What happens to AI answers when you discontinue or rename a product

An outdoor gear brand retires its best-reviewed three-season tent to make room for an updated version with the same name but a new frame design. Six weeks later, someone asks ChatGPT for a three-season tent under $300, and the retired model gets named, specs and all, with no mention that it is gone or that a direct replacement exists one click away on the same site. A few months later, a different company finishes a rename it planned for a year, announced with a press release, and updated across its own site the same week. Prospects researching the new name are still being told by Perplexity that the company operates under the old one. Two different events, discontinuing a product and renaming a company or product line, and the same underlying failure: the thing you track internally as “done” is not done anywhere an AI engine looks.

Two different ways of being stuck in the past

A discontinued product and a renamed one break for different reasons, even though both look like “the AI is out of date.” A discontinued product usually fails because of recrawl lag: the page describing it, or the page that used to sell it, has not been revisited since it changed, so the engine is still working from what it saw before. A rename fails for a slower, less technical reason. The old name is not sitting in one page waiting to be recrawled, it is distributed across years of training data, backlinks, directory listings and entity records like Wikidata, and a 301 redirect on your own domain does not touch any of that. Fixing a recrawl problem is mostly patience and a clean signal. Fixing a rename problem is closer to an outreach campaign.

The product is gone from your catalog, not from the index

Removing a page does not remove what AI engines learned from it, and a well-documented takedown shows how slowly that unwinds. Meikai tracked citation volume for 126 brands after Alibaba’s /product-insights/ article pages started returning soft-404s, generic error content served with a normal 200 status instead of an honest 404 or 410, on 27 March 2026. Measured against the pre-removal baseline, Google AI Mode’s citations fell below half within one week and then stopped citing the pages entirely, Grok fell below half within five weeks before its own coverage lapsed, and both ChatGPT and Copilot took 13 weeks to fall below half. Perplexity never dropped below half at all, and by mid-July was citing the removed pages at over its original rate (Meikai, 2026). That is one takedown of one site section, but the shape matches a broader pattern: Otterly’s study of more than 20 million URLs cited across seven AI search engines over a month found that 19.3% no longer resolve to a live page, with ChatGPT citing dead pages at the highest rate of the engines it measured, 25.1%, against Google AI Overviews at 12.6% (Otterly, 2026). A citation only clears once the engine happens to recrawl that exact URL again, and a discontinued product page, which nothing new links to and which you have no reason to keep updating, is exactly the kind of page that gets recrawled least.

The soft-404 detail in the Meikai case is worth sitting with, because it is a mistake most catalogs make without noticing. A product that returns 200 with placeholder or error content, rather than a real 404, a 410, or a redirect to its replacement, looks alive to a crawler even though it is not. Standard ecommerce SEO guidance is specific about this for feeds: a product replaced by a new version should be deleted from the feed entirely so it doesn’t confuse Google Shopping, and the page itself should either redirect to the closest living replacement or return a real 410 if nothing replaces it, not sit there marked OutOfStock, a schema.org value meant for something that might come back (Yoast, 2026). The same logic applies to the page itself. If the model is gone for good, say so with a real status code or point the URL at its replacement. Leaving a soft-404 up is how a product keeps looking current long after it should not.

A new name has to outweigh years of the old one

A rename is not a recrawl problem because the old name was never confined to one page you control. One vendor’s breakdown of the mechanism lists the layers that keep the old name in place: models trained before the rename have no record of it at all, a name used for years simply appears in more training documents than a name used for months, entity records like Wikidata and Wikipedia often still list the old name as primary, and years of backlinks still use the old name as anchor text. The same piece notes that Perplexity tends to reflect a rename fastest because it leans on Wikidata, Gemini draws more on Google’s Knowledge Graph, and AI Overviews depend heavily on your redirects and canonical tags actually being in place, which is a different fix for each platform rather than one fix that covers all of them (Passionfruit, 2026). A 301 redirect is still necessary, it is how a crawler finds the new page at all, but it resolves a URL, not whatever a model already associated with the old name before the redirect existed. Our guide to building the entity signals AI engines trust covers keeping an entity’s identity consistent in the first place; a rename is that same problem run in reverse, trying to move an identity that sources have already agreed on.

What narrows the gap in both cases

Neither failure mode resolves itself on your schedule, but each has a specific, checkable list rather than a vague “give it time”:

• For a discontinued product, make the page’s status unambiguous. Redirect it to the closest living replacement if one exists, or return a real 404 or 410 if it does not, and pull it from any product feed entirely rather than leaving an out_of_stock flag that implies it is coming back.
• For a rename, update the records that carry more weight than your own site: your Organization or Product schema’s alternateName and legalName, and the entity profiles (Wikidata, major directories) that AI engines treat as a primary source rather than a claim you are making about yourself.
• For both, go find the pages you do not control that are still repeating the old fact, a comparison article, a directory listing, an old review, and ask for the update. Our piece on the sources AI engines trust covers why a crowd of old, consistent mentions can outvote the one page you already fixed, which is the same mechanism at work when a price or spec you’ve already corrected keeps getting cited anyway.
• Audit the page itself, not just the decision you made about it. Our AI-readiness audit guide covers checking that structured data and visible content agree, which matters most exactly when something just changed and the two have a real chance of disagreeing.
• Check what is actually being said now, on a schedule, rather than waiting for a customer to forward a screenshot. Tracking citations and sources across engines catches a lingering mention of a dead SKU or an old name while it is still one or two queries, not the whole picture of how your brand shows up.

Check it instead of guessing

The reliable move after a discontinuation or a rename is the same either way: stop assuming the change has landed and go look at what engines are actually saying. A free Visibility Check gives you a same-day read across the engines your buyers use, and if the old product or the old name is still showing up, Monitor turns that one check into a trend you can act on, instead of finding out again the next time someone forwards you a screenshot.