Original research

Does AI recommend what you sell today?

A test of 48 answers shows how AI recommends a discontinued product, with the manufacturer’s website as its source.

Marc Seefelder 8 min read
YOUR PRODUCT PAGE AI RECOMMENDATION BABY SKINCARE RANGE FRAGRANCE-FREE NO LONGER AVAILABLE STATUS MISSING DOES NOT ARRIVE

One observed case · paraphrased

The answer was checked directly against the manufacturer’s page. This illustration summarizes the case; it is not a screenshot. [S1]

Three facts run from your product page to the AI recommendation. Product and property arrive. The status no longer available drops out on the way.

AI can recommend your brand while suggesting a discontinued product. In our test, eleven of 48 answers cited the manufacturer’s product pages. None of those eleven mentioned the “no longer available” notice. One recommended the discontinued skincare range as a suitable choice. For the brand, that initially looked good: it was found, explained and cited. For the customer, the information that changed the choice was missing. [S1]

01

We checked 48 AI answers about one skincare range

In September 2026, we put 65 buyer questions to two AI systems for a skincare manufacturer: Claude and Perplexity, both through their APIs with web search enabled. We asked every question twice, producing 260 answers. [S1]

This article focuses on one task: finding a fragrance-free skincare range for babies and children with sensitive skin, comprising a wash, lotion and nappy cream. We used twelve question variants: six without a brand name, three comparing competitors and three naming the brand. That produced 48 answers. [S1]

We also checked the manufacturer’s website during the same period. Four product pages for the range said “no longer available”. They had no buy button, named no successor and carried no structured product data, although they included a product identifier. [S1] We keep the brand anonymous. The finding concerns how availability is described, not product quality.

02

AI recommends the discontinued range and cites the manufacturer

Eleven of the 48 answers cite the discontinued product pages: ten from Perplexity and one from Claude. Five came from questions naming the brand, four from comparisons and two from questions without a brand name. None of these eleven answers mentions the status. In twelve further answers, the pages appeared only in search results, without a citation. [S1]

The recommendation matched the customer’s request. In one question without a brand name, we described the range as a customer might describe it to a shop assistant: a fragrance-free wash, lotion and nappy cream from one range. Perplexity suggested several suppliers and identified two ranges as the clearest matches. One was the discontinued range. Its answer explained which products went together and linked to the manufacturer’s product page. [S1]

If you were responsible for that brand, this would initially look good. Your company appears, the products are explained and your website serves as a source. These are the kinds of results that appear in AI visibility reports.

The cited page says “no longer available”. The answer does not. It presents the range as a suitable choice without mentioning the manufacturer’s status. None of the eleven citing answers includes that status, though this does not mean that all eleven make the same recommendation. [S1]

The one answer among all 48 that mentioned the status came from a question naming the brand. The manufacturer’s page appeared only in its search results. Instead, it cited two online pharmacies that, according to the answer, carried a renamed successor with the same product identifier. The manufacturer’s page named no successor. We did not confirm the retailers’ claim with the manufacturer, so its accuracy remains unresolved. [S1]

A separate agent test of the same task reached a similar obstacle: the page was readable and the status recognizable, but there was no purchase path, so completion was blocked. [S1]

We submitted test questions and reviewed the answers. We did not observe an actual customer making a purchase. The case nevertheless shows what a customer would have had to establish independently: whether the recommended offer was still available, and under which name.

03

A cited product page does not guarantee a correct recommendation

Your product page can be the source and still be misunderstood. The AI used information from the manufacturer and described properties useful for choosing a product. The limitation that changed that choice did not come through. The saved answer does not reveal whether the system read only an excerpt or lost the notice when summarizing. We can establish only that the page contained the notice and the answer did not.

Product status is a specific fact that changes a customer’s decision. Write it so that it remains understandable alongside the product name. Our suggestion is to keep status, date and a confirmed next step together. Whether AI carries that information through needs to be tested again. [S3]

The brand was found. An outdated offer was recommended. The checked manufacturer pages offered no purchase path; we did not investigate retailers’ remaining stock. Counting citations therefore leaves an important question unanswered. Citation Conversion measures how often a brand mention comes with a citation of its website. [S4] To establish whether that citation supports a correct recommendation, check the actual offer.

A recent study provides wider context. Researchers published their own posts on public platforms and monitored ten AI search systems. Of 1,004 answers, 17.5 percent cited one of those posts; 7.5 percent reproduced an identifying marker embedded in the text. Being cited and having content carried into the answer are different measures. [S2] The study does not establish whether a particular product recommendation is current. That requires checking the offer itself.

There are regular reasons to do this: a new product generation, changed package contents, an expired tariff or a new name for an existing product. Each change is a reason to look at the answers customers receive.

04

Explain what is current and what comes next on the product page

Consider the discontinued range. An old product page can still help people who already own the product. It needs to explain which version of the offer it describes and what people can do now.

Put the status next to the product name. If there is a confirmed successor, link to it and show both the old and new names. If there is none, say so and offer a contact route or other help. A similar name alone does not make a product a replacement.

