How to investigate an outdated product price in an AI answer

A worked tutorial from Beacon by EVAA. Prepared 2 October 2026.

When an AI answer gives the wrong product price, preserve the answer and establish which offer applied at that moment. The model may have used an old page, confused a variant or selected another market. Your own source may also be wrong. Each explanation leads to different work.

This tutorial follows one fictional Singapore product from an old price claim to a documented correction plan. Every product, price, source record, answer and timestamp in the example is invented. No live AI system or customer was tested, and no successful correction or citation increase is claimed.

Start with an exact claim

The fictional question was asked at 10:00 Singapore time on 1 October 2026:

What is the current Singapore price of black Demo Lamp 20, variant DEMO-L20-BLK-SG, excluding delivery?

The fictional answer was:

It costs SGD 79.00, excluding delivery.

It cited the previous Singapore price record. The question supplies several parts of the claim: the black variant, a specific SKU, Singapore, the current offer and exclusion of delivery. Preserve that context alongside the short answer. A price is difficult to judge accurately after those qualifiers disappear.

Keep one row for this price claim. If the answer also states brightness, warranty and availability, give those separate rows linked to the same answer. They may have different sources and different outcomes.

Build the source timeline before choosing a label

The fixture provides these authoritative teaching records:

The new price took effect one hour before the answer. In UTC, the change occurred at 2026-10-01T01:00:00Z and the answer at 2026-10-01T02:00:00Z. Retain one unambiguous timestamp format in the record and display the buyer's local time separately when helpful.

The tax treatment is unspecified in the fixture. Do not add “tax included” because the price happens to be in Singapore dollars. The same restraint applies to delivery time, stock and promotion terms that the record never states.

This exact historical record is what supports an outdated label. A conflicting “79” headline without currency, market or a valid period would not establish that the answer used an old valid price. It might describe another offer or be unassessable with the available evidence.

Judge the fact and the citation separately

For the worked claim, the fact judgment is outdated: the current same-scope offer is SGD89, and retained history establishes that SGD79 previously applied.

The cited old record establishes the historical amount. It does not support its use as the current price at 10:00. Under the clarified companion rubric, record partial citation support and expired temporal applicability. Exclude it from any count of fully supporting current-price citations.

That separation also catches the reverse problem. Another fictional answer gives the correct SGD89 price but cites the product specification, which contains no price. The fact is supported by the commerce record; the chosen citation provides no support for that price. A correct number does not make the cited document correct.

Use a precise reason rather than a red or green score alone:

The amount matched version 11 for the same SKU and market, but its validity ended at 01:00 UTC. Version 12 set SGD89 from 01:00 UTC. The answer at 02:00 UTC asked for the current price. The old citation supports the historical amount only.

A reviewer can reconstruct that conclusion without knowing who wrote the answer or which tool collected it.

Check whether there is a source defect to fix

In this fixture, the current authoritative commerce record already says SGD89. Changing it again would not address an established defect. The old record is truthful historical evidence and should not be erased simply because a fictional answer used it out of context.

In a real investigation, inspect the surfaces that could be presenting the current offer:

• The product page for the exact variant and seller
• Any visible offer text and matching structured data
• The active catalogue or merchant feed
• Relevant support, campaign or retailer pages
• Historical pages that may be mistaken for a current offer

Record each source's URL or identifier, retrieval time, effective time and visible value. If an active page still says SGD79, that is a concrete candidate for correction. If the current sources agree on SGD89, investigate retrieval, citation choice and stale copies rather than inventing a catalogue problem.

Google's guidance for AI features recommends making important information available as text, keeping structured data consistent with the visible page and maintaining current Merchant Center information. It also states that inclusion in AI Overviews and AI Mode does not require a special AI file or special schema. Those are useful checks for Google Search eligibility, not a promise that a particular answer will change. Google Search guidance

Write a correction the source owner can review

A useful correction request names the affected record and the exact proposed change. For this fictional exercise, a proposed active-offer statement would be:

Demo Lamp 20, black, SKU DEMO-L20-BLK-SG. Singapore offer: SGD89.00, excluding delivery. Effective from 1 October 2026 at 09:00 Singapore time.

