How to make content citable: writing passages an AI assistant can lift

A citable passage is a block of text that still means something after an AI assistant lifts it off the page. It names its own subject rather than leaning on the heading above it, and every number in it carries the source and the date that number came from.

This is lesson 6 of Learn GEO. The earlier lessons were about measurement, because measurement is the part we can stand behind. This one is about editing, which is the part where most of the published advice on the subject stops being evidence and starts being folklore, so one section of this page sorts our own recommendations into what has been tested and what has not. The engines we measure are ChatGPT, Gemini and Perplexity, described on AI visibility tracking; the terms and the surfaces they refer to are in lesson 1, what answer engine optimization means.

On 2026-09-20 we opened two of the pages currently ranking for this topic and read them end to end. Both present a numbered list of tactics, and both present every item at the same level of confidence. Neither said which of its recommendations had been tested and which were the author's judgement. That distinction is the whole of the difficulty here, so it is what this page is organised around.

The five edits

Each one is cheap, and each one is judged by the same test: cut the passage out, show it to somebody who has not seen the page, and see whether it still says anything.

Put the definition in the first sentence. Not the background, not the market trend, not the problem statement. The thing being defined, named, in the subject position, in the opening clause. Forty to sixty words is enough to define almost anything, and a passage that length is the right size for something to quote whole.

Make the paragraph independent of what precedes it. A pronoun whose antecedent is two paragraphs up, or a subject carried down from a heading, breaks the moment the paragraph is read on its own. Repeating the noun reads slightly clumsy to a human who is reading in order. It is the correct trade.

Use a table when you are comparing things. Prose that alternates between two options ("the free check runs three questions, whereas paid monitoring runs the full set") forces a reader, or anything reading on their behalf, to hold both halves in memory and pair them up. A table has already done the pairing.

Write FAQ headings as complete questions. "Pricing" is a label. "How much does weekly monitoring cost?" is a question, and the answer underneath it can be matched to somebody who asked roughly that. The heading is the retrieval key; a one-word key retrieves nothing in particular.

Put a source and a date on every number. A number with no provenance is unusable to a careful reader and unverifiable to anyone else. This is also the rule that constrains us hardest, and the reason several claims that would have fitted nicely in this article are not in it.

Four rewrites from our own pages

These are real passages from this site, quoted as they currently stand, not invented examples. Where a passage was already in the right shape we say so rather than manufacturing a fault.

1. The opening of our own service page

Before, from /en/geo/: "People now ask ChatGPT, Gemini and Perplexity the questions they used to type into Google — and the answer names a handful of brands. We measure whether yours is one of them: every week we ask your buyers' questions three times on each engine and record the mentions, the position in the list and the sites each answer cites."

After: "AI visibility tracking is a weekly measurement service: we ask your buyers' questions on ChatGPT, Gemini and Perplexity three times each, and record whether your brand is named, where it sits in the list and which sites the answer cites. The questions never contain your brand name."

What changed. The first sentence now defines the service instead of describing the market. The pronouns went: "whether yours is one of them" only works while the previous sentence is still on screen. The second sentence is no longer a continuation, so either sentence can be taken alone.

Worth noting where the fault came from. The Traditional Chinese version of the same page opens with a definition, and the English one does not. The two were written to the same brief. Nobody decided to make the English lead context-first; it happened because English marketing copy reaches for a hook, and the hook displaced the definition.

2. A comparison table with no row label

Our free-versus-paid table on /en/geo/ has three columns. In Traditional Chinese and Japanese the first column header reads 項目, "item". In English and German it is an empty string.

Before	After
(blank), "Free check", "Paid monitoring"	"What you get", "Free check", "Paid monitoring"

What changed. One header cell. The rows in that column are "Engines and questions", "Frequency", "Results", "Competitors", "Trend", and until the header names them, nothing in the table says what kind of thing those are. A blank cell is invisible to a person, who infers the answer from the layout in about a quarter of a second. It is not invisible to anything parsing the table as data.

