Content for AI citation · Worked examples

What does a quotable passage actually look like?

Four before-and-after passages, each built to fail one way and then rebuilt, with the property named every time. The reasoning behind each lives in its own article.

Content for AI citation worked examples12 min readUpdated 30 Aug 20268 sources, checked 30 Aug 2026

Back to the roadmap Next: Retrievability as a ceiling

On this page

The short answer

A quotable passage is one that is still true, still checkable and still about your subject after somebody cuts it out of your page and reads it with nothing around it. That is a property you can test on writing you already have, and the 2026 critical survey states the criterion exactly: not to “add numbers,” but to provide “relevant, verifiable, dated, and properly attributed evidence.”¹ A study of 55,393 trending queries found Google AI Overviews activating on 64.7% of question-form queries against 13.7% of queries overall, which is why every heading below is a question.²

Key takeaways

The test is extraction, not style: cut the passage out, read it cold, and ask whether a stranger could check every claim in it.
A fabricated statistic “may increase reuse while degrading epistemic quality”. The cheap version of every technique here works, which is why quotability alone is not the goal.
The same property helps or hurts depending on the question. A date earns its place in a time-sensitive answer and adds nothing to a stable definition.
This is a method for writing one good passage. Google says producing content for every fan-out variation, primarily to manipulate its systems, violates its scaled content abuse policy.

The standard

What makes a passage quotable rather than merely well written?

Surviving the cut. Read a passage with nothing above or below it and three things have to hold: it answers a question somebody would ask, every fact in it can be checked by a stranger, and the subject is named inside it rather than inherited from the page. The 2026 critical survey describes statistics, definitions, comparisons, prices, dates and references as forming “units that the generator can select and attribute”.¹ What that unit is is what engines quote; how much evidence stands behind each property is the levers with evidence.

One sentence in that survey is the constraint this page is built on: “The criterion is therefore not to ‘add numbers,’ but to provide relevant, verifiable, dated, and properly attributed evidence.”¹ An invented figure is more extractable than an honest hedge, so quotability alone is not the goal, and nothing here is advice to sprinkle numbers into copy.

Google’s guidance, first-party and last updated 10 July 2026 (checked 30 Aug 2026), sets the same priority in plainer words: “Creating content that people find unique, compelling, and useful will likely influence your website’s presence in generative AI search in the long run more than any of the other suggestions in this guide.”³ What follows makes a real answer easy to lift. It will not make a thin one look liftable.

Examples 1 and 2

What does a passage that fails the cut-out test look like?

Every passage here is our own construction, written to fail one way and then rebuilt. None is quoted from a real page or a real company.

Before. “Question formats. This means the difference is much bigger than most teams assume, and it is the main reason the approach below works so well.”

Why it fails. Four expressions point outside the passage: “this”, “the difference”, “most teams” and “the approach below”. Lifted out, it says something is bigger than something else, names neither, and can be attributed to nobody. Fine for a reader who just finished the paragraph above it; unusable to anything reading it alone.

After. “How often do AI Overviews appear for question-shaped queries? A study of 55,393 trending queries collected between 13 March and 21 April 2026 found Google AI Overviews activating on 64.7% of question-form queries, against 13.7% of queries overall.² Phrasing a heading as a question changes the odds that an AI answer is generated at all, before any question of which sources it cites.”

The property: self-containment. Every noun is defined inside the passage, the first sentence carries the whole answer, and the measurement arrives with its sample, window and source.

Before, failing differently. “Our platform cuts review time dramatically, and customers tell us the difference is night and day.”

Why it fails. Fluent, confident, unquotable. “Our platform” loses its referent the moment the passage leaves your site; “review time” has no definition and no baseline; “customers tell us” is not an attribution anyone can check; “dramatically” is not a measurement. Strip out what cannot be verified and nothing is left to quote.

After, with invented figures. “How long does a first-pass draft review take in Kestrel Review? Across 380 reviews in July 2026, Kestrel Review’s product logs recorded a median of 11 minutes from draft submission to first reviewer comment, against 42 minutes for the same teams in the month before adoption.” Kestrel Review and every number in it are invented. The shape is what transfers: a named subject, a defined metric, a sample, a window, and a stated origin for the figure.

The property: attributable evidence. The survey rates extractable evidence at moderate to strong support, conditional on “compatible truthfulness, attribution, and intent”.¹

Example 3

How do you add evidence without just adding numbers?

Before. “Companies that optimise for AI search get three times more citations.”

Why it fails, and why it works anyway. No subject, no comparison group, no period, no measurement definition, no source: the shape of evidence with none of the content. The uncomfortable part is that the shape performs. Adding a fabricated statistic, the survey finds, “may increase reuse while degrading epistemic quality”.¹ Raising reuse while lowering truth is a liability dressed as a lever, which is what the survey’s second white-hat test catches: are statistics, reviews and references verifiable?

