In May 2026, if you asked AI engines what a certain UK employee wellbeing platform cost, the answers varied by a factor of ten. Different engines quoted different figures, sourced from third-party marketplaces, stale reviews, or nothing at all. The company's own pricing page said, in effect, "contact us for a quote" - so the machines went and found numbers elsewhere, and the numbers were wrong.
In August 2026, I asked ChatGPT the same pricing question in three separate runs. All three answers were accurate. All three cited exactly one source: the company's own pricing page. The tiers were right - £8,000, £12,000, £30,000 - the £9.99 individual plan was right, the add-ons were right. The tenfold spread was gone.
What changed in between is the least technical fix in this whole series: the company published its real prices.
spread across AI answers
accurate, own page cited
§ Why "contact us" hands your pricing to strangers
Here is the mechanic that pricing-page debates always miss. The question "what does X cost" does not go unanswered just because you declined to answer it. Buyers ask it, and now they ask machines, and machines are built to produce an answer from whatever sources exist.
If your own page offers no number, the engine's source hierarchy simply moves down a rung: review marketplaces, old forum threads, a competitor's comparison table, a cached quote from 2023. You have not kept your pricing private. You have delegated it - to the least accurate and least current sources on the internet, with no correction mechanism. The May audit's tenfold spread was not an AI hallucination problem. It was a sourcing vacuum the company had created itself.
The moment a real number exists on the official page, the hierarchy inverts. Engines prefer the primary source when one exists; in the August runs, the official page was not just cited, it was the sole source in every run. One page, published once, ended the guessing across the entire question.
§ The loop closed further than citation
The second measurement made this more than a citation win. The client's analytics showed an AI-assistant referral channel forming - sessions arriving from chatgpt.com and claude.ai - and the pages those sessions landed on were the homepage and, repeatedly, the pricing page. Half of those sessions were engaged visits, on par with the site average.
Read the full loop: a buyer asks an engine about cost, the engine quotes the real prices from the official page, and some of those buyers click through to the page that answered them. Cited, then clicked. The pricing page went from a dead end that machines routed around to the top of an acquisition path, and each stage of that loop is separately measurable - the citation in sweep runs, the referral in analytics.
The honest caveats travel with the numbers: AI referral counts are a floor, because referrer stripping undercounts them, and this channel is currently small - a dozen labeled sessions a month against hundreds from organic search. The point is not the volume. The point is the direction and the mechanism, both of which were at zero while pricing was quote-gated.
§ What this does and does not prove
It does not prove every company should publish prices. There are sales-motion reasons to quote-gate, and that is a commercial decision, not an SEO one.
What it proves is narrower and more useful: the cost of quote-gating now includes a visibility term that did not exist five years ago. Machines answer the pricing question either way. The choice is whether they answer it from your page or from the internet's memory of you. Whoever makes the pricing-page decision should make it knowing that.
And one warning from the same engagement, big enough to earn its own case study (#37 in this series): when you do publish prices, publish them everywhere machines look. This site's visible page carried the new prices while its structured data still said "pricing on request" - the invisible metadata layer contradicting the visible page ChatGPT was quoting. Machines read both layers. Update both, the same day.
§ The checklist
- Ask the engines what your product costs. Multiple runs, not one. Note every figure and every source. If the figures vary or the sources are not your page, you have found the vacuum.
- If prices are publishable, publish real numbers on your official pricing page, in plain text a crawler can read without executing anything.
- Update the page's structured data in the same release. Visible price and schema price must agree, or you have traded one contradiction for another.
- Re-run the same pricing queries a few weeks later, again in multiple runs. Accurate + own-page-cited in every run is the exit criterion.
- Watch your analytics for AI-assistant referrals landing on the pricing page, and treat the count as a floor. That is the cited-then-clicked loop, and it is the number that makes this legible to a CFO.
Numbers dated as of the May and August 2026 audits: pricing citations from 3-run scraper sweeps verified against the live page HTML, referral data from the client's analytics over a 30-day window. The site is anonymized as a matter of client confidentiality.
The one-line version: AI engines were never going to stop answering the pricing question. This company just took over the answer - by giving the machines the one source they wanted all along.