How to Triage AI Citation Gaps When Dozens of Pages Aren't Cited

The first time you run a real prompt set against AI answer engines, the result is demoralising. We ran 25 prompts across ChatGPT, Perplexity, and Google AI Overviews on our own domain this week, 148 runs in total. An AI Overview appeared on 146 of them. We were cited in zero, and not even mentioned in one.

That is the normal starting point, and the useless response to it is a spreadsheet of every page that isn’t cited. Fifty rows of “not cited” is not a work plan. The question that matters is narrower: for this specific prompt, what kind of failure is this, and is it the kind I can fix with content at all?

Here is the classification we use, why the distinctions matter, and the order to work them in.

First, separate “invisible” from “mentioned but not cited”

Before classifying anything, split your results three ways:

Cited. Your URL is linked inside the answer. The engine used you as evidence.

Mentioned. Your brand is named, no link. The engine knows you exist and chose someone else’s page as the source.

Invisible. Neither. You are not in the answer at all.

Most teams collapse mentioned and cited into one number, which hides the single most actionable segment you have. Mentioned-but-not-cited is the cheapest gap to close. The engine has already decided you’re relevant; it just didn’t find a page of yours worth pointing at. That is usually a content-shape problem on a page you already own, not an authority problem.

Invisible is a different animal, and often not a content problem at all. See Offsite Seed below.

The seven gap types

Every uncited prompt resolves to one of these. The value is in the discipline: you name the type before you decide the fix, which stops you defaulting to “write another blog post” for everything.

1. Update Existing: the page exists and is wrong for this audience

You have a page that should win this prompt. It doesn’t, because it’s written for a different reader than the one asking.

This is the most commonly misdiagnosed gap, because the page looks fine. The tell is an audience mismatch between the query and the page’s framing. A comparison page written for a neutral first-time evaluator will lose a query from someone already using a competitor and asking whether to switch. Different reader, different question, same topic.

Evidence to look for: the cited competitor page covers the same topic but speaks directly to the querent’s situation. If the engine cited a page that says “if you’re already on X,” and yours says “X vs Y at a glance,” that’s your answer.

Fix: reposition the existing page. Do not write a new one, or you’ll cannibalise yourself.

2. Add Evidence: the claim is there, the proof isn’t

Your page asserts the right thing but offers nothing an engine can quote as substantiation. No figures, no named sources, no dated methodology.

AI engines strongly prefer extractable, attributable claims. “Faster onboarding” is not citable. “Onboarding completes in under 10 minutes, measured across 40 accounts in Q2 2026” is.

Fix: add specific, sourced, dated evidence to the section that addresses the prompt. Not the whole page, just the section.

3. Create Comparison: no owned asset for a head-to-head query

A prompt explicitly compares you against a named alternative, and you have no page that does so.

Important caveat, learned the hard way: check whether the page already exists before you classify it this way. We spent a week acting on nine high-priority recommendations telling us to build a comparison page we’d already published, because the check wasn’t there. If you own the page, this is Update Existing, not Create Comparison.

4. Create How-To: a procedural query with no procedural page

The prompt asks how to do something and you have only conceptual content. Engines answering “how do I…” want ordered steps, prerequisites, and a stated outcome. A thought-leadership essay on the same topic will not be cited for it.

5. Create Diagnostic: the reader wants to figure out what’s wrong

Distinct from How-To, and frequently conflated with it. A how-to assumes you know the destination. A diagnostic assumes you don’t: the reader has a symptom and needs to identify the cause.

“How to improve AI citation rate” is a how-to. “Why isn’t my page being cited” is a diagnostic. They want different structures: the diagnostic needs branching, symptom-to-cause mapping, and elimination logic.

6. Improve CTA: cited, but nothing happens

The rarest and most annoying. You’re cited, traffic arrives, nothing converts. The page earns the citation and wastes it.

Worth naming separately because it’s invisible to any tool that only tracks citation rate. If you aren’t tying AI referral traffic to conversion, you’ll never see this gap.

