Why does AI say my product has features it never had?

An AI answer about your product comes from one of two places. Either the model answers from memory (its training data, frozen at a knowledge cutoff), or it searches the web and builds the answer from pages it retrieves. Invented features appear when neither source gives it a clear, specific statement to use.

The process runs in four stages:

• Parametric: what the model learned in training. If your product was small or new at the cutoff, the model knows little and fills the gap with typical features for your category.
• Access: whether the engine's crawler can fetch your pages at all.
• Retrieval: whether your page enters the answer for a given question.
• Generation: which retrieved text the model uses and cites.

A failure at any stage leaves the model guessing. Generative models guess fluently, so an invented integration or export format reads like fact. If you're unsure whether you're looking at a fabrication or stale data, the difference between hallucinated and outdated answers changes what you fix.

These are the patterns to check first. None is a confirmed finding about your site until you test it.

• No dedicated feature page. Features scattered across blog posts and release notes give the engine nothing to quote in one place.
• Features that only appear after JavaScript runs. Tabs, accordions and comparison tables loaded by script may not exist in the raw HTML that AI crawlers read.
• Confusion with competitors. Say you sell scheduling software with a name close to a bigger rival's. A model with thin data on you may attach the rival's features to your brand.
• Wrong third-party pages. Review sites, listicles and old comparison posts sometimes credit products with features they never had. When an engine cites those pages instead of yours, it repeats the error.
• Blocked crawlers. A robots.txt rule or a firewall that turns away search crawlers removes your own pages from the candidate set, leaving only other people's descriptions.

The same mechanics drive wrong pricing in ChatGPT, so one fix often covers both.

The strongest evidence concerns access, not wording.

Vercel found that ChatGPT and Claude crawlers fetch JavaScript files (11.50% and 23.84% of their requests) but don't execute them, so they can't read client-side rendered content. Googlebot does render JavaScript, so this gap matters most for ChatGPT, Claude and Perplexity.

OpenAI documents separate crawlers: GPTBot for training and OAI-SearchBot for ChatGPT search. Its bots page says each setting is independent of the others. Blocking OAI-SearchBot keeps your pages out of ChatGPT search answers. Allowing GPTBot does not get you cited.

Google states that a page must be indexed and eligible to show with a snippet to appear as a supporting link in AI Overviews or AI Mode, with no additional technical requirements.

On content tactics, a 2026 critical survey of 45 GEO studies found that no reviewed technique shows a stable, cross-platform causal effect on discoverability. The honest position: clear facts on a fetchable page give the engine a sentence it can quote. That is not a promise it will.

1. Publish a plain feature list. One page, server-rendered, listing each feature in a specific sentence. Add what the product doesn't do where buyers often assume it does (for example, "no native mobile app").
2. Check the raw HTML. Load the page with JavaScript disabled, or view the source. If the feature list isn't there, move it into the server response.
3. Check crawler access. Confirm robots.txt allows OAI-SearchBot, PerplexityBot and Claude-SearchBot, then confirm your CDN or firewall doesn't return 403 to them. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
4. Remove contradictions on your own site. Make the pricing page, help docs and feature page agree.
5. Correct the third-party pages engines cite. Ask review sites and comparison authors to fix wrong claims. Being named accurately on cited sites is associated with appearing correctly in answers.
6. Re-check each engine separately. Results differ by platform. Our guide on finding out what AI says about your company covers how to test.

• The feature existed once. That's an outdated answer, not an invention. See how to handle dropped features.
• Only one engine gets it wrong. Each engine uses its own index and models. Gemini has its own correction path, covered in fixing Gemini's wrong facts.
• The answer came without citations. It probably came from training data. Nothing you publish this quarter changes that memory. OpenAI lists a February 16, 2026 knowledge cutoff for its GPT-5.6 models, and an August 31, 2025 cutoff for the Instant path, as verified in August 2026.
• The question was leading. "Does it integrate with X?" invites a yes. Test neutral questions too.