Why AI answers can quote old SaaS prices

An old-looking SaaS price in an AI answer may be superseded, valid for a different offer, or impossible to assess from the available context. Preserve the answer and its date, align the plan and billing terms, then inspect the cited and current vendor sources. That sequence produces a finding your team can act on without guessing how the answer was generated.

You can start with a provisional QuotedWrong scan or an answer your team has already captured. Before investigating, gather the exact prompt, full response, observation time, visible platform information and any citation URLs. Keep the original capture even if the next answer changes.

1. Identify the offer and question date.
2. Capture the current first-party terms.
3. Read the source actually cited.
4. Classify the evidence and choose an action.
5. Repeat a defined sample after recording changes.

1. Identify which offer the answer describes

Begin with the prompt’s time frame. “What did Team cost in 2024?” and “What does Team cost today?” require different source comparisons. If the question lacks a date, record that limitation. Do not evaluate a historical answer exclusively against the price for new customers today.

Next, align the company, product and plan. A renamed plan can carry different features or eligibility. A legacy customer may retain terms that are no longer offered to new customers. A promotional price may be accurate only during a stated period. These are alternative offers to investigate, not assumptions that make every difference acceptable.

Record currency, unit, market, payment cadence and commitment. A monthly equivalent displayed under an annual selector can look smaller than a month-to-month price without being older. The currency and billing comparison guide gives a worksheet for these fields. Complete that comparison before assigning a freshness problem.

This step is finished when you can write a precise target, such as “Team, new customers, US offer, USD, per seat, annual prepayment, as displayed on the captured pricing page.” If two offers remain plausible, list both and carry the ambiguity forward.

2. Capture the applicable first-party source

Open the vendor page in the relevant market and select the billing option that applies. Save the visible plan text, source URL and capture time with a timezone. Include a screenshot when the selector state affects interpretation, but preserve readable text as well so a reviewer can compare the amount and terms directly.

Distinguish a capture date from an effective date. A capture proves what your recording shows at that moment. It does not prove when the price became available. A dated change announcement or versioned plan documentation may help establish an effective date; a generic “updated” footer does not establish which sentence changed.

If your company has internal release records, use them to investigate the timeline, then identify the public material a buyer could use. Internal approval of a new price and publication of that price are separate events. A draft pricing card in a design file is not evidence that the public page displayed it.

Do not overwrite the earlier capture with the new one. Give each snapshot a stable identifier and record the relationship between them. A reviewer should be able to inspect the before-and-after material without relying on your memory of the change.

3. Read the source the answer actually cites

Open every citation that appears to support the price. Find the relevant plan and amount within the page, and note whether the cited passage covers billing conditions. If a citation is inaccessible, record the access failure. An inaccessible source creates a verification gap; it does not establish that the answer invented the price.

OpenAI’s search documentation, accessed 26 September 2026, describes citations in search-supported answers. A visible citation gives you a source to inspect. It does not establish that every nearby sentence is supported or reveal every influence on the answer.

Compare three things separately: what the answer says, what the cited source says now, and what the current applicable vendor source says. If you have a dated capture of the cited page from the observation time, add that as another piece of evidence. Without it, avoid asserting what the cited page contained when the answer was generated.

4. Assign one of four diagnosis outcomes

This matrix is an editorial workflow for vendor review. Its labels describe the evidence available for a claim; they are not a diagnosis of a model’s internal mechanism.

An unsupported answer is not automatically an incorrect amount. The same price might appear in another valid source. Conversely, a citation that accurately repeats an old page can still fail to answer a question about the current offer. Keep factual alignment and citation support as separate columns in the review.

A hypothetical timeline, with five different dates

The following dates and prices are invented for explanation. This example was prepared on 26 September 2026; it is not a customer finding.

• 1 September, 12:00 UTC: ExampleCo publishes a new-customer Team price of USD 30 per seat per month, paid monthly.
• 2 September, 09:00 UTC: A vendor reviewer captures the new pricing page.
• 10 September, 14:00 UTC: A sampled answer quotes USD 25 for new customers and cites a comparison page.
• 11 September, 10:00 UTC: The reviewer inspects that comparison page and finds USD 25, but has no capture from 10 September.
• 26 September, 09:00 UTC: A repeat sample quotes USD 30 after the vendor has clarified its own pricing text.

The first sample has an amount that conflicts with the supplied current-offer evidence. The citation inspection supports a narrower observation: the comparison page still displayed the older amount when inspected. The repeat sample is consistent with the new price, but it does not prove the vendor’s edit caused the change or that future answers will match.

5. Repair material you control, then retest

Check your pricing page, plan documentation, comparison pages and public help articles for conflicting current-offer statements. Preserve historical documentation when it remains useful, with a clear date and applicability label. Removing all old prices can make legacy customer support harder without establishing that assistants will refresh faster.

Give the same offer a consistent description across owned pages. Keep amount, currency, unit and payment terms together. If a third-party page is inaccurate, prepare a source-linked correction request for your team to approve. A draft request is not a sent correction, and a sent correction is not evidence that the publisher changed the page.

Google’s recrawl guidance, accessed 26 September 2026, says a request does not guarantee immediate inclusion or inclusion at all. That guidance concerns Google; it provides no timetable for a ChatGPT answer to change.

For the repeat sample, reuse the question, locale and collection surface, and record any unavoidable differences. Report “the later sample matched these fields” rather than “the problem is fixed.” The QuotedWrong method keeps provisional sampling separate from human-reviewed findings.

Questions about old AI prices

How often does ChatGPT update data?

There is no single refresh schedule you can apply to every pricing answer. Learned information, supplied context and search-supported responses are different inputs. Capture the observed answer and its sources instead of assigning it a universal update date.

Can I tell whether the price came from training data?

Usually the answer itself does not establish that. A missing citation does not prove a training-data origin. The guide to pricing-answer evidence explains which observations you can preserve and which causal explanations should remain hypotheses.