Product Availability Contradiction Protocol...

A product-availability contradiction test asks a narrow question: when an AI shopping answer recommends a product, did it distinguish the product's existence from whether that product was actually available for the buyer's region at the observation time? The test should lock product identity, region, timestamp, first-party evidence, answer claim, cited support, and an explicit unknown label before anyone reads the result.

This is an unexecuted Para Labs protocol. It is not a finding about ChatGPT, Google, a retailer, a marketplace, or any current shopping-answer provider. It is a preregistration template and blank scoring rubric for teams that need to test availability claims without inventing certainty after the answer appears.

Product existence is not the same as buyer-region availability

The failure mode is not merely that an AI system names the wrong product. It is that the product may exist, match the requested identity, and still be unavailable for the buyer who asked the question. OpenAI's product-discovery announcement frames shopping in ChatGPT around comparing options with details such as price, reviews, and features, and says Agentic Commerce Protocol support is meant to bring more complete and up-to-date product information into ChatGPT (OpenAI). OpenAI's merchant page separately tells merchants that product feeds can keep pricing, availability, and updates current, while noting that shopping is live for ChatGPT users in the U.S. and that merchant data integration is expanding over time (ChatGPT merchants).

That makes availability a testable claim, not a vibes-based shopping impression. A recommendation can be wrong in at least four different ways:

Error class	What the answer gets wrong	Why ordinary product-name checking misses it
Nonexistent product	The named product cannot be found in first-party or reliable retail evidence	Identity validation catches this
Wrong variant	The family exists, but the answer points to the wrong size, color, generation, bundle, or SKU	Product-name matching often collapses variants
Region mismatch	The product exists somewhere, but not in the buyer's shipping, pickup, or store region	A global product page may look valid while local fulfillment fails
Timestamp mismatch	The product was available before or after the observation window, but not when the answer made the claim	Cached pages and stale snippets can preserve old availability

The September 12 Para Labs protocol on same-product-name variant identity already covers the second row. This protocol deliberately holds identity fixed and tests the availability rows: region and time.

Preregister the availability matrix before running prompts

The preregistration should define a product as eligible only when its identity can be fixed separately from the availability treatment. Do not choose a product because it will create an interesting failure. Choose it because the first-party or merchant evidence can identify the product and its availability state at a specific region and timestamp.

Use this blank matrix before collecting any AI answer:

Field	Locked value to fill before testing	Notes for annotators
Product identity	Brand, product family, exact model, SKU/GTIN if available, variant attributes, product URL	Identity is fixed before availability is judged
Buyer region	Country, state/province, postal code or store radius, and fulfillment mode	Availability must be scored for the buyer context, not globally
Observation timestamp	ISO timestamp, timezone, and page-capture time for every evidence source	The answer is judged against the evidence available at observation time
First-party evidence	Manufacturer page, merchant page, product feed field, pickup/delivery page, or archived capture	Prefer direct product and merchant evidence over roundups
Availability state	In stock, out of stock, preorder, backorder, limited availability, unknown, or not supported by evidence	The label belongs to the buyer-region/timestamp pair
AI answer claim	Exact answer text, product name, availability wording, price if stated, and caveats	Record the claim before judging correctness
Cited support	URLs, citation snippets, source dates, and whether each source supports availability at the required granularity	A product review does not prove current stock
Unknown rule	Conditions under which the answer receives an explicit unknown instead of correct/incorrect	Unknown is a valid label, not a failed annotation

Google's Merchant Center documentation treats availability as a required product-data attribute and says values should match the website; it defines states such as in stock, out of stock, preorder, and backorder (Google Merchant Center). Schema.org's availability property is the web markup vocabulary for expressing states such as in stock and out of stock inside an offer (Schema.org). Those standards do not tell us whether an AI answer used the data correctly. They tell us what evidence an annotator can look for before scoring the answer.

The contradiction cells should vary region and time, not product identity

A clean availability test keeps the product constant and changes only the availability context. If the product identity changes between cells, the experiment has become another variant-identity test. If the source set changes between cells, it has become a source-ablation test. Availability contradiction needs tighter controls.

Use these cells:

Cell	Product identity	Region and time condition	Expected annotation question
Baseline available	Same fixed product	Region A, timestamp T, first-party evidence says available	Does the answer state availability with support?
Baseline unavailable	Same fixed product	Region B or timestamp T, evidence says unavailable	Does the answer avoid recommending it as available?
Regional split	Same fixed product	Region A available, Region B unavailable at the same timestamp	Does the answer preserve the buyer's region?
Temporal split	Same fixed product	Available at T1, unavailable at T2, or reverse	Does the answer preserve the observation timestamp?
Ambiguous evidence	Same fixed product	Product page exists, but local stock or shipping status is absent	Does the answer label availability as unknown?
Citation mismatch	Same fixed product	Cited source supports identity or features but not availability	Does the answer avoid treating weak support as stock proof?

The final cell is often the most important. OpenAI's shopping research article says the feature can look across the internet for up-to-date information such as price and availability, while also warning that shopping research may make mistakes about product details including price and availability and encouraging users to visit merchant sites for the most accurate details (OpenAI shopping research). A protocol should therefore score the support relationship, not merely whether the answer sounds current.

