Why isn’t my content appearing in AI-generated answers?
There are only two ways a page fails: it was never retrieved, or it was retrieved and not used. Telling them apart takes one test and determines everything you do next.
This is a page-level question rather than a brand-level one, and it has a satisfyingly narrow answer space. A page fails to appear in AI answers in exactly two ways, and telling them apart is most of the work.
Only two failure modes
Generative systems retrieve candidate pages, then select passages from them. So:
Retrieval failure. Your page was never in the candidate set. The model had no opportunity to use it, and nothing about your writing is relevant to the outcome.
Selection failure. Your page was retrieved, read, and passed over in favour of someone else’s passage. The content reached the model and lost.
These look identical from outside — you are absent either way — and their remedies share nothing at all. Rewriting a page that is never retrieved is wasted effort. Fixing rendering on a page that is retrieved and rejected changes nothing.
The test that separates them
Ask an assistant a question so specific that only your page could answer it. Your own product name plus a detail documented nowhere else — a specific limit, a particular feature behaviour, a number from your own documentation.
If it answers correctly: your content is being retrieved and read. You have a selection problem. Skip to the selection section.
If it cannot answer, or invents something: your content is not reaching the model. You have a retrieval problem, and it is mechanical.
This takes two minutes and saves a great deal of guessing. Run it against ChatGPT and Perplexity separately, because the answer can differ — Perplexity retrieves more aggressively and more transparently.
If retrieval is failing
Four causes, in frequency order.
1. The content only exists after JavaScript. The most common cause by a wide margin. AI crawlers fetch raw HTML and do not execute scripts — an analysis by Vercel and MERJ covering more than 500 million GPTBot fetches recorded zero JavaScript execution, with the same behaviour from ClaudeBot and PerplexityBot.
Check: disable JavaScript in your browser and load the page. If the content is not there, this is your answer. Fix: server-render the content that matters.
2. The crawler is blocked. Usually at the CDN or WAF rather than in robots.txt, through a bot-management preset nobody chose deliberately.
Check: request the page from outside your network with an AI crawler user agent and confirm a 200. Fix: allow the crawlers you want, deliberately.
3. The page is not indexed or not relevant enough. Retrieval leans on organic signals. A page that is not indexed, or is buried with no internal links, is unlikely to be pulled into a candidate set.
Check: URL Inspection in Search Console, plus whether the page has internal links pointing at it. Fix: ordinary SEO.
4. The server is too slow. AI crawlers use tight timeouts, often one to five seconds. A slow response can produce no crawl at all, silently and with no error anywhere.
Run both checks at once
A free scan tests AI retrievability on your pages and shows what the assistants actually answer for your category.
Run my free scanIf selection is failing
Your page reached the model and was passed over. Four editorial causes.
1. No self-contained passage. Models lift two or three sentence chunks. If every paragraph on your page depends on the previous one, on an unresolved pronoun, or on a chart, there is nothing to extract cleanly.
Check: copy three consecutive sentences from the middle of the page and read them cold. Do they answer anything? Fix: name subjects instead of referring back, and put the answer in the same sentence as the question it answers.
2. The answer arrives too late. Preamble is fatal. If the direct answer is in paragraph six, the model uses someone else’s paragraph one.
Fix: move the answer to the first two sentences under each heading, then add nuance.
3. The page answers too many questions. A page covering six topics answers none of them cleanly enough to be preferred over a page devoted to one.
Fix: split it. One complete answer per page beats six partial ones, which inverts a decade of content-hub thinking.
4. Nothing is specific enough to attribute. Vague claims get paraphrased and lose the citation. The Princeton and IIT Delhi GEO study found Statistics Addition, Quotation Addition and Cite Sources were the three highest-performing methods tested, with roughly 30–40 percent relative visibility improvements.
Fix: add a dated number, a named source, or an original data point. Something a model cannot reproduce without crediting you.
Page-level checks
A short list to run against any page that should be cited and is not:
- Does the content appear with JavaScript disabled?
- Does the page answer one question, or several?
- Is the answer in the first two sentences of its section?
- Can three sentences be lifted and still make sense?
- Is there a number, date or named source in the extractable passage?
- Does the heading match how a person would phrase the question?
- Does structured data match the visible text, or contradict it?
- Is the page linked from your other relevant pages?
Read who won instead
The most efficient diagnostic step and the most frequently skipped: look at the source that was cited.
Whoever holds the citation is showing you exactly what the model preferred. Open their page and compare it against yours on the checks above. In most cases the difference is visible within thirty seconds — they answered in two sentences and you answered in four paragraphs, or they had a number and you had an adjective.
