What Content AI Assistants Skip

Every guide to AI visibility says to publish more helpful content, and the content most teams publish sits in the exact categories an assistant discards before it reads a word. We run a 260 post outbound library, and last August Google AI Mode quoted our own site back at us and then recommended 2 competitors in the same answer. Below, the 3 gates that do the skipping, the 8 categories that never get through, and what survives.

What Content Do AI Assistants Skip?

The word skip is doing something specific here. It does not mean an assistant read your page and preferred another one. It means your page was never a candidate. Three separate systems drop content, in sequence, and by the time a model is choosing what to quote, most of the web has already been removed from the pool.

Here is the full list of what gets dropped, roughly in order of how often we see it on real B2B sites.

1. JavaScript rendered body copy. The crawler gets an empty container and moves on.
2. Anything behind a form, a login, or a cookie wall. A crawler will not fill in a form to reach your best material.
3. Text inside images and untagged diagrams. A chart with the finding baked into the pixels is a finding nobody can quote.
4. Undated pages. A claim with no date attached is hard to trust on anything that changes.
5. Unattributed pages. No named author, no bio, no profile behind the byline.
6. Summaries of other people's research. The assistant recognizes the claim as borrowed and cites the source you borrowed it from.
7. Buried answers. The fact exists but it sits in paragraph 9 of a section that never names its own subject.
8. Orphan pages. Nothing on your own site links to them, so nothing walks the crawler in.

Only 2 of those 8 are about writing quality. The rest are structural, which is the good news, because structural problems get fixed once and stay fixed.

Why Do AI Assistants Skip Most of the Web?

An answer engine runs 3 gates, and content has to clear all 3. Most teams only ever work on the third one.

Gate 1: Fetch

A crawler requests the URL and receives raw HTML. Whatever is not in that response does not exist as far as this gate is concerned. No rendering, no waiting, no second attempt.

Gate 2: Retrieval

The page is split into short passages, each passage is turned into a vector, and a query pulls back the handful that sit closest to it in meaning. The model normally sees the passage without the article around it.

Gate 3: Citation

Among the passages that came back, the system decides which are worth attributing. Recency, attribution, source reputation, and whether the claim originates here all get weighed at this step.

Retrieval augmented generation is the plumbing under gates 2 and 3, and the unit it moves is the passage, not the page. That single detail explains most of the skipping. Your article is not competing as an article. Each section of it is competing on its own, stripped of the heading above it and the paragraph before it.

Gate 1 is where the biggest and quietest losses happen, and it is the gate almost nobody audits. A page that fails at fetch produces no error anywhere in your analytics. It ranks fine in Google, it looks fine in a browser, and it is simply absent from the corpus that ChatGPT and Claude read from. We wrote up the mechanics of gate 2 separately in internal linking for LLM retrieval, because the link graph is what decides which passages a retriever can reach at all.

The reason this matters more every quarter is that citation and ranking have come apart. Ahrefs studied 863,000 keywords and 4 million AI Overview URLs and found that only 38 percent of cited pages also rank in the top 10 for the same query, down from 76 percent seven months earlier. Roughly a third of citations now go to pages that do not appear in the top 100 at all. Holding position one is no longer evidence that you are in the answer, which is the whole argument in LLM citation versus SEO traffic.

Which Formats Get Dropped Before Retrieval Even Starts?

Gate 1 is mechanical, so the failures are mechanical too. The single largest one is JavaScript.

Vercel measured this across its own network and published the numbers. In its analysis of AI crawler behavior it logged 569 million GPTBot fetches and 370 million Claude fetches in a single month and found that none of the major AI crawlers render JavaScript. They download JavaScript files as text, ChatGPT at 11.5 percent of its requests and Claude at 23.84 percent, and then never execute them.

The same dataset carries a second finding that gets less attention and costs just as much. ChatGPT's crawler spent 34.82 percent of its fetches on 404 pages and Claude's spent 34.16 percent, against 8.22 percent for Googlebot. A third of the crawl budget on your site is being burned on URLs that do not exist. Every stale link, every redirect chain, every dead entry in an old sitemap is taking a slot that a live page needed.

Onely's guidance on LLM friendly content lands in the same place from the SEO side, which is that content buried in long paragraphs and built in the browser does not make it into machine readable blocks. Google is the partial exception here, since Gemini shares a rendering pipeline with Googlebot, but building your AI visibility on the one crawler that renders is a strange bet when the other three do not.

Here is how the common formats actually behave at gate 1.

