If a tool scores your page and won't show the maths, it's guessing.

The idea

An AI answer engine does not read a page the way a person does. It chunks the page, pulls out passages, and reproduces the ones that answer the question on their own. Ellipsis put it in one line in August 2026: a page earns its citation one sentence at a time. A claim that only makes sense inside its paragraph never travels.

So Lift splits your text into sentences and asks of each one: if this were the only thing the engine saw, would it understand it, and would it use it? The three verdicts:

Sentence checks

Stands alone

Fails when the sentence opens with a pointer to something outside itself: this, these, that, it, they, here, the former, as mentioned above. Also fails when it opens with we, our or I, because an engine cannot attribute "we" to a brand it hasn't been told about. Soft-flags a sentence that uses "we" anywhere without naming the brand.

Source: Ellipsis, GEO vs SEO (Aug 2026): "A claim that only makes sense in the paragraph around it cannot be lifted out of that paragraph, so nothing gets quoted." The blind test on the front page is the same rule run by an actual model.

Continues a thought

Soft flag for a sentence that opens with a connective (but, however, also, for example, in addition, this means). It reads as a hinge to the previous sentence. Usually the fix is deleting the first word.

Source: same as above. Connectives are the mildest form of leaning on context.

Length

Pass at 12 to 45 words. Soft flag under 12 or from 46 to 60. Fail under 6 or over 60.

Source: MaxAEO, 3,200 cited passages across eight engines (early 2026): median quoted passage 28 words, 70% of verbatim pulls between 15 and 45 words. Ellipsis describes citations as "usually around twenty words, standing alone".

Carries something checkable

Pass if the sentence contains a statistic, a number with a unit, a year, a currency amount, a definition shape ("X is a…") or a proper noun beyond the first word. Soft flag if it names a subject but makes no measurable claim. Fail if there is nothing a reader could verify.

Sources: Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024): adding statistics, quotations or cited sources improved visibility by 30-40%; keyword stuffing did nothing. Kevin Indig's analysis of 1.2M ChatGPT responses (via ALM Corp): cited passages average 20.6% proper nouns against 5-8% in ordinary text, and "X is" constructions are cited about twice as often. MaxAEO: statistic lines pulled 3.4x, definition sentences 3.1x, table rows 2.7x, list items 2.2x more than plain narrative.

Names its subject

Fails when nothing in the sentence says who or what the claim belongs to: no proper noun, no defined term, no named source. "The standard VAT rate is 20%" is true and checkable and still fails, because alone it doesn't say which country. Soft flag when the sentence points at a source without naming it ("according to our survey").

Source: Ellipsis, AI isn't a black box: owned claims are the ones "anyone can go on your website and check", attributed 85% of the time or more; unowned claims get credited to the category instead. In Lift's own blind tests, "name the company, country or source" was the model's single most common reason for not quoting a sentence.

Unsourced figure

Soft flag for a statistic with no named source and no named subject. The number will be paraphrased into the answer; the sentence won't be quoted.

Source: Aggarwal et al.: citing sources and adding statistics were the two strongest methods, and they compound. Indig: cited posts that combined an answer capsule with original data were cited most.

Words AI drops

Flags evaluative language: best-in-class, cutting-edge, seamless, world-class, industry-leading, innovative, robust, powerful and about sixty more. Fail if the sentence has nothing checkable to go with it; soft flag if it also carries a number.

Source: Ellipsis, AI isn't a black box (May 2026): concrete, falsifiable claims were attributed to the brand 85% of the time or more; "enterprise-grade" 42%; "best-in-class" 11%; "cutting-edge" 6%. The engine "generates its own evaluation vocabulary" instead of quoting yours.

Unquantified

Soft flag for many, most, numerous, countless, significant, a range of when no figure is present. A number in that slot would make the sentence a statistic line.

Source: the statistic-addition result in Aggarwal et al. above.

Throat-clearing

Fail for openers that delay the claim: In today's…, It's worth noting…, In this article…, Let's…, Have you ever…, Welcome to…, When it comes to… and similar.

Sources: HiBob's SEO team, reported by Tom Orbach (Marketing Ideas, Feb 2026): "We stopped writing intros." Pages answering in the first sentence were cited more. Ellipsis: "AI pulls hardest from the top of a page and from the sentence straight after a heading."

Call to action

Fail for contact us, book a demo, get started, learn more, sign up and similar. An instruction to the reader is not a claim about the world. Keep them; just don't count on them.

Question

Fail for a sentence ending in a question mark. Questions are quoted as headings, not as answers. The check on headings below rewards the opposite.

Source: Indig (above): question-form headings were cited about twice as often, and 78% of those came from H2s.

Written to the reader

Fail for a sentence that addresses "you" and carries no statistic or definition; soft flag if it does. Engines reuse claims about the world. Conversation with a reader gets paraphrased away, if it's used at all.

Source: Ellipsis, Best AI SEO Agencies: descriptions written as self-contained statements about the subject are the ones that travel. Indig: the optimal subjectivity score for cited text was 0.47, meaning cited passages are mostly factual rather than opinion or address.

Contains a link

Soft flag. Links inside the sentence are noise to the extractor.

Source: Indig (above): about 91% of cited answer capsules contained no links.

Position

Not a pass/fail. The first and second sentence under a heading, and the first sentence of a paragraph, are noted because that is where most pulls happen. Position also feeds the ranking of "most likely to be quoted".

Sources: MaxAEO: 64% of quoted passages were the first or second sentence of their section or paragraph. Indig: 44.2% of citations come from the first 30% of a page.

Page checks

Live URL checks

These only run in URL mode, because they need the real page.

The model read

This is the main event. The whole page goes to Claude Opus 5 with one instruction: you are an answer engine and you have retrieved this page as a source; say what you would do with it. It returns the sentences it would quote verbatim and the question each one answers, the claims it wanted to use but wouldn't as written (with a rewrite that keeps every fact and invents none), and the questions the page is close to owning.

In parallel, a blind test sends up to sixteen sentences (the strongest quick-read candidates plus the highest-value failures) to the same model with no surrounding text at all, one at a time. That is the pure form of the mechanism: does the sentence make sense with nothing around it? Where a quoted sentence also passed the blind test it is marked "survives alone"; where it only worked with the page around it, the card says so.

Your text is sent to Anthropic's API once for the read and is not stored by Lift. The rules and the model disagree sometimes; when they do, trust the model, then tell me, because that is how the rules get better.

What Lift can't tell you

Whether AI actually cites you. That takes tracking real answers over time against real buyer prompts, which is what products like Ellipsis's FALCON do. Lift is a pre-publish edit check: it tells you which sentences are built to travel, not whether they arrived. All of the research above is correlational, the engines change monthly, and a sentence that lifts clean can still lose to a better page. Treat every threshold as a starting point for an argument, not a rule.