We asked ChatGPT and Perplexity what 20 business tools cost, and logged every source they cited. In four out of five citations, the answer points back at the vendor’s own pricing page — the page whose headline number is a marketing decision.
A pricing page is written to sell. The number in large type is usually the annual rate, the promotional rate, or the rate for a plan most buyers will outgrow — and the conditions attached to it sit in smaller type, in a footnote, or on a different tab. That is not deception; it is marketing, and every vendor does it.
The problem is what happens when an AI engine answers “what does this cost” by reading that page back to you. The framing survives the summary and the conditions do not. When 79% of the sources behind those answers are the vendor themselves, the answer inherits the vendor’s framing almost by construction.
| Engine | Answers | Without a source | Share |
|---|---|---|---|
| Chatgpt | 20 | 1 | 5% |
| Perplexity | 20 | 14 | 70% |
An answer with no citation is not necessarily wrong, but it cannot be checked. We saw one in our own test run: asked what SaneBox costs, Perplexity returned a plan name that does not exist, a price that does not match any tier, and a trial length that was wrong — with an empty source list.
| Domain | Citations |
|---|---|
| kinsta.com | 27 |
| payoneer.com | 25 |
| easydmarc.com | 18 |
| reclaim.ai | 17 |
| hubspot.com | 15 |
| support.pipedrive.com | 14 |
| meetgeek.ai | 12 |
| pipedrive.com | 9 |
| help.reclaim.ai | 9 |
| webflow.com | 8 |
| close.com | 7 |
| airwallex.com | 5 |
The pattern holds across the top of the table: the vendor’s own domain, followed by the vendor’s own help centre. 91 distinct domains appear in total, but the distribution is heavily concentrated.
This is one run of 40 answers across 20 tools and two engines, in a US locale, on a single day. It is enough to show a pattern and not enough to quantify it precisely. We are publishing it because the direction is stark and the method is cheap to repeat — not because 40 answers settle anything.
We did not test whether the answers were correct. Measuring that at scale would mean re-verifying every quoted price by hand, which is the work our own corpus already does for 20 of these tools, but not on the engines’ answers.
And the result that is least flattering to us: this site was cited 0 times. We publish verified prices with reading dates, and none of that reached these answers. That is a finding about our own visibility, and it belongs in the report.
Queries of the form [tool] pricing for 20 business tools we research ourselves, submitted asynchronously to the Cloro API against Chatgpt and Perplexity, US locale, on 2026-08-19. Every response was stored with its full source list. Counting rule: a citation counts as “the vendor’s own site” when the registrable domain matches the tool name, so a vendor help centre on the same domain counts as the vendor.
Raw responses, including the source lists: available on request and re-runnable from the method above. The underlying pricing corpus is published as an open dataset at /data/ under CC BY 4.0.
AIBuilder Marketplace, “Who do AI engines cite for software pricing?”, Report 02, 2026-08-19. Free to quote and reproduce with attribution (CC BY 4.0). If you want the raw response set for your own analysis, get in touch.