The gated PDF row is the one that starts arguments, so it is worth being blunt about it. A gated whitepaper is a trade. You are exchanging every citation the argument could have earned for a list of email addresses. That trade made sense when the only distribution channel was search traffic you could retarget. It makes much less sense when the same argument, published openly, can be the thing an assistant quotes for the next 2 years. Gate the tool, gate the template, gate the audit. Do not gate the thinking.

Why Does a Summary of Someone Else's Research Lose to the Research?

This is the category that catches good writers, because the content is genuinely useful and it still does not get cited.

An assistant cannot invent a fact and then attribute it to you. It can only point at a source. So when it encounters a claim in your post that clearly originated somewhere else, the honest move, and the one these systems are built to make, is to cite the origin. Your roundup of 10 industry statistics is a map of 10 other people's citations. You did the work of assembling it and they get the link.

The scale of the concentration this produces is worth sitting with. Digital Applied analyzed 1,000 AI Overviews and found the top 1 percent of cited domains capture 47 percent of all citations, with Wikipedia at 24.3 percent and Reddit at 21.6 percent. Around 1,100 other domains split the remaining 25.4 percent between them.

Read that as a strategy note rather than a complaint. You are not going to out-encyclopedia Wikipedia on a definitional query, and there is no amount of publishing that gets you there. The ground that is actually winnable is the ground where no encyclopedia entry can exist, because the claim did not exist until you made it. A 100 page study of AI Overview citation sources by CXL reaches the same practical conclusion from a different angle, which is that the pages winning citations tend to be the ones answering a narrow question completely rather than a broad question adequately.

Quality signals sit on top of that. Lumar's work on how AI systems judge source quality describes experience, expertise, and trust functioning as a gate rather than a ranking factor, which matches what we see. A named author with a real profile behind them clears it. A byline reading "our team of experts" does not, and neither does an undated page, which is the reason how often to refresh content for AI search turned out to be a more important question than it sounds.

What Does Get Cited Instead?

Everything that survives all 3 gates has one property in common. It cannot be sourced anywhere else.

There are only 4 kinds of content that reliably have that property, and every one of them is harder to produce than a blog post about a topic somebody else already covered.

• A number you measured. Your own campaign data, your own benchmark, your own failure rate. Nobody else has it, so if an assistant wants the claim it has to point at you. Our reply rate benchmarks and our inbox placement testing results are not repeatable by a writer who has never run a send.
• A transcript of a conversation nobody else recorded. A 45 minute interview with a named operator produces several thousand words of specific, quotable, first party language that exists in exactly one place on the internet.
• A named opinion. A real person taking a position, with a profile and a track record behind the name. This is what entity consistency for AI search is protecting, because a position only accrues to you if the systems can tell that all of it is the same you.
• A process you actually run. Step by step, with the parts that go wrong included. Sanitized process content reads as generic and gets skipped. The 3 failure modes you hit last month do not.

The second one is the reason our content and our acquisition stopped being separate projects. A recorded conversation produces a long, specific, first party asset and a sales conversation at the same time, from the same 45 minutes. We walk through the mechanics of that in podcast transcripts as AI search fuel and turning episodes into answer content, and the markup side in schema markup for podcast pages.

The catch is that a recorded conversation only clears gate 1 if you publish the transcript. An episode page carrying an embedded player and a 2 line description is the video row of the table above. It fails at fetch, and the most quotable asset you own never enters the pool. That failure is common enough that we wrote how to get your podcast cited by AI assistants as its own piece.

How Do You Tell If Your Own Content Is Getting Skipped?

Three checks, in this order, because each one rules out a different gate. None of them need a tool you have to buy.

1. Turn JavaScript off and reload the page. If the article body disappears, gate 1 is where you are losing and nothing downstream matters until it is fixed. Do this on a template, not on one page, because the failure is almost always a template.
2. Grep your server logs for the crawlers by name. GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, and Google-Extended. The user agent strings are documented in the OpenAI bots reference and in Anthropic's crawler documentation. You are looking for 2 things, which pages they fetched and what share of their requests returned a 404.
3. Ask your real buyer questions in the assistants themselves. Pick 10 questions a buyer would actually type, run them in ChatGPT, Perplexity, Claude, and Google AI Mode on a fixed schedule, and record who gets named and in which sentence. This is the only measurement that reflects the outcome you care about, and it is what tracking AI search visibility and auditing your brand in ChatGPT are built around.

One more check that costs 10 minutes and finds more than people expect. Pull your published URL list, pull every internal href on the site, and diff them. Every URL with zero inbound internal links is an orphan. On our own corpus that diff surfaced pages that had been live for months, connected to nothing, cited by nothing.

