Why AI Search Can Misstate a SaaS Price—and How to Check

An AI search answer may give a SaaS price without the condition that makes it true: billing term, seat count, currency, region or plan tier. It may also…

An AI search answer may give a SaaS price without the condition that makes it true: billing term, seat count, currency, region or plan tier. It may also repeat an old offer or name a figure the vendor does not publish. To assess the answer, compare its exact claim with a first-party offer for the same purchase conditions. Keep a dated copy of both.

Why a plausible price can still be wrong

“The Team plan costs X per month” could mean a monthly subscription, a monthly equivalent paid annually, a per-user charge or a base account fee. Those are different offers even if the number matches. A result for one country may also appear in a search made from another.

The US National Institute of Standards and Technology describes confabulation in generative AI as the confident presentation of erroneous or false content; it also notes that an output can contain a misleading citation. Open any cited page and inspect the relevant passage. A disputed price is not necessarily invented by the model: an older page, ambiguous table or mismatched billing unit may explain it.

Capture what the assistant actually said

Save the complete answer, including the sentences around the price. Record the date and time with a time zone, the search or assistant product, any displayed model version, locale and country settings, and the exact prompt. Keep the destinations of any cited links and note which part of the answer each appears to support. A screenshot preserves presentation; copied text makes the wording searchable.

Transcribe the plan name and qualifiers exactly. “Business” and “Business Plus” are different tiers. “From”, “per seat” and “billed annually” change the meaning of a figure. Note whether the assistant gave a firm amount, range, estimate or “quote required”. If it answered a question about the current offer with an older figure, keep the question: the error may be about time rather than arithmetic.

Capture the first response before asking a follow-up. A second prompt may reveal what the assistant thinks its source or billing unit was, but it changes the conversation. Record tests in other locales or fresh sessions separately; combining them can conceal inconsistent answers.

Compare the same offer on the vendor's site

Open the vendor's pricing page directly. Record its URL, access time, selected billing toggle and nearby footnotes. Follow first-party explanations of usage allowances, seat rules and regional terms. If a dated announcement explains a price change, record its publication date separately from your access date. An undated page shows what was visible when checked; it does not prove when the offer began.

Compare the exact plan, buyer type, country, currency, billing period and unit of sale. A monthly equivalent under an annual commitment is not automatically a month-to-month price. A per-user amount is not a whole-account total. A base subscription may exclude metered use. GitHub’s billing documentation, for example, distinguishes fixed plan costs, additional subscriptions and metered usage. That illustrates why one price field may not describe a complete bill; it does not establish another vendor's terms.

Check whether the tier actually fits the buyer's needs. Seats, usage limits, feature access and eligibility can matter more than the headline amount. If the relevant tier requires a sales quote, record “quote required” rather than borrowing an estimate from another site. Label negotiated or logged-in prices as account-specific and keep them separate from the public offer. Note tax treatment only if the page states it.

Do not treat matching currency symbols as proof of a match. The same symbol can represent different currencies, while a converted search result may use a different exchange rate or date from the vendor's displayed amount. Write down the currency code when the vendor supplies it. If the vendor offers regional selectors, record the selected region and check whether switching it changes the plan, billing term or price. Avoid converting the figure yourself unless the comparison explicitly calls for an estimate; a converted estimate is not the vendor's quoted offer.

Describe the discrepancy precisely

  • Stale: The assistant repeats an offer documented on an older first-party page or announcement, while the current vendor page shows different terms. Preserve both dates. A third-party page alone cannot prove when the vendor offered that price.
  • Wrong unit: The number appears on the vendor page, but the answer changes what it buys. It might call an annual per-seat equivalent a cancellable monthly account fee, or describe a usage rate as a flat subscription. Write the stated unit and documented unit side by side.
  • Unsupported: The figure or condition cannot be substantiated in the first-party material checked. Start with “unsupported in the checked sources”. Use “fabricated” only when the evidence supports the stronger claim; an inaccessible regional page or unpublished agreement may remain unexamined.
  • Unresolved: The vendor page is ambiguous or a relevant condition is missing. State what evidence would settle the question.

More than one description may apply. An older annual per-seat figure presented as a current monthly account fee is both stale and wrong-unit. The labels should identify the next check, not force a verdict from incomplete material.

Keep a comparison record

Use one record per assistant claim and one comparison record per vendor offer. Include:

  • The exact prompt, full answer, observation date, time zone, product and locale.
  • The quoted price sentence, plan name and every condition attached to it.
  • Each assistant-cited URL, the relevant passage and whether it supports that sentence.
  • The vendor URL, access date, any stated effective date, billing toggle and visible offer wording.
  • Currency, country, billing term, sale unit, seat minimum, included allowance and usage charges where disclosed.
  • The finding, evidence for it, remaining uncertainty and any later correction or recheck.

Write the finding as a sentence a colleague can verify: “The answer calls this an account-wide monthly fee; the vendor page labels it per user under annual billing.” That identifies the mismatch without inventing a replacement total. If seat count or likely usage is unknown, leave the resulting bill uncalculated.

When evidence conflicts, keep the competing records instead of silently choosing the newest-looking snippet. A search preview may omit a footnote that appears on the full page; a saved announcement may explain a former price but not the current offer. State which source answers which question. For a buying decision, use the terms shown in the vendor's actual purchase flow or supplied quote, and check them again before agreeing. For an editorial correction, cite the public page and the conditions visible there, without implying that private terms apply to everyone.

Correct the source and check the result

If you control the vendor page, first fix contradictory visible copy, pricing tables and linked explanations. Make the billing state clear and check that plan names and limits agree across relevant pages. Google’s SoftwareApplication documentation describes a price field for a paid app and recommends a currency field. If the page uses that markup, inspect it against the visible offer; it should not encode a number stripped of its conditions.

Log the erroneous wording, affected URL, change made, publication date and how the live page was checked. If the assistant cited a third-party page, record it separately and use its correction process where appropriate. Feedback submitted to a search or assistant product is a submission, not proof that its answer changed.

Later, repeat the original prompt under the recorded settings and save the new response. Google’s structured data guidelines say valid markup does not guarantee a rich result. A page correction likewise should not be reported as a corrected AI answer until a later observation shows it. Close the record with what the assistant said, what the matching first-party offer showed, and what remains uncertain.

Continue this line of research with QuotedWrong.