AI citability checker for your content

This page analyzes the text. If your question comes earlier, whether the crawler can reach the page at all, the GEO readiness checker measures the access layer for the whole site, and the llms.txt generator handles the file at the root of the domain. If what is missing is markup, the FAQ schema generator and the Article schema generator build the JSON-LD. Practical order: access first, citable text second, markup last. Marking up a paragraph nobody can lift solves nothing.

What generative search does to your text

An AI system does not read your page the way a reader does. It cuts the content into chunks, stores each chunk separately and, when answering, retrieves the chunks that look like they answer the question. What reaches the final answer is not your article: it is a 40 to 60 word snippet that has to make sense on its own, away from the site, with no section heading, no previous paragraph and no chart caption to lean on.

From that comes the only question worth asking before you publish: is there a piece of your text that answers the question and stands on its own? In most marketing posts the answer is no. The information is there, but scattered: the promise sits in the title, the condition sits in the third paragraph and the number sits in a table with no caption. For a patient human reader that works. For a snippet, there is nothing citable left.

How to use it

1. Write the target question the way a user would type it, not the way you would headline it. It is worth 12 of the 100 points and it is what exposes answers written entirely in synonyms.
2. Paste one section at a time, not the whole post. One H2 and the block under it is the right unit: that is roughly how systems chunk.
3. Read the extracted snippet before you read the score. If you cannot understand that block without context, no model will either.
4. Work the fix queue top down. It arrives sorted by points lost, so the first item is always the best return per minute.
5. Paste the corrected text back and compare the two scores. The gap is the only honest evidence that the rewrite changed anything.

How it works: the 9 items and their weights

The score follows the order in which a system crosses the text: first it looks for the answer, then it tests whether the chunk stands alone, then whether it fits in a citation, and only then whether it deserves trust.

The opening carries 28 points because it decides whether a snippet exists at all. The stands alone item punishes two things: starting with a connector (this, therefore, also) and referring to another part of the page (as we saw, the table above). Both are invisible to someone reading the whole post and fatal to someone reading only the chunk. Structure is worth 10 because a list or heading gives the system a clean cut point, and source is worth 8 because an attributed claim is cheaper to verify than a floating one.

Worked example (reproduces the default result)

The text loaded by default is a real case from a conversion blog: three honest paragraphs, written to be read in full, and for that exact reason impossible to lift. It runs 101 words, 3 sentences and 3 paragraphs, and the score lands at 41 out of 100, band D.

• Answer in the 1st sentence: it opens with In this article we are going to, so the 10 warm up points are gone. It runs 31 words, which yields 6 x (45 minus 31) / 20 = 4.2 of 16.
• Question terms: both test and run show up in the snippet, so 12 of 12.
• Stands alone: no connector in the opening and no reference to the rest of the page, 12 of 12.
• Sentence: 33.7 words per sentence, close to the ceiling of 38, leaving 2.9 of 12. Paragraph: 33.7 words, inside the ideal of 60, 10 of 10.
• Verifiable data: not a single number in the text, 0 of 12. Attributed source: none, 0 of 8. Structure: no list, heading or table, 0 of 10.
• Firm language: 5 vague terms in 101 words is 4.9 per 100, nearly double the 2.5 limit, so 0 of 8.
• Total: 4.2 + 12 + 12 + 2.9 + 10 = 41.1, which rounds to 41.

The diagnosis is plain: the text knows the answer and never says it. Now the same content, same information, reorganized by the items in the queue:

## How long should an A/B test run

An A/B test should run for whole weekly cycles, until it reaches your planned sample size.

For most stores that lands between 2 and 4 weeks. Three rules close the count:

- Run at least 1 full weekly cycle, Monday to Sunday: whoever buys on Tuesday is not whoever buys on Saturday.
- Stop when you reach the sample size you calculated, never at the first sign of a win.
- Do not go past 4 weeks. Beyond that, cookie loss and seasonality eat the reading.

Example: a store with 20,000 visitors per week and a 2.5% conversion rate needs 21 days to detect a 10% relative lift at 95% confidence. Source: the Donnu A/B sample size calculator.

This second version runs 127 words, 8 sentences, 15.9 words per sentence on average, 9 numbers and one declared source, and it scores 100 out of 100. Nothing was invented: the gain came from cutting 31 words of warm up, breaking the long sentences, turning the three conditions into a list and swapping a few weekly cycles for between 2 and 4 weeks. Same research, written to be lifted.

How to read it, and where the score misleads

The score measures form, not merit. A wrong text, well structured and full of numbers, scores 100 without trouble. No client side tool verifies whether your data is true, and that part stays human work: the number you cite has to exist, and the source you attribute has to say what you claim it says.