The same table has a second problem we can point at but not fix in this article. One cell reads "Mentioned or not + overall score", and the word "score" appears nowhere else on the page. The reader is never told what the score is out of, or what it counts. The rewrite is to define it or drop it, and that is a decision about the product rather than about the prose.

3. An FAQ answer that needs its own question

Before, from /en/geo/: Q. "Why ask each question 3 times?" A. "AI answers vary between runs. A single answer may or may not mention you by chance; repeated asking shows what is typical."

After: A. "We ask each question three times on each engine because AI answers vary between runs. A single answer may or may not name you by chance, and three draws are enough to show that the answer moves, though not enough to put a tight interval around the rate."

What changed. The answer now contains the number that was only in the question. An FAQ answer travels separately from its heading more often than any other kind of passage on a site, and this one, lifted out, did not say how many times anything was asked. The added clause at the end is a limit rather than a feature, which is the other thing a self-contained answer has to carry: the reason three is the number, and what three cannot buy you.

4. A sourced number, and the same number unsourced

The positive example is real, from our methodology page. The weak version is constructed for the comparison, and we have not published anything like it.

Constructed weak version: "Small samples need a better statistical approach, and studies show the standard method breaks down at low sample sizes."

Published version: Wilson's interval "can be safely employed with small samples and skewed observations" (Wikipedia, Binomial proportion confidence interval, retrieved 2026-09-20).

What changed. A named interval instead of "a better statistical approach", a named source instead of "studies show", a retrieval date, and a quotation short enough to be checked in one click. "Studies show" is the tell that no study was opened.

What was already right

We went looking for faults in two more places on /en/geo/ and did not find any. All nine FAQ headings on that page are already complete questions ending in a question mark, so there was nothing to rewrite. The "How we measure" section is already a four-step list with a bold label on each step, rather than the paragraph of prose that section usually is. We mention it because the interesting finding from an exercise like this is the pattern: on this site the openings and the table headers are where the fault sits, and the FAQ is not.

What the research supports, and what is only our hypothesis

One peer-reviewed study has tested content edits against a generative engine and published the result: Aggarwal et al., GEO: Generative Engine Optimization (arXiv 2311.09735, KDD 2024). We opened it. Every figure below is quoted from the paper itself.

Read the setup before the numbers. The paper's own engine was assembled by the authors from the top five Google results and gpt-3.5-turbo, and the headline result is that the best methods "improve upon baseline by 41% and 28%" on its two visibility metrics. That is an upper bound measured inside a reconstruction, not a result from a shipped assistant. The authors also ran a smaller check against Perplexity.ai as it existed then, where quotation addition gave a "22% improvement over the baseline" on one of the two metrics. The paper is from 2023. None of it was measured on the 2026 versions of the engines we track.

Technique	Status	What the evidence actually is
Add quotations from credible sources	Tested in the GEO paper	Best of the nine methods tested. In the group the authors describe as reaching "a relative improvement of 30-40% on the Position-Adjusted Word Count metric", inside their reconstructed engine
Add relevant statistics	Tested in the GEO paper	Same 30-40% group, same reconstructed engine
Cite your sources	Tested in the GEO paper	Same group. The paper reports "Lower-ranked websites, which typically struggle for visibility, benefit significantly more from GEO"
Keyword stuffing	Tested, and it failed	The paper finds such methods "offer little to no improvement on generative engine's responses"
Special files or markup for Google's AI features	Answered by Google, for Google's surface only	"You don't need to create new machine readable files, AI text files, or markup", and "There's also no special schema.org structured data that you need to add" (Google Search Central, AI features and your website, retrieved 2026-09-20). This is about AI Overviews and AI Mode, which are outside what we measure. It says nothing about ChatGPT, Gemini or Perplexity
Definition in the first sentence	Our hypothesis	Untested. Derived from what the retrieval step has to do, not from a measurement
Passages that survive being lifted	Our hypothesis	Untested. Same reasoning
Tables instead of comparison prose	Our hypothesis	Untested by us. The paper tested content edits, not layout
FAQ headings as complete questions	Our hypothesis	Untested. Plausible because the heading is what a question gets matched against, which is an argument rather than evidence
Source and date on every number	Our hypothesis, with an adjacent result	The paper's statistics and citation methods are close relatives, but they tested adding statistics, not attributing them

Most of this article is in the second half of that table. We have no before-and-after data of our own on any of these techniques, on any of the three engines we measure, and we are not going to imply otherwise by putting the paper's percentages next to our own advice and letting them blur together.