7. Offsite Seed: the citation pool doesn’t contain you

The engine is drawing from third-party listicles, review platforms, directories, and community threads. None of them mention you. No page on your own domain will fix this, because your domain isn’t in the retrieval set for this query.

This is the one to be honest about. In our own run, 21 of 32 open recommendations were Offsite Seed. That is a distribution and PR problem (getting into the roundups, claiming the review profiles, earning the community mentions), and treating it as a content problem wastes a quarter.

How to tell them apart, fast

The classification comes from evidence in the answer itself, not from inspecting your own site:

1. Read what actually got cited. The cited URLs tell you what shape of content the engine wanted. Owned competitor page → you need one. Third-party listicle → Offsite Seed. Their page on your topic, framed for this reader → Update Existing.
2. Check whether you own a page on that topic at all. If yes, the default is Update Existing or Add Evidence. Creating a second page competing with your own is a common and costly mistake.
3. Read the query’s grammar. “How do I” → How-To. “Why isn’t” / “what’s wrong with” → Diagnostic. “X vs Y” / “should I switch” → Comparison. “Best tools for” → usually Offsite Seed, because those SERPs are dominated by roundups.
4. Check mentioned-vs-invisible. Mentioned means you’re in the consideration set, so favour fixing an existing page. Invisible on a high-volume commercial query usually means Offsite Seed.

The order to fix them in

Not by count. By cost-to-close against likelihood-of-moving.

First: Update Existing and Add Evidence on prompts where you’re already mentioned. Cheapest possible wins. The engine already knows you; you’re changing a page you already own; and the feedback loop is short enough to learn from.

Second: Create How-To and Create Diagnostic for high-demand prompts with no owned asset. More expensive, but this is where durable informational coverage gets built. Prioritise by measured demand, not by how interesting the topic is to you.

Third: Create Comparison. High commercial intent, but usually a small number of prompts, and you must verify the page doesn’t already exist.

Fourth, and in parallel with everything: Offsite Seed. It has the longest lead time, so start it early even though it pays back last. It is also the only category where writing more content on your own site does nothing at all.

Improve CTA whenever it appears. It’s rare, and it’s pure waste while it lasts.

The honest part

Two things worth saying plainly.

Classification is only worth doing if you then measure whether the fix worked. Shipping a change and watching an aggregate dashboard drift is not measurement. You need the same prompts re-run on the same engines before and after, or you’re guessing with extra steps.

And a gap you classify as Offsite Seed is a gap content cannot close. The discipline of naming it correctly is what stops you spending a quarter writing posts for queries where your domain was never in the running.

---

We run this classification on our own domain and publish what it finds, including the parts that don’t flatter us. The run behind this post returned zero citations across 75 attempts. If you want the same analysis on yours, start with a free account: 5 prompts, 1 platform, no credit card.

Frequently asked questions

How do I classify AI citation gaps when dozens of pages aren't cited?

Resolve each uncited prompt to one of seven types (Update Existing, Add Evidence, Create Comparison, Create How-To, Create Diagnostic, Improve CTA, or Offsite Seed) using the URLs the engine actually cited as your evidence. Then sequence by cost-to-close, starting with prompts where you're already mentioned but not cited.

Which AI citation gaps should I fix first?

Prompts where you're mentioned but not cited, fixed on pages you already own. The engine has already judged you relevant, the change is an edit rather than a new asset, and the feedback loop is short.

Why isn't my page cited by ChatGPT even though it ranks well on Google?

Ranking and citation are different selection problems. Common causes: the page is framed for a different reader than the one asking, it makes claims without extractable evidence, or its structure doesn't match the query's shape, such as a conceptual essay answering a procedural question. Start by reading which URLs were cited instead of yours.

Is it worth writing new content for every gap?

No. In our own analysis, 21 of 32 open gaps were Offsite Seed, where no amount of owned content helps. Writing for those is the single most common way teams waste a quarter on AEO.