This is suggested teaching copy, not an implemented update or a complete legal price disclosure. The owner must confirm any missing tax, promotion, availability or commercial terms before public use.

If the issue is confusion between a historical page and the current offer, a proportionate proposal might be to make the archived period explicit and link to the current record. If an active merchant feed has the wrong number, correct the feed through its owner and verify that the landing page agrees. Keep those proposals separate; the fixture has not diagnosed either real defect.

For Merchant Center, Google documents automatic updates that can use landing-page data for price, availability and other specified attributes in Shopping ads and free listings. That is a bounded product-data mechanism. Enabling or fixing it does not establish that ChatGPT, Claude or every Google AI answer will adopt the same correction. Merchant Center automatic updates

Assign one accountable role to each proposed change. Product operations may own price truth, the web team the product page and a distributor its own listing. An agency should record whether it can implement a change or only recommend it. Do not mark a third party's page corrected because an inquiry was drafted or sent.

Keep the workflow state honest

A correction record should distinguish what has happened from what the team wants to happen.

A “resolved” review label means reviewers resolved the classification. It does not mean the product source was updated or the AI answer was fixed.

Preserve the original row after any later update. The new source version and new answer should be separate evidence, linked by stable identifiers. Otherwise, a later reviewer may be unable to tell which information supported the first decision.

Recheck the unchanged question after an approved update

When a real source change is approved and verified, rerun the same question under documented conditions. Keep the product, market and question text fixed. Record the assistant surface, visible location, sign-in state, search mode and any query rewrite. Retain unavailable answers separately.

The expected price in this fictional case is SGD89 excluding delivery. An expected result is only an acceptance target. Do not insert it into the answer ledger as though a model produced it.

Several outcomes are possible in real work:

• The answer changes to SGD89 and cites an appropriate current price source
• The number changes, but the citation still fails to support it
• The answer remains SGD79
• The answer avoids giving a price or cannot be collected
• A different product or market appears, making the comparison unsuitable

Record the actual outcome and its limitations. If an answer improves after a page change, other causes remain possible, including changes in the provider, sources or collection conditions. Repeat a fixed panel over a useful period before claiming durable improvement.

For a connected branded assistant, test the catalogue lookup separately from unconnected search. Verify which record it retrieves, how it handles an expired offer and what it says when the current price is unavailable. Passing that test does not establish independent search recommendations.

Avoid labelling every other problem “outdated”

The companion fixture contains seven cases to check the boundaries of the review method:

These outcomes depend on the supplied evidence. A missing page, inaccessible source or conflicting authoritative record can make a clear claim unassessable. A reviewed, defined corpus with no relevant support can justify “unsupported within this corpus.” Neither label should become a claim that the reviewer searched every possible source.

What the small calibration exercise established

Two separate automated reviewers received the seven shuffled claims, fact-only reference records and label definitions. The proposed answer key and the other reviewer's judgments were withheld. They agreed on all seven fact labels and six of seven citation labels.

The disagreement concerned the expired SGD79 record. One reviewer labelled it contradictory to current validity; the other labelled it partial support for a historical amount. Both found that it failed to support the current price. The clarified rubric now records historical support and temporal applicability separately. Both original judgments are retained.

This checks how two automated reviews handled seven designed examples. It does not establish human reviewer reliability, real-world product accuracy or a statistically representative agreement rate. No independent post-clarification review has been run, so the resolution is not reported as new seven-of-seven citation agreement.

Use the exercise to make review instructions clearer. For a real engagement, sample actual permitted answers, retain the source versions, obtain appropriate reviewer input and preserve disagreements before adjudication.

Finish with a record someone else can inspect

Before closing a real correction, gather the original answer, current and historical source evidence, classification reason, approved source-change receipt and the actual follow-up result. Mark anything missing as open.

For the worked price example, the review and proposed correction plan are complete, while source changes and AI remeasurement are unperformed. That is a useful result: the team knows which price applied, why the cited old record is insufficient and what evidence it still needs before claiming a fix.

Related reading

• Separate facts, completeness and source relevance
• AI citation measurement and denominators