Then reconcile the information wherever your company provides its offer: the product catalogue, retailer feeds, structured product data on the page and any function through which an AI assistant checks availability. Keep two facts distinct: the manufacturer has discontinued the product; a retailer may still have stock. Each should state its own current position.

We call the layer of content, data and functions an AI agent reads and uses the Agent Surface. [S5] Hyperize builds it. Our free Snapshot offers an initial look at what AI says about your offer: a buyer question, measured answers and a next step. Whether AI takes the status into account afterwards can only be checked by asking again.

The Universal Cart answer explains how to prepare your product data and purchase routes for Google's shared cart. How do I prepare my store for Google's Universal Cart?

Request my free Snapshot

05

Check a real customer question with your team

You do not need a complete picture of every AI answer about your brand to begin. Take a question from sales or customer service, ideally one where a person is trying to choose: Which product fits my needs? Which service do I need? Can I get it by a particular date?

Ask the assistant you want to test. Then work through five checks:

  1. Which product does AI mean? Read the answer as a customer would. Is the model, variant and name correct?
  2. Which source supports the decisive claim? Open it. A retailer page or independent review is not automatically a problem. Check that it concerns the right product and supports the claim.
  3. Is the claim still current? Check status, price, terms and time period against your current offer.
  4. Does the recommendation fit the need? Look for a missing condition that would change the decision.
  5. Can the customer or agent take the next step? Does the link lead to the right product? Can someone make an enquiry or select an available item? If an agent is meant to take the step, test that too, within the permissions granted. A link in an answer does not establish that an agent can use it.

You can now discuss the finding as a team. Brand leaders check whether the company is represented correctly. SEO and GEO teams identify which page is found and which information carries through. Agent commerce teams check whether the selection leads to the next permitted action. Everyone works from the same customer question.

06

After the change, test whether AI describes your offer correctly

Save the question and answer before making a change. Afterwards, ask the same question several times using the same settings. Check whether the missing information now appears and whether the next step can be completed. One better answer is a reason to investigate further, not evidence of a lasting result. [S6]

The AI in our example could still have mentioned the skincare range. A useful answer would have explained its status and offered a confirmed next step, such as a manufacturer-confirmed successor or a contact route.

That is where you can start: an offer that recently changed and a question your sales team regularly receives about it. Check whether AI tells a customer what you would tell them today.

Questions about checking your offer

Does AI recommend what I sell today?

Not reliably in our test. Eleven of 48 answers cited product pages for a discontinued skincare range. None mentioned the “no longer available” notice on those pages. A citation of your website does not establish a correct recommendation. [S1]

Is an external source in an AI answer a problem?

Not by itself. An independent review or retailer page may support a recommendation. Check that the source concerns the right product, the right period and the actual claim. In our case, one answer included retailers’ claims about a possible successor. We did not confirm those claims with the manufacturer. [S1]

How can I spot a recommendation for a discontinued product?

Open the source supporting the decisive claim. Check its status, name and product identifier against your current offer. Follow the proposed next step: does the link lead to something a customer can obtain today?

How do I check whether a change helped?

Save the question and answer before the change. Ask the same question several times afterwards with the same settings. Compare whether the missing fact appears and the next step works. [S6]

Sources and study scope

Measured on 25 September 2026: one anonymized skincare range, 48 API answers from Claude and Perplexity Sonar. No observed customer purchases, retailer-stock checks or medical assessment. Original answers and saved pages remain in the private research record.

First-party research · Evidence tier: proprietary · Confidence: B

S1

Skincare Snapshot study, anonymized

Hyperize Original Research · · First-party study

  • 65 questions × 2 systems × 2 runs = 260 answers. This article examines 12 question variants and 48 answers about one skincare range.
  • 11 answers list discontinued product pages as citations (Perplexity 10, Claude 1); none mentions the status. Another 12 contain them only in search results. One answer across all 48 mentions the status using retailer sources.
  • Four manufacturer pages say “no longer available”. A separate agent test finds no purchase path on those pages. Raw evidence remains private; the link leads to the study scope, not a public dataset.

API tests with web search; saved answers checked against manufacturer pages. n = 48.

Read the study scope above
S2

From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search

Qi Liu et al. · arXiv · · Research paper

  • 1,004 answers in the publication experiment: 17.5% cite a placed post, 7.5% reproduce a marker. Ten AI search systems; measurements in March and April 2026. The study does not test our proposed product correction.
Read source
S3

Which information does AI cite? (German)

Hyperize · · Guidance

  • Background on presenting individual facts clearly. Not evidence that the change proposed here will work.
Read source
S4

How is AI brand dominance measured? (German)

Hyperize · · Definition

  • Definition of Citation Conversion: the share of brand mentions accompanied by a citation of the brand’s own website.
Read source
S5

Agent Surface

Hyperize · · Methodology

  • The layer of pages, data and permitted functions an agent reads and uses.
Read source
S6

Is one ChatGPT measurement enough? (German)

Hyperize · · Measurement guidance

  • Repeat questions under the same settings; one improved answer establishes neither a lasting nor a causal effect.
Read source
Updated: Next review: Article as JSON