Not long after ChatGPT launched, I asked it which headphones to buy. I had spent years building Headphones.com, and the model never brought us up. In that early experience there was no live web search to inspect. A familiar way of measuring discovery had stopped being enough.
The useful distinction for an operator is between what a model has learned and the information an assistant retrieves while answering. It is a simplified way to organize the work, not a complete description of a recommendation system. Shopping answers can also use structured product data and the buyer's current request and context. OpenAI documents merchant product feeds as an input for its shopping experiences.
Audio taught me why the public record matters long before AI existed. Head-Fi threads, subreddit discussions and reviewers helped buyers decide who was credible. At Headphones.com, reputation accumulated through the products we sold, the advice we gave and the policies we stood behind. That material can be useful to an assistant when it is available in the sources the system uses. We cannot inspect a model and establish that it absorbed every thread, review or article.
Retrieval creates a different, more immediate check: what information can the assistant reach now? Some answers use live search, but that behavior depends on the product and mode. Perplexity, for example, documents controls that let a user turn web sources on or off. A result collected with one setting should not stand in for every experience.
The traffic from these systems is observable even though a recommendation's causes are not always visible. Adobe reported a 1,200% increase in traffic from generative AI sources to US retail websites between July 2024 and February 2025. That is a historical growth comparison over a specific window, not a forecast for an individual store or a measure of every AI-assisted purchase.
For a merchant, access is an actionable starting point. Inspect the page a shopper or search system would receive, and check the actual crawler rules. OpenAI distinguishes OAI-SearchBot for search from GPTBot for model-training purposes. Cloudflare likewise documents controls by bot purpose. Review the intended search access separately from the training policy instead of assuming one blanket setting describes both.
The accessible page then has to answer the buyer's question. Consider an illustrative listing for closed-back wired headphones. A shopper trying to avoid disturbing colleagues needs facts about the design, connection and intended use, plus evidence for any claim about sound leakage. A memorable model name alone does not establish suitability. Clear specifications help a reader assess the product, but they do not guarantee that an assistant will choose it.
Check consistency across the visible page, its structured data and any feed supplied to a shopping platform. A correct description paired with an outdated price or unavailable variant can still produce a poor buying experience. Follow the exact item through the purchasing path instead of stopping at a screenshot of a recommendation.
Press can contribute useful independent evidence, too. A published story may be retrievable if the assistant's search system can access and select it. Publication does not guarantee inclusion in a future training set or a place in an answer. I would judge coverage first by what it establishes for a reader and whether it reaches the people the business serves.
To see where you stand, choose a small set of real buying questions with budgets and use cases. One in audio might be “closed-back headphones under $500 for an office.” Ask under recorded conditions, keep the complete answers and cited pages, and repeat across comparable sessions and days. Record both appearances and omissions. The sampling guide explains why one good answer is not a stable visibility result.
Lantern grew out of wanting a way to inspect that changing picture. The useful work is to see how answers describe the brand, identify evidence gaps and test specific improvements. A recommendation tool can help prioritize the investigation; an observed change still needs to be checked against the same measurement conditions.
I expect assistants to take on more of the buying process. The timing and the winning interfaces remain uncertain. What a merchant can do now is concrete: make the product understandable, keep the offer accurate, build a credible public record and retain evidence of what the assistants actually answered.
Common questions
Do recommendations come from training data or live information?
Both can contribute, and some shopping experiences also use structured product feeds. The inputs vary with the assistant, feature and request, so the training-versus-retrieval distinction is a useful starting point rather than an exhaustive mechanism.
Does blocking GPTBot remove a site from ChatGPT search?
OpenAI documents separate controls for training and search crawlers. Review the named user agent and current documentation rather than assuming that a training restriction automatically prevents search discovery.
Why can a brand rank on Google and still be absent from an AI answer?
Different systems can use different sources, context and selection processes. Access, product facts and cited evidence are useful things to inspect, but an omission alone does not identify its cause.
How long does a product-data fix take to improve visibility?
There is no dependable universal timetable or guarantee. Verify that the corrected information is available first, then compare repeated observations under consistent conditions before attributing a change to the fix.