Occasionally the cited source is not a company page at all but a forum thread, a review site or an independent roundup. That is also a finding, and a more important one: it means the citation is not winnable with on-site work alone, and the real gap is that those third-party sources do not describe you. Our guide to increasing AI citations covers that half.
When both failures are happening
A common and confusing situation: some assistants cite the page and others never do.
This is usually not a mystery. Google AI Overviews draw from Google’s index and Googlebot renders JavaScript, so a client-side-rendered page can be cited there while remaining completely invisible to ChatGPT and Perplexity, whose crawlers do not render.
If you see that pattern — present in AI Overviews, absent everywhere else — you have identified the rendering problem without needing any other test. It is one of the cleaner diagnostic signals available, and it is only visible if you measure more than one engine.
The reverse pattern is also informative. Named by ChatGPT but never cited in AI Overviews usually means your brand is well described by third parties while your own pages do not rank well enough to be pulled into Google’s candidate set — a corroboration success sitting on an SEO weakness.
Deciding whether the page is worth fixing
Not every page that fails to be cited should be rescued. Three questions decide it.
Is the question actually being asked? Run the prompt yourself. If assistants rarely produce a citation-bearing answer for it — because it is answered from general knowledge with no sources — then no page will be cited for it and the effort belongs elsewhere.
Is a competitor being cited for it? If yes, the citation is winnable and you can see exactly what winning looks like. If no source is cited by anyone, that is a much weaker opportunity.
Would the citation matter commercially? Being quoted on a definitional question aimed at students is a nice chart movement and nothing else. Prioritise the questions your buyers ask before choosing a supplier.
The order to fix a batch of pages
When several pages are failing, resist working through them one at a time. Do the mechanical check across all of them first — if rendering or crawler access is the cause, it is the cause for every page simultaneously and one fix resolves the whole set.
Only once retrievability is confirmed does per-page editorial work make sense, and then it should be ordered by commercial value of the question rather than by how annoying the page is.
How long to wait
Before concluding a fix has not worked, allow for the propagation lag, which differs sharply by engine.
Perplexity: days. It cites live retrieval openly and is the earliest indicator that something landed.
Gemini: a few weeks, on Google’s recrawl cadence.
ChatGPT: several weeks at minimum, because it depends on recrawling and often on third-party sources catching up too.
And in all cases, judge on a trend rather than a single answer. Generated responses vary between identical runs, so one absent result after a fix proves nothing — run the question several times across a few weeks before deciding the work failed.
Sources
- Vercel and MERJ — analysis of 500M+ GPTBot fetches finding zero JavaScript execution; same behaviour for ClaudeBot and PerplexityBot.
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande — “GEO: Generative Engine Optimization”, ACM SIGKDD 2024 (arXiv:2311.09735); ~30–40% relative visibility improvement from statistics, quotations and cited sources.
- Google Search Central — URL Inspection and index coverage documentation.
Frequently asked questions
Why isn’t my content appearing in AI-generated answers?
There are only two possibilities. Either the page was never retrieved — because it is not in raw HTML, the crawler is blocked, or it is not relevant or authoritative enough to be pulled in — or it was retrieved and the model chose someone else’s passage. The remedies share nothing, so identifying which one applies is the whole diagnosis.
How do I know if AI can see my page at all?
Ask an assistant something so specific that only your page could answer it — a detail documented nowhere else. If it answers correctly, your content is being retrieved and read, so the problem is selection. If it cannot answer or invents something, the content is not reaching the model and the problem is mechanical.
My page ranks first on Google but is never cited. Why?
Most often because the content only exists after JavaScript runs. Googlebot renders JavaScript and AI crawlers do not, so a client-side-rendered page can rank perfectly and show GPTBot an empty shell. The next most common cause is that the page answers many questions vaguely rather than one completely.
Does page length affect whether AI cites it?
Not directly, and padding hurts. Models extract passages of two or three sentences, so what matters is whether any passage answers the question completely and stands alone. A long page with no self-contained answer is less citable than a short page with one.
How long after publishing before a page can be cited?
Perplexity often picks up a new page within days because it cites live retrieval openly. Gemini follows Google’s recrawl cadence, typically a few weeks. ChatGPT is slowest, since it depends on recrawling and on third-party sources referencing the page. Four to eight weeks before drawing conclusions is reasonable.
Should I republish or rewrite a page that is never cited?
Rewrite the structure before considering republishing. In most cases the page has the right information arranged in a way that cannot be extracted — the answer arrives in paragraph six, or is spread across three sections. Restructuring answer-first fixes more of these cases than new content does.