-party site has become a competing source for your buyers' questions. Our reading (not something we measured): in some of those answers, your own pages may not show up at all.
Does an old source mean a wrong answer?
Not always, and we can't prove it did in our data. But outside research says the risk is real.
ChatGPT looks pages up and then writes its answer from what it found. HoH, a 2025 benchmark that tests exactly this setup, reports that outdated information substantially reduces answer accuracy by distracting models from the correct information. It can also push models toward potentially harmful answers, even when current information is available. A Google Research paper on conflicting sources adds that models "often struggle to appropriately resolve conflicts between sources."
Think of a colleague handed two undated versions of a price list: sometimes they'll quote the old one.
The mix also keeps moving. In separate internal research across the AI assistants we track, we found that 78% of newly cited pages weren't cited again by the same prompt within the next two weeks. So a single check, including ours, is only a snapshot.
Your playbook
- Ask your buyers' questions, repeatedly. Use real questions about fees, plans and features, in each market and language you sell in. Run them over several days, because one answer is one sample.
- Open every cited PDF. Check who published it and its date. Sort each one into "old copy of ours" or "someone else's document."
- Publish the current version on your own site. Put prices, fees and terms on an easy-to-find HTML page with a visible "last updated" date. Make sure robots.txt doesn't block OpenAI's search crawler from it.
- Deal with the copies. If old copies of your own documents are being cited, ask the hosting site to remove the ones you have the rights to. If competitors' files answer your buyers' questions, answer those questions clearly on your own site.
- Re-test every month. This shift took two days, so a once-a-year check will miss the next one.
If you want to know which documents AI assistants cite when your buyers ask about your prices, start with steps 1 and 2. Genezio runs this kind of repeated, persona-based check if you'd rather not build it yourself.