Also worth knowing before you set expectations. Answer engines pull more sources than most people assume, and Semrush found ChatGPT citing an average of 15 sources per response in its index of 126 million AI search prompts. Being 1 of 15 is a far more winnable game than being position 1, which changes what a realistic target looks like.

What Should You Publish Instead of Skippable Content?

The order matters here, and it is the opposite of what most content plans do. Fix the gates before adding pages, because publishing more into a broken corpus just makes the corpus bigger and no more visible.

1. Fix gate 1 across every template. Server rendered body copy, real anchor tags with real href attributes per Google's crawlable link guidance, and a 404 sweep. This is a one time engineering task and it unlocks everything else.
2. Ungate the arguments. Move the substance of your gated assets into open HTML pages. Keep the gate on tools, templates, and audits, where somebody is trading an address for something they use rather than something they read.
3. Put the answer at the top of every section. Name the subject in the first sentence of the body, not just in the heading, because the heading gets stripped. That is the entire idea behind answer capsules, and it is the cheapest formatting change with a real payoff.
4. Add the machine readable layer. FAQ schema, an llms.txt file, and a dedicated info page for AI assistants. These are how you state plainly what you are, in a format that does not depend on a model inferring it from your prose.
5. Say the category affirmatively. Never define yourself by what you are not. We learned this the expensive way, and the failure is in the next section.
6. Then produce original assets on a cadence. A measured number, a recorded conversation, or a named position. One a week beats 5 explainers a week, and it is not close.

Step 6 is where most teams stall, because original assets are the ones that require doing something rather than writing something. That is exactly why they work. The barrier that makes them expensive for you makes them expensive for everyone, which is the definition of a moat, and it is why cold email versus organic content is a false choice once the outbound is producing the recordings that feed the content.

What This Looked Like on Our Own Site

In August 2026 we asked Google AI Mode what High Ticket AI Systems is. It described us as not a traditional B2B podcasting production company, then recommended 2 named competitors to anyone who wanted a podcast agency.

It was quoting us. Our own machine readable files carried a negation of our category, and the qualifier did not survive compression. What reached the buyer was the shortest version of our own sentence, which read as an admission that we were not the thing they were looking for. Never publish a negation of your own category. The adjective gets dropped and the noun is what lands.

The second lesson from that audit was harder to swallow. We had argued the podcast case in 43 blog posts while the structured data on all 302 pages said we were a cold email agency. Prose does not outvote schema. The schema won, on every page, including the 43 that argued the opposite.

The fix was not new content. It was a sweep: one affirmative category noun everywhere, the same entity naming in schema and in prose, dates and authorship on every page, and a rebuilt link graph so the deep pages stopped sitting alone. Volume was never the problem. We already publish more than our 2 closest competitors combined, and it did not settle the category question by itself, which is the uncomfortable finding in why AI answers cite some companies and not others.

Give a sweep like that 4 to 8 weeks before judging it. Crawling, indexing, and re-embedding all lag, and a corpus of a few hundred pages does not get recrawled in a day. The signal to watch is not a rank position. It is whether your category noun appears in the first sentence of the answer, and whether a competitor is still being recommended over you on your own category query. We track ours the same way we tell clients to, which is the routine in how to rank in AI search and how to get cited by ChatGPT and Perplexity.

The Practitioner Takeaway

The instinct when you are not being cited is to write more. It is the wrong first move almost every time, because the content you already have is usually failing at a gate that has nothing to do with how well it was written.

Work backwards through the gates instead. Can a crawler read the page with JavaScript off. Does each section name its own subject in the first sentence. Is the claim in that section yours, or are you restating somebody else's finding and handing them the link. Three questions, and they will tell you more than a content calendar will.

The strategic version is simpler still. The web is filling up with content that summarizes other content, and assistants are built to route through the summary to the origin. Every month that continues, the value of being an origin goes up and the value of being a summary goes down. There is no version of the next 2 years where restating other people's research wins.

Which is why we stopped treating content and acquisition as separate budgets. One recorded conversation with a real operator produces a first party transcript that nothing else on the internet contains, a set of quotable specifics, and a live sales conversation, out of the same 45 minutes. The content problem and the conversation problem have the same solution, which is the case we make in how to get clients from a podcast and how to turn podcast guests into clients.

If you want the conversation side handled, that is what we install. We run the list, the email invitations, and the follow up, we edit and publish every episode, you own the recording on your own show, and we guarantee 30 recorded conversations with your ideal buyers in 90 days or your money back.