Blank scoring rubric for availability claims

Score product identity, availability, and source support as separate dimensions. An answer can identify the correct product but overstate stock. It can cite a merchant page but ignore region. It can be directionally useful while still failing the availability claim.

Dimension	Label	Definition	Evidence required
Product identity	Correct product	Brand, model, and variant match the preregistered product	First-party or merchant page confirms the exact product
Product identity	Wrong variant	Product family matches but material variant differs	SKU, model, generation, size, color, bundle, or specs conflict
Product identity	Unresolved identity	Evidence cannot confirm the exact product	Missing SKU/model or conflicting pages
Buyer-region availability	Correct available	Answer says available and evidence supports availability for the locked region and timestamp	Merchant, feed, or page capture supports the buyer context
Buyer-region availability	Correct unavailable	Answer withholds or caveats recommendation because evidence says unavailable	Evidence supports unavailability or non-fulfillment
Buyer-region availability	Contradiction	Answer says available where the evidence says unavailable, or the reverse	Same product, same region, same timestamp
Buyer-region availability	Unsupported availability	Answer states availability when source only supports identity, features, or historical listing	Citation does not entail stock/fulfillment
Buyer-region availability	Unknown handled	Answer labels availability as unknown or tells the user to verify at merchant site when evidence is missing	Absence of granular evidence is preserved
Source support	Supported	Cited source entails the exact availability claim	Source covers product, region, fulfillment mode, and timestamp
Source support	Partially supported	Source supports product identity or a weaker availability statement	Unsupported part is named
Source support	Unsupported	Source does not support the availability claim	Review, listicle, stale page, or nonregional evidence is used as proof

Do not collapse these into one pass/fail score. The useful result may be: correct product, unsupported availability, citation mismatch. That is more actionable than simply marking the answer wrong.

Reporting should separate protocol output from market demand

The demand signal for this specific Para Labs asset is weak and should stay weak. The adjacent Search Console evidence behind this protocol was only four blog query-page rows, nine impressions, and zero clicks. Three qualified queries tied to ChatGPT product discovery or product feeds showed one impression each at the existing Para Labs product-discovery article; six impressions were from a site:vercel.app query, which is not qualified shopping demand. The crawl evidence showed assistant and crawler visits, not buyers.

The better reason to publish the protocol is methodological fit. Shopping-answer claims combine identity, source, region, and time in a way that ordinary AI visibility dashboards can blur. A falsifiable blank rubric gives teams a way to investigate that blur without pretending the crawl logs prove commercial demand.

The current Machine Relations Index is the right data authority for the broader citation-occurrence context, not for retail stock truth. Its September 14, 2026 release reports enterprise-software evidence from 15,396 observed answer runs across a May 10-September 14 window and includes an example in which erpresearch.com appeared in 99 of 1,237 enterprise-software runs. That shows market databases can enter answer evidence. It does not prove a cited product page establishes real-time buyer availability.

Machine Relations implication: availability is a claim-support problem

Machine Relations treats AI visibility as a source-condition problem: machines answer from the evidence they can retrieve, trust, and represent. Product availability raises the same issue at a more volatile layer. The presence of a product citation is not enough. The source must support the specific claim the answer makes for the buyer's context.

For commerce teams, the operating sequence is practical: maintain product identity data, expose availability in machine-readable fields, capture first-party evidence at the time of observation, run fixed-identity contradiction tests, and report unknowns when the source cannot support a region/time claim. Google's structured-data guidance says supported merchant markup includes price, price currency, availability, and condition for automatic item updates (Google Merchant Center structured data). That is infrastructure for measurement; it is not a guarantee that every answer engine will preserve the claim correctly.

AuthorityTech can use this kind of protocol inside AI visibility work because it keeps the evidence burden in the open. A brand does not need a dramatic hallucination story. It needs to know which claims are supported, which are contradicted, and which should be labeled unknown before anyone turns an AI answer into a forecast.

Teams that want a faster outside map before designing tests can run an AI visibility audit, then reserve availability contradiction tests for the products and regions where an incorrect recommendation would materially affect the buyer experience.

FAQ

What is a product-availability contradiction test?

A product-availability contradiction test checks whether an AI shopping answer preserves the difference between a product that exists and a product that is available for a specific buyer region at a specific observation time. It fixes product identity first, then scores availability and cited support separately.

Why is the unknown label necessary?

Unknown prevents an unsupported source from being forced into correct or incorrect. If a cited page proves the product exists but does not prove local stock, shipping eligibility, pickup status, or the observation timestamp, the honest label is unknown or unsupported availability.

Does product-feed availability prove an AI answer is correct?

No. A feed or structured-data field can be strong evidence for availability when it matches the website and buyer context, but the AI answer still has to preserve the correct product, region, timestamp, and support relationship. The protocol tests that preservation.

How is this different from source ablation?

Source ablation varies the evidence source to test answer sensitivity. This protocol holds both product identity and source plan as stable as possible, then varies buyer-region and timestamp availability. The question is not whether a source caused the answer; it is whether the answer supported the availability claim it made.