Why am I not cited in AI answers?

How do you find out what is wrong?

The cause almost always sits on one of four layers, and the layers build on each other: reachable, readable, citable, relevant. If one layer fails, every later one is pointless. The best answer block is worthless if robots.txt shuts the crawler out, and the most generous crawler access is worthless if the page holds no answer.

That is why the order matters. Layers one and two take minutes, they are configuration questions with a clear answer. Layers three and four are editorial work and take weeks. Start writing before you have checked access and you may be investing in text no crawler ever sees.

Layer 1: does the crawler even arrive?

Start with robots.txt, because this is where the most common mistake sits: blocking AI crawlers wholesale also blocks the search crawlers that produce the citations in the first place. The major providers separate their bots by purpose, and that separation is what decides your case:

• OpenAI: OAI-SearchBot powers the search features in ChatGPT, GPTBot collects material that may be used to train the models, and ChatGPT-User fetches pages when someone asks for them in a chat (OpenAI docs, 2026). Blocking GPTBot does not keep you out of ChatGPT answers, blocking OAI-SearchBot does.
• Anthropic: ClaudeBot collects web content that may contribute to training, Claude-SearchBot works on the quality of search results, and Claude-User fetches pages on user request (Anthropic Support, 2026).
• Perplexity: PerplexityBot indexes pages for search results and, according to Perplexity, is not used to train foundation models, while Perplexity-User fetches pages when a user question requires it (Perplexity docs, 2026).
• Google: Google-Extended governs training and grounding in other Google systems, not inclusion in AI Overviews. What counts there is the regular search index (Google Search Central, 2026).

The second place where requests fail is not in robots.txt at all: your CDN or web application firewall. In July 2026 Cloudflare split the controls into three classes, Search, Agent and Training, and from September 15, 2026 new domains get new defaults. Search stays allowed, Agent and Training are blocked on ad-supported pages, and multi-purpose crawlers that combine search and training fall under the training block (Cloudflare Changelog, July 1, 2026). So the dashboard belongs in the diagnosis, not just the file.

For Google there is a third point: there is no extra optimisation for AI Overviews and AI Mode. If you are not in the search index you cannot appear there, and nosnippet, max-snippet, data-nosnippet and noindex restrict previews and with them the AI features (Google Search Central, 2026). How the selection works beyond that is covered in How do I get into Google AI Overviews?

The hardest check is your server log: filter the last 30 days of requests by the bot names above. If a name does not appear at all, the diagnosis is finished and everything else can wait.

Layer 2: can the crawler see the content you mean?

A crawler sees the HTML your server delivers, and nothing else. The common AI crawlers do not execute JavaScript, so anything that only comes into being in the browser does not exist for them. The fastest test takes ten seconds: open “view page source” and search for your most important sentence. Do not use the developer console, it shows the rendered result and paints too friendly a picture.

Four cases where a page looks perfect to humans and empty to machines:

• Content loaded via JavaScript: tabs, accordions and product data from an API are the usual suspects.
• The key statement sits in an image or a video: numbers in the infographic, the table as a screenshot, the argument in an embedded video. None of that is text on the page.
• Something stands in front of it: consent banners, logins or paywalls that serve an interstitial instead of the article. What counts is what is in the HTML, not what a human sees after two clicks.
• The status code is wrong: error pages returning 200, long redirect chains, geoblocking, or a rate limit that hands crawlers one 429 after another.

Layer 3: is the passage citable?

From here on it is editorial. AI systems do not cite articles, they cite passages, and a passage has to stand on its own. The typical symptom on this layer: your page is demonstrably crawled and ranks in search, yet the answer names a competitor whose page is weaker on substance.

What is usually missing is the answer format, not the knowledge: no direct answer in the first paragraph, headings as keywords instead of questions, numbers without source and date, three paragraphs of preamble before the actual statement. How do I write an article that AI cites? covers the fix in detail, and How do I get cited in ChatGPT? sets out the criteria ChatGPT applies.

Layer 4: are you answering the question people actually ask?

The fourth cause is not technical: nobody asks the question you want to be cited for, at least not in that form. Questions in chat interfaces are longer, more specific and tied to a case. Not “disability insurance cost”, but “what does disability insurance cost if I work as a carpenter?”. Serve only the short version and you have nothing to offer for the long one.

The process, on the other hand, is always the same: collect the questions your audience actually asks, check which of them your website already answers, and write only the gaps. These tools find the questions people ask ChatGPT and Perplexity shows how to find them. One rule holds throughout: one question, one page that closes it. Three pages half answering the same question compete against each other.

What if everything checks out and you still do not appear?

Then three explanations remain, none of them a mistake:

• The answer fluctuates: the same question returns different sources across two runs. A single sample is not a measurement, it is a snapshot.
• Other sources are better placed for this question: for comparison and recommendation questions, answer engines often draw on overviews, trade media and review platforms rather than vendor pages. That makes it a question of mentions beyond your domain, not of your page.
• The change has not landed yet: sources fetched live take effect with the next crawl, model knowledge only with a new training run. GEO explained: why AI answers are the new ranking sets out why those two paths are separate.

How do you check whether anything changed?

With a fixed list of questions you ask repeatedly, across several runs. What matters is not whether one chat mentions you, but how often you are mentioned over a week, in what context and with which page. How do I measure my visibility in AI answers? covers the metrics worth tracking and how to collect them cleanly.

That is the chain flize brings together: a GEO score for access and readability, scouting for the open questions, the answer check against your website, and content production for the gaps, meaning articles written to be cited.