AI Still Recommends These Dead Companies — a Stale Training Data Case Study
Ask the major AI assistants for the best virtual try-on software and they will confidently recommend companies you cannot buy from — we measured it. This is what stale training data looks like in the wild, why it happens, and how to find the same blind spots in your own market. Vonca runs this measurement automatically across ChatGPT, Perplexity, Gemini and Claude from $99/month, and the free 60-second scan on the homepage gives you a first reading.
Disclosure: the data below comes from our own monitoring runs — we build the tool that produced it. Every company-status claim is linked to a primary source.
The measurement
In mid-July 2026 we put the same eight buyer questions to ChatGPT, Perplexity, Gemini and Claude — questions like "best virtual try-on software for fashion brands" — and logged all 32 dated answers. Three of the recommended vendors stood out:
- Zeekit appeared in 3 of 32 answers across two engines — including as ChatGPT's #1 recommendation, described as "enterprise-level." Walmart acquired Zeekit in May 2021; TechCrunch reported at the time that Walmart "will not continue to operate Zeekit's existing business", and the technology became Walmart's own in-house try-on feature. No brand can license it.
- Fits.me was named by Claude among the best options. Rakuten acquired Fits.me in 2015 and later shut the company down, absorbing the technology.
- Metail appeared in 3 answers as a shopper-facing fitting room. This one is subtler — and it's where most "dead tools" lists get it wrong: Metail still operates after its 2019 acquisition by TAL Apparel, but pivoted to 3D garment technology for production teams. The product the AI describes is not the product the company sells.
Three recommendation slots — a meaningful share of every answer — pointing buyers at doors that are closed or lead somewhere else entirely.
Why models keep doing this
Models learn from web content that ages. The 2019–2022 "best virtual try-on" lists where these names were correct still outweigh reality in the training corpus, and unless a strong, current, dated source contradicts them at answer time, the old consensus wins. Training cutoffs trail the present by months to years, so this is structural, not a glitch — every market has stale spots.
The mechanism cuts both ways, though. When an engine does retrieve live sources and they contradict its training-data belief, it has to lean on the correcting source to say so. That makes a verified, dated correction page one of the most reliably citable page types on the web — provided the correction is airtight. We nearly published "Metail shut down" ourselves before checking Companies House; a wrong correction would have burned the page's credibility entirely. Verify before you correct.
What this means for your brand
Stale spots are openings. If the engines in your market still recommend a vendor that folded, pivoted, or got acquired, the query stream that error generates — "is X still available?", "what happened to X?" — has almost no current answers competing for it. The brand that publishes the verified correction inherits that traffic and, more importantly, the citation.
The confusion has an economy. Residual demand around a defunct name doesn't disappear; it gets absorbed — by the acquirer, by lookalike products, by whoever answers the question first. Watching which stale names your market's engines still emit tells you where that demand is sitting unclaimed.
Your own brand can be the stale data. If you've repositioned, changed pricing, or renamed, the engines may still be describing the old you. The same measurement that finds competitors' stale spots finds yours.
How to find the stale spots in your market
Manually: ask all four engines the questions your buyers type, log every vendor named, then check each vendor's pricing page for a date within the last year. It works, and it's slow — which is why we automated it. Vonca runs your buyer questions across ChatGPT, Perplexity, Gemini and Claude every day, and its Stale-Data Radar flags recommended vendors that are no longer available, each with the primary source that proves it — turning the model's error into your correction brief.
Frequently asked questions
Why does ChatGPT recommend discontinued products?
Because its training data includes years of articles written while those products were active, and without a current dated source correcting the record at answer time, the outdated consensus resurfaces.
How do I check whether AI recommends outdated information in my market?
Run your buyers' real questions through the four major engines and verify every vendor named — or let Vonca do it daily and flag the stale ones automatically.
Can stale AI answers be corrected?
Yes — with a verified, dated correction page. Engines that retrieve it must reconcile it against their training-data belief, which is exactly the situation in which they cite. Accuracy is non-negotiable: verify every claim against a primary source first.
How fast do corrections show up in AI answers?
It varies by engine and query. Low-competition correction queries ("is X still available?") can surface within days of indexing; competitive "best X" queries take longer. There is no honest way to guarantee a timeline — anyone who does is guessing.
See what the engines say about your market right now. The free 60-second scan on the Vonca homepage runs your first questions across the engines — no signup, no card.