There is no version of this where doing the edits gets you cited. The engines decide, they change without notice, and a page can be retrieved, read and summarised with the brand behind it never surviving into the answer text. What the edits do is remove the failure modes you control: a passage that cannot be quoted because it does not parse alone, a number that cannot be repeated because it has no source, a comparison a machine has to reassemble from prose.

How to check this on your own pages

You do not need us for any of this. Take the last page you published and do the following.

1. Cut the first paragraph out into an empty file. Does it define the subject? If the first sentence starts with a trend, a question or the word "In", it does not.
2. Do the same with three paragraphs from the middle. Mark every pronoun with no antecedent inside the paragraph, and every subject that arrived from the heading.
3. Find each passage that compares two things in prose. Count how many times the reader is moved between the two. More than once is a table.
4. List your FAQ headings on their own. Any that is not a sentence ending in a question mark is a label.
5. Search the page for digits. For each number, write down where it came from and when it was read. Every number you cannot complete that line for either gets a source or comes out.

The output is a list of edits, not a score. Whether any of them moves your mention rate is a question for measurement rather than for an editor, and the way we measure is set out in full on our methodology page, including the sample sizes and what they cannot support.

Questions

What makes a passage citable by an AI assistant? It has to mean something on its own. That means the subject is named inside the passage rather than inherited from a heading, the pronouns resolve within it, and any figure in it carries a source. Whether a citable passage is more likely to be cited is a separate question, and one we have not measured.

How long should the definition at the top of a page be? Forty to sixty words is the range we write to. It is long enough to define a subject with a qualifier or two, and short enough to be taken whole rather than trimmed by whatever is doing the quoting.

Does adding statistics and quotations really improve AI visibility? It did in the one study that tested it. Aggarwal et al. found quotation addition, statistics addition and citing sources reaching "a relative improvement of 30-40% on the Position-Adjusted Word Count metric" (arXiv 2311.09735, retrieved 2026-09-20). The engine that result came from was built by the authors from the top five Google results and gpt-3.5-turbo in 2023, so treat it as evidence that content edits can move citation behaviour, not as a setting to dial in.

Does keyword stuffing work on AI assistants? No, on the available evidence. The same study reports that such methods "offer little to no improvement on generative engine's responses". It is the one technique in this area with a tested negative result.

Do I need schema markup or an llms.txt file to be quoted? For Google's AI Overviews and AI Mode, Google states that you do not: "You don't need to create new machine readable files, AI text files, or markup" (Google Search Central, retrieved 2026-09-20). That statement covers Google's surface, which is outside what we measure. No equivalent statement exists for ChatGPT, Gemini or Perplexity, so for those engines the honest answer is that nobody has published one.

Will these edits get my brand cited? Nobody can tell you that, and a supplier who does is selling you a guarantee they cannot honour. The edits remove reasons a passage cannot be used. What happens after that is the engine's decision, and the only way to find out is to measure before and after with the same question set.

How do I tell whether a rewrite worked? Measure the same question set before and after, and compare the confidence intervals rather than the point estimates. If the intervals overlap, the change is not yet evidence of anything. The arithmetic is on our methodology page.

What to do next

Run the five checks above on one page, not on your whole site. One page tells you which of the five you habitually break, and that pattern is usually worth more than the individual fixes.

Then measure before you rewrite anything else, because a rewrite with no baseline in front of it cannot be evaluated afterwards. What a baseline consists of, and what it costs, is on AI visibility tracking.