After. “How much do published GEO tactics move citations? C-SEO Bench, a benchmark of ten conversational-SEO methods run across two tasks and six domains, found only three of 54 method and domain combinations significantly positive in its main experiment, and none of those in question answering.⁴”

The property: checkability, not numeracy. The replacement figure is smaller, less flattering, and confirmable by anyone who follows the link: the passage became attributable rather than more numeric, and a definition with no digits satisfies the same criterion.¹ What that benchmark implies for a roadmap is argued in why content tactics test null.

Example 4

Does the same passage work for every question?

No, and the survey says so: “A recent date helps with a time-sensitive query, but not necessarily with a stable definition. A quotation may help answer a historical question, yet burden a product listing.”¹ Its support table rates recency, prices and dates at moderate support, conditioned on time-sensitive or commercial queries. A factorial experiment of 252,000 trials across six models and eighteen factors, reported in that survey, found effects for explicit prices and recent dates while formatting alone had weak effects.⁸

For a definition question. “What is query fan-out? Query fan-out is the generation of several concurrent related queries from one user question, so that retrieval runs against the parts rather than the whole; Google’s documentation defines it as a set of concurrent, related queries the model generates to fetch additional relevant results.³” A publication date there would age the passage without making it truer.

For a commercial question. “What does Kestrel Review cost? Kestrel Review, the invented product used throughout this page, charges a flat $40 per editor per month with no annual commitment, as of 30 August 2026.” The price and the date are the answer. A version reading “pricing is flexible and scales with your team” is not a worse-written passage; it is one that cannot be the unit that answers.

The property: evidence chosen by the question. One rule removes the date from the first passage and requires it in the second, which is why a fixed template yields a page of passages each slightly wrong for what was asked.

The check

How do you check a passage you have already written?

Six checks, run on writing that already exists. Only two rest on a measured effect, and knowing which two is worth more than running all six.

Cut it out and read it cold

Paste it into an empty file and read it with nothing above or below. If you cannot tell what it is about, neither can anything else.

Name the subject in sentence one

No pronoun or definite phrase whose referent lives outside the passage. “Our platform” and “the approach above” both fail on their own.

Give every figure four things

The number, the sample, the date or window, and where a reader can check it. That is the survey’s criterion in full, and it asks for no extra numbers.¹

Match the evidence to the question

A date for a time-sensitive query, a price for a commercial one, a definition for a stable term. Copying all three everywhere is the template error.

Make one claim, not three

Drafting advice rather than a measured effect: a passage carrying three claims can be quoted as making the weakest of them.

Run the four white-hat tests

Semantic preservation, evidentiary authenticity, separation of content from instruction, disclosure and fairness. Fluency is not a pass.¹

The four tests in check 06 are the survey’s, and they are cumulative rather than alternative: a rewrite has to pass all four. Its summary of where the line falls is blunt: “Reorganizing paragraphs or adding a verified primary source will generally satisfy these tests. A model-directed sequence, fabricated testimonial, or instruction to favor a brand will violate them.”¹

Be honest about which checks are load-bearing. Checks 03 and 04 map onto levers the survey grades, extractable evidence at moderate to strong and recency and prices at moderate and conditional. Checks 01, 02 and 05 follow from the unit being a passage rather than a page, but their separate effects have not been isolated in any study this curriculum can link.

The ceiling

What does a good passage still not buy you?

Not retrieval, which decides whether the passage is ever a candidate. A 2026 benchmark that restored retrieval and reranking found body-text-only optimisation scoring below the untouched baseline on all three of its averaged measures.⁶ A better-written passage can lose more upstream than it gains downstream, which is the rewrite that costs you retrieval; whether the page is fetchable at all is retrievability as a ceiling.

Not accuracy. In a human evaluation of four generative search engines, 51.5% of generated sentences were fully supported by their citations and 74.5% of citations supported the sentence they were attached to.⁵ A passage built exactly to this standard can still be quoted for a claim it does not make.

And not volume. Google warns that creating separate content for every variation of how people might search, fan-out queries included, “primarily to manipulate rankings or generative AI responses in Google Search violates Google’s scaled content abuse spam policy”.³ This method produces one good passage at a time. Whether length, schema or an llms.txt file belong near the work is settled in length, schema and llms.txt.

