The Content That Still Works in AI Search

Published July 9, 2026 · ClickRadius Blog

Here is the uncomfortable thing I've had to tell more than one client this year: most of the content on your site is now invisible to the engines that decide what your buyers read. Not penalized — invisible. The answer gets synthesized without ever naming you. But some content still gets cited, consistently, across ChatGPT, Gemini, Perplexity, Claude, and Grok. I've spent a lot of hours staring at the difference, and it is not subtle once you see it. This post is that difference, laid out plainly: what still earns a citation, what quietly stopped working, and why.

Why most content went dark

The mechanism changed, so the reward changed. Google's May 19, 2026 announcement described a redesigned Search experience and made Gemini 3.5 Flash the default model inside AI Mode. It did not establish AI Mode as the global default or formally demote ordinary results; normal SEO foundations remain relevant. AI Overview coverage varies by measurement source, query cohort, collection date, locale, device, and feature definition. This page does not assert a universal current percentage. When there is no list of links to click, "content that ranks" is no longer the same thing as "content that gets seen."

That reframes the whole job. The engine still reads your page. The question is whether it finds anything worth crediting you for, or whether it simply absorbs the gist and moves on. A page that repeats what a hundred other pages already say gives the model nothing to cite — it can generate that sentence itself. So the content that "went dark" mostly didn't get worse. The bar moved, and generic content is now below it.

An answer engine only names a source when that source offers something it can't just synthesize. Citation is not a reward for optimization. It is a reward for being irreplaceable on a specific point.

— The ClickRadius team

What still works, in order of impact

1. Original data and first-hand experience

This is the top of the list for a reason: it is the one thing a model cannot manufacture. A number from your own operations, a documented before-and-after, something you tested or measured or saw — that is content an engine has to attribute, because it exists nowhere else. You don't need a formal study. First-hand experience is original data. The moment your page contains a fact the model would otherwise have to leave out, you've given it a reason to say your name.

2. The Princeton triad: quotations, statistics, citations

According to the Princeton-led GEO research (KDD 2024), interventions involving quotations, statistics, and source citations affected the selected source's relative visibility in the study's generated responses, with changes up to 40% in some settings. These elements make a useful editorial checklist: is there an attributable statement, are claims backed by specific numbers, and are sources named and linked? Three yeses make a page easier for a reader to inspect; they do not establish that an engine will retrieve or cite it. The benchmark began with the top five Google sources, randomly selected one source for optimization, and measured that source's relative visibility/impression in generated responses; it did not measure retrieval or citation-inclusion probability, and reported effects varied by domain.

3. Extractable structure

Engines lift answers in chunks, so content that is already chunked wins. That means tight, self-contained passages that make sense pulled out of context; a real FAQ; clean lists; a direct answer near the top of a section rather than buried under three paragraphs of throat-clearing. I've watched the same fact get cited when it's stated crisply in one sentence and ignored when it's smeared across a paragraph. The engine is doing extraction, so make your best points extractable.

4. Question-form headings

People query AI in natural questions, and the engines match against content shaped the same way. Headings phrased as the actual questions your buyers ask — and answered immediately underneath — map directly onto how answers get assembled. It is a small edit with outsized leverage, and it doubles as the honest way to structure a page anyway.

5. Genuine, verifiable expertise

Expertise can be demonstrated on the page through depth, specificity, and correct sourcing, while independently controlled records can corroborate the entity behind it. Those are distinct evidence surfaces. Neither guarantees a citation, and their causal shares have not been measured.

Write the thing only you could write, package it so a machine can lift it, and make sure the machine can confirm who you are. That is the entire content playbook for AI search.

— Douglas Brown, founder, ClickRadius

What stopped working

The mirror image is just as clear. Here is the content I now tell people to stop producing:

• Thin generic SEO filler. Keyword-padded articles that restate the obvious. They could rank when ranking meant clicks; now they get absorbed without credit.
• Word-count padding. The 2,000-word article that says what could be said in 300. Length was a proxy for depth to old algorithms; engines extract substance, so padding just dilutes the citable parts.
• Unsupported claims. Assertions with no number, no source, and no first-hand basis. There is nothing to attribute, so there is no citation.
• Me-too content. The article written because competitors have one, adding no unique fact or perspective. If the model can generate it, it will — without you.

None of this is a penalty in the old sense. It is simpler and more brutal than a penalty: it is silence. The page loads, the engine reads it, and nothing about it earns a mention.

How to retrofit a page you already have

1. Find the one thing only you know. A number, an experience, a result. If the page has none, that is the real problem — add it before anything else.
2. Install the triad. Add a quotable expert statement, back every claim with an attributed statistic, and cite your sources by name with links.
3. Make it extractable. Convert headings to questions, put the direct answer first, break dense paragraphs into clean passages, and add a genuine FAQ.
4. Confirm the entity. Make sure the expertise on the page is backed by a consistent, verifiable identity off the page.

ClickRadius estimates five families of readiness inputs on a 0–100 scale and records monthly answer and citation observations from ChatGPT, Gemini, Perplexity, Claude and Grok. It may prepare an eligible on-site correction, but it executes that exact revision only with current authorization and a connection proven capable of that effect; otherwise the correction remains for review or customer implementation. A later citation observation does not prove that the correction caused it, and an external engine controls each answer. You can still perform the first content pass by hand with the four steps above.

Frequently asked questions

What kind of content gets cited by AI engines now?

Start with original data, first-hand experience, and clearly attributable evidence that a reader can inspect. The Princeton-led GEO research (KDD 2024) reported selected-source visibility changes for quotations, statistics, and source-citation interventions in a bounded benchmark; it did not measure retrieval or citation inclusion. Clear headings and self-contained answers improve readability, but whether an engine later names or cites the page must be observed separately. The benchmark began with the top five Google sources, randomly selected one source for optimization, and measured that source's relative visibility/impression in generated responses; it did not measure retrieval or citation-inclusion probability, and reported effects varied by domain.

What content stopped working in AI search?

Thin, generic SEO filler — the keyword-padded, say-nothing articles written to satisfy a crawler rather than a reader. When the results page returned ten links, mediocre-but-optimized content could still rank and collect clicks. Now the answer is synthesized directly, so a page that adds no unique fact, data point, or perspective simply gets absorbed without credit. It is not penalized so much as it is invisible: there is nothing in it worth citing.

Do I need to publish original research to get cited?

Original research helps, but you do not need a formal study. First-hand experience counts as original data — what you tested, measured, saw, or learned that no competitor can copy. A specific number from your own operations, a documented before-and-after, or an expert judgment stated plainly all give an engine something attributable. The bar is not academic rigor; it is that the content contains verifiable, non-generic substance the model would otherwise have to omit.

Curious which of your pages are already citable? Get your free AI Readiness Score — five evidence families, graded, with fixes prioritized — or see plans and pricing.