The second trap is farming points. Stuffing irrelevant numbers to inflate data density produces a worse text and a better score, which is the worst of both worlds. Same for lists: turning prose into bullets with no real hierarchy makes the text easier to cut and harder to understand. If the fix does not improve the text for a person, it is not worth the point.

The third is the limit of the method. The analysis is lexical, so it does not catch irony, does not know whether your this has a clear antecedent and does not judge whether the answer is right. It is very accurate on structural diagnosis, which is where almost everyone loses points, and it misses on edge cases. Treat the result as a reviewer that never gets tired, not as a verdict.

The fourth is confusing citation with traffic. Being cited in a generated answer does not guarantee a click, and in search with the answer already on screen a good share of sessions end right there. The gain from being in the answer is brand presence plus the slice of people who click to verify. If your goal is sessions alone, the effect of clickless search is detailed in zero-click search optimization.

From text to evidence

Rewriting for citation is a bet on how third parties treat your content, and a bet with no measurement turns into team folklore. The short path:

• Record the score before and after each rewrite, with the date. With no baseline, any later variation becomes free interpretation.
• Track mentions and citations, not just position. The full method is in generative engine optimization, and the content structure that earns the most citations is in FAQ content for AI citation.
• When the rewrite touches what the visitor sees, and not only what the machine reads, measure it with an A/B test before rolling it out site wide. Cutting the introduction moves conversion, not just retrieval.

To turn the first item in the queue into a testable hypothesis, use the hypothesis generator. To size the experiment before you switch anything on, the sample size calculator settles it in a minute. And the relationship between both fronts is in does A/B testing affect AI citations.

FAQ

What makes content citable by AI?

Being liftable. An AI system does not cite your page, it cites a piece of it, and that piece has to answer the question on its own, without leaning on the previous paragraph or the section heading. In practice that is five things: the answer shows up in the first sentence, the snippet holds up away from your site, the sentence is short enough to fit in a citation, there is a verifiable number anchoring the claim, and the language asserts instead of hedging. That is exactly what this tool measures.

Does the score predict that AI will cite me?

No, and be suspicious of any tool that promises that. No AI system publishes its real source selection criteria, and the choice depends on things outside your text: domain authority, what competitors published, the history of the user conversation. What the score does is measure the form of the text, which is the only part of the equation you control. A text scoring 100 still competes with others; a text scoring 30 does not even enter the competition, because there is no snippet to lift.

Why is the first sentence worth so much?

Because it becomes the snippet. When a retrieval system chunks the page, the top chunk carries the promise of the title and is the one most likely to be read in full. If the first sentence is warm up, the model has to hunt for the answer mid text and compete with whoever answered up front. The 16 points split into 10 for the absence of warm up and 6 for length: up to 25 words takes all of it, and it hits zero at 45.

Does the tool use AI to analyze my text?

No. The analysis is lexical and structural: word counts, sentence and paragraph splitting, matching against a closed phrase list per language, and detection of numbers, lists and source markers. No model is called and nothing leaves your browser. That carries an advantage an AI analysis would not have: the result is reproducible. The same text always returns the same number, so you can use the score as a baseline before and after a rewrite.

What counts as vague language?

Phrases that weaken the claim without adding information: generally, tends to, usually, somewhat, depends on many factors, in most cases. They exist to protect the writer, not to inform the reader, and a system that has to pick one source prefers the one that asserts. The list is closed and different in each language. The item hits zero once the text passes 2.5 vague terms per 100 words. The fix is almost never deleting the caveat: it is swapping the adverb for a number and a condition.

Do I really have to declare the target question?

You do, because without it 12 of the 100 points stay at zero and you lose the most useful diagnosis in the tool. Coverage compares the question terms against the snippet, and the failure pattern is always the same: the text answers the question in synonyms and never writes the words the person actually typed. Writing the question before writing the answer also fixes half the structure problems, because it forces you to decide what the section promises.

Should I analyze the whole post or one section?

One section at a time, which is how systems read it anyway. Pasting 3,000 words returns an average that hides the thing that matters: the opening of each block. The use that pays off is taking every H2 in the post, pasting that section with the question it promises to answer, and fixing whatever the queue points at. A post with eight sections becomes eight short analyses, and each one is a snippet candidate.

Does writing for AI make the text worse for humans?

No, and the evidence predates AI search. Everything the score rewards, answer up front, short sentences, a number instead of an adjective and paragraphs that fit on a screen, is what good web writing already asked for. What the score punishes is the habit of padding the introduction so the piece looks more complete. Cutting the warm up wins on both fronts: the human reader finds the answer before giving up, and the machine finds a snippet to lift.

Keep going

With the text liftable, the next step is the layer it does not solve: crawler access, markup and authorship signal outside your domain. Read the full method in generative engine optimization and what changes in conversion work in CRO in the age of AI.