The honest limit of these examples

Every passage above is an illustration written for this page, so none of them is evidence of anything. The clearest published description of what quoted passages look like is a vendor dataset this curriculum will not link, so these properties are corroborated by peer-reviewed work rather than demonstrated by it. Nothing here has been shown to move traffic or revenue, and the survey rates the claim that citation scores predict clicks or revenue at very low confidence.¹ Treat the six checks as cheap hygiene with a plausible mechanism, and measure retrieval, citation and fidelity separately, which is the survey’s own closing advice.¹

Where a product fits, and where it does not

Nothing here needs software: a passage that survives extraction is an hour with a text editor and the willingness to delete a sentence you like. Bavior does not write passages, does not grade them, does not crawl your site and cannot tell you which of two drafts is better. It covers the measurement after the writing: a fixed prompt set run across five engines on a schedule, recording which sources each answer cited, so you can see whether a rewritten passage changed who gets quoted. Where a cited source is a live discussion thread rather than a page you own, it drafts a reply on an account you control, which you approve before anything posts. Read metrics that hold up before trusting your own before-and-after. The free AI visibility check and the free GEO audit need no paid plan; paid plans are from $99/mo billed monthly, or $79.17/mo billed annually (as of 30 Aug 2026).

Sources, all checked 30 Aug 2026

“Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023 to 2026)”, 15 Jul 2026, arXiv:2607.14035 (survey preprint); §7.2 the “add numbers” criterion, §9.1 the four white-hat tests, Tables 4 and 5: arxiv.org/abs/2607.14035
Xu et al., 2026; 55,393 trending queries collected 13 March to 21 April 2026; AI Overview activation 13.7% overall, 64.7% for question-form queries (preprint): arxiv.org/abs/2605.14021
Google Search Central, “Google’s Guide to Optimizing for Generative AI Features on Google Search”, last updated 10 Jul 2026 (unique and useful content, query fan-out, scaled content abuse; first-party): developers.google.com/search/docs/fundamentals/ai-optimization-guide
Puerto et al., “C-SEO Bench: Does Conversational SEO Work?”, NeurIPS 2025 Datasets & Benchmarks Track, arXiv:2506.11097; three of 54 combinations significantly positive: arxiv.org/abs/2506.11097
Human evaluation of four generative search engines, 2023, arXiv:2304.09848; 51.5% of sentences fully supported, 74.5% of citations supporting their sentence: arxiv.org/abs/2304.09848
“SAGEO Arena”, KDD 2026, arXiv:2602.12187; end-to-end benchmark restoring retrieval and reranking; Table 2 body-text-only averages below baseline: arxiv.org/abs/2602.12187
Aggarwal et al., “GEO: Generative Engine Optimization”, KDD 2024, arXiv:2311.09735; the founding benchmark later studies test against: arxiv.org/abs/2311.09735
Vishwakarma et al., 2026; 252,000 trials across six models and eighteen factors; relevance and position as primary determinants of the first citation, effects for explicit prices and recent dates (reported in the survey at note 1)

Continue reading

Learn GEO GEO roadmap Learn GEO Previous: Rewrites that cost retrieval Learn GEO Next: Retrievability as a ceiling

FAQ

Frequently asked questions.

Does putting statistics into a passage make it more likely to be quoted?

Verifiable ones do, and the distinction matters more than the numbers. The 2026 survey rates extractable evidence at moderate to strong support but states the criterion is not to "add numbers,“ rather to provide relevant, verifiable, dated and properly attributed evidence. It also warns that a fabricated statistic may increase reuse while degrading epistemic quality, so an invented figure can raise your citation odds and your liability at the same time.

Should every heading on my site be a question?

Use question-shaped headings where a person would ask a question, and stop there. Question-form queries triggered AI Overviews on 64.7% of a 55,393-query sample against 13.7% overall, which is a real difference. But Google's guidance warns that creating separate content for every search variation, fan-out queries included, primarily to manipulate its systems violates its scaled content abuse spam policy. One good passage beats forty shaped ones.

What do I do with a passage that is true but has no number in it?

Leave it alone if it is a definition or a comparison, because those count as extractable evidence in their own right. The survey lists statistics, definitions, comparisons, prices, dates and references together as units a generator can select and attribute. A precise definition with no digits in it satisfies the standard; a vague sentence with a number bolted onto it does not, and is the more common failure.

How do I tell whether a rewrite actually improved anything?

Measure retrieval, citation and fidelity separately, which is the survey's own closing recommendation, because a rewrite can move them in opposite directions. One 2026 benchmark that restored retrieval and reranking found body-text-only optimisation scoring below baseline on all three of its averaged measures. A single blended visibility number will show you the sum of two effects cancelling, then report no change.

Bavior Editorial

The team that researches and maintains Bavior’s writing on Reddit marketing and AI search visibility. Every figure here is attributed to a named source with the date it was checked, and none of our links are affiliate links.

Found a number that looks wrong? Tell us and we will re-check it: support@bavior.com