AI describes features we dropped last year. How do I get it to update?

An AI answer about your product comes from one of two places. Either the engine fetched pages at answer time and summarized them, or the model answered from what it learned in training. Each path fails in its own way.

• Old pages on your site. A 2024 feature page, an old changelog, a help article or a comparison page can stay live and indexed long after the feature is gone. If an engine retrieves it, the engine repeats it.
• Third-party pages. Review sites, listicles, partner directories and old launch coverage describe your product as it was when they were written. When engines cite other sites instead of yours, those sites shape the answer.
• Training data. A model knows nothing after its knowledge cutoff. OpenAI documents a February 16, 2026 cutoff for its GPT-5.6 models and August 31, 2025 for the Instant path used in ChatGPT. A feature you dropped in late 2025 can still exist in that model's memory.

The first two are fetchable, so you can change them. The third is frozen. The work is figuring out which one you're looking at, then fixing what you can reach. If you're unsure whether the engine is inventing or just behind, telling a hallucination from an outdated answer comes first.

Start with the engine that is wrong, and look at what it cites.

1. Ask the question that produced the stale answer. Use a logged-out session so memory and past chats don't skew it. Logged-in and logged-out results can differ for the same prompt.
2. Open every cited link. Perplexity shows its sources by number. Check which one mentions the dropped feature.
3. If a citation is yours, note the URL. That page is the fix.
4. If a citation is a third party's, note the site and the sentence.
5. If there are no citations, the answer probably came from training data. That is a retrieval-off answer, and it behaves differently (see below).

Repeat with a few phrasings. Engines rewrite your prompt into their own search queries, so one question doesn't show everything they retrieve. The free Perplexity brand check runs this kind of snapshot for you.

Retire old pages on your own site

An engine can't cite what its index no longer holds. In Otterly's 2026 experiment, Google's AI surfaces stopped citing two sites once Google removed them from its index, while Bing kept them and Copilot kept citing them. The same mechanism works in your favour for a dead feature page.

• Redirect an old feature page to the current product page if the URL still gets traffic or links.
• Remove it if there's nothing to redirect to.
• Add noindex to pages you must keep live, such as archived release notes. Google documents noindex, nosnippet and max-snippet as controls that keep content out of search results and AI features.

Expect each engine to drop the page on its own schedule. Check Google Search Console and Bing Webmaster Tools separately, because one index being clean tells you nothing about the other.

Publish the current truth where crawlers can read it

State the current feature list as plain sentences on a page that loads without JavaScript. Vercel's crawler study found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. If your feature grid renders client-side, those crawlers may see an empty page and fall back on older sources. A direct line such as "Product X no longer includes offline sync, which we retired in 2025" gives an engine a sentence it can quote.

Then confirm the crawlers can reach that page. Perplexity's index uses PerplexityBot and ChatGPT search uses OAI-SearchBot, and each is allowed or blocked separately in robots.txt. A firewall or CDN can still turn them away when robots.txt allows them. The AI Crawler Checker requests a page as each AI crawler and shows which ones get blocked.

Get third-party pages corrected

For each review or directory that an engine cites, contact the publisher with the specific outdated sentence and the correct one. Update any listings you control yourself, such as marketplace and directory profiles. Being named accurately on the sites engines already cite is associated with appearing in the answer, though nobody can promise the effect. Older press coverage works the same way, as covered in whether old news shapes AI answers.

Retrieval is optional on every platform. ChatGPT searches at the model's discretion. Claude.ai search is a user toggle. When the engine doesn't search, it answers from training data, and no page edit reaches that answer.

That has three consequences.

• You'll see mixed results. The same question can be right one day (search on, fresh page fetched) and wrong the next (search off, old training data).
• The timeline depends on the vendor. The stale claim fades only when a model with a later cutoff replaces the current one. Routing matters too: ChatGPT's Instant path carried an older cutoff than its reasoning models, per OpenAI's documentation.
• Your fixes still count. Fetchable fixes help every search-on answer now, and pages that crawlers can read are the material any future training run could draw on.

Whether allowing training crawlers like GPTBot changes what future models recall has never been measured, so treat that as a separate decision. For realistic timeframes, see how long AI takes to update brand information.

Re-run the same questions after each fix and record which engine said what, with or without citations. Measure each platform separately. A single blended score hides the fact that one engine may have updated while another still quotes your 2024 changelog.

The free checker on this site gives a snapshot of whether the main engines describe you correctly for three buyer questions. If feature errors sit alongside pricing errors, the same tracing method applies to ChatGPT getting your pricing wrong.

Can I ask Perplexity or ChatGPT to correct a wrong answer directly? No vendor documents a process for publishers to submit corrections to answers. Feedback buttons exist, but nothing published says they change what the engine tells other users. The reliable route is the source: fix or retire the page the engine cites, and get third-party pages updated. Search-on answers then have current material to draw on.

Will adding schema markup or an llms.txt file make AI pick up the new features faster? The evidence doesn't support it. Google says structured data isn't required for its generative AI features and that Search ignores llms.txt. Ahrefs' matched test found AI Overview citations fell 4.6% after pages added schema. Existing markup does no harm, but the useful fix is a plain, fetchable sentence stating the current feature set.

Should I delete old release notes that mention dropped features? Not necessarily. Release notes are useful history for customers. Keep them if they matter, but add noindex so search engines and their AI features don't serve them, or add a clear line at the top saying the feature was retired and when. Delete or redirect pages that only describe the old feature.

Why does one AI engine have the right features and another doesn't? Each engine uses its own index, its own crawler and its own model with a different knowledge cutoff. In Otterly's experiment, Copilot kept citing sites that Google had removed, because Bing still held them. One engine fetching your new page says nothing about the others, so check and fix each one separately.