Which page an AI search answer draws on cannot be observed from outside; what a site can measure is whether its own sentences read correctly once lifted off. The work therefore lands in fields that get printed again without the page: card summaries, definition lines and FAQ answers travel alone. A claim we cannot source stays off the page, and one of our comparison tables still carries an empty case cell.
Is the sentence still true and complete once lifted off
Whether a sentence that travels without its page still reads true and complete on its own is the only thing measurable on the page’s side, and it takes no tool beyond closing the page. One sentence goes in, yes or no comes out. No dashboard, no report, no ranking figure.
Where a given sentence ends up, and next to what, is something we cannot see, so we produce no claim about it. The same limit is written into our own pages. In one comparison table on this site, a row we left empty because no source could be shown is still empty. An empty cell carries information; a filled cell with nothing behind it looks exactly like a sourced one once it has been lifted out.
The test does not apply to every sentence. Inside a body paragraph, sentences lean on each other and should: “that distinction” and “so” are fine there, because a body paragraph only ever lives on its page. Units that can be read without a body are different. The definition line, the card summary and the FAQ answer each exist as their own field in the schema and get printed again where the page is absent: the definition line in llms.txt, the card summary in listing cards, the answer in the FAQPage schema node. A takeaway is a schema field too, but the only place printing it is the article page itself. Whether a first sentence or a table cell is read without its body is not something we can see, so we write both to be correct without one. A field printed without its page has to be correct without its page.
Four conditions for a lifted sentence
A sentence lifted off its page stands on four conditions: its reference does not stay behind on the page, it carries a fact rather than an adjective, it names the system it is about, and it holds a single idea.
The first condition is about reference. A sentence opening with “this approach”, “as described above” or “on this platform” leaves its referent behind the moment it is lifted. The rule is mechanical: a demonstrative plus an abstract noun does not belong in a sentence carrying evidence. Our own writing rules call this a backward reference and ban it there.
The second condition is about facts. “Offers a powerful archive” answers nothing when you ask what happens. “Today’s forecast is not deleted when it expires; it becomes an archive row” does answer it: a record, a behaviour, and the direction the behaviour runs in. Only one of the two leaves information behind when it is cut loose.
The third condition is about the subject. Not “the system” but which system. In an evidence sentence the project or brand name is the subject and the sentence opens with it, because a name landing mid-sentence gets parsed as an aside rather than as the thing the claim rests on. One of our build checks looks for that shape directly.
The fourth condition is one idea per sentence. Join two facts with “and … because …” and the sentence produces a causal link that exists nowhere in the record; cut it in half and what survives is a claim nobody made. Çatalca Merkez Kasap’s project record keeps two such facts as separate entries, and merging them makes the connection the writer’s rather than the record’s.
The four conditions do not apply to body paragraphs and they are not a score. “Which languages do you work in?” answered with “Turkish and English.” stands without naming a subject.
A meta description earns the click, a definition line defines
A page’s meta description is written to earn a click from a search result card and is read together with the page; the same page’s definition line says what the page is with no page around it. Neither substitutes for the other.
Both sentences sit side by side on the project page where the definition line and the meta description can be compared directly. The meta description names nobody and opens on an anonymous “a weather site”: one bringing forecasts, dated records and climate summaries together. The definition line makes the site itself the subject: “gunlukhavadurumu.com.tr is a weather site publishing 15-day forecasts, dated historical records and climate summaries computed from that archive, per date and location.” Lift the first one off and it is unclear whose site holds what. Lift the second one off and it is still a definition.
On this site llms.txt is generated the way the sitemap is: a title, an address and one sentence per page, produced at build time and never typed by hand. The file is not a standard or a requirement; the format is one llmstxt.org proposes, and who reads it is not something we assert.
On the insights collection the definition line is a required schema field, and a missing one stops the build. While the field was optional, a missing line filled itself with the card summary and nobody noticed; the fallback hid the defect. The check catches identical text in the two fields and nothing finer: a difference in meaning is not measured.
The rule for a page’s opening sentence follows from this. It defines; it does not tease. A rhetorical question or a scene-setting anecdote postpones the answer until somewhere inside the page, and the lifted opening then leaves a tone of voice behind rather than a subject.
A definition line does not do the click’s work either: rewrite a meta description as a definition and the search result card goes flat. Even here the rule is not applied to every collection equally.
An FAQ answer travels without its body
An FAQ answer is read without its body in the FAQPage schema node, which prints the same text as the accordion and carries nothing else. An answer beginning “as explained above” loses its reference the moment it is printed there.
The first sentence answers, and it answers by echoing the verb of the question: “Does the user have to wait on screen?” is met with “No one has to.” rather than a bare yes or no. A page opens with a definition; an FAQ answer must not. One introduces a page, the other closes a question, and the same opening move fails in the second job.
The answer carries four steps: the direct answer, what the reader gets from it, the evidence, and the next move. In the evidence step the project name is the sentence’s subject. Drop a step and the answer gets shorter, which makes the problem look like length; what is missing is a step. There is a floor and a ceiling per language, and an answer under the floor is never padded to reach it.
Tense rides inside the answer too. In an article whose card summary and Secretum answers each carry their own tense marker, every answer about the product says the arrangement is one we are putting together now. Secretum, which we are building at the moment, is not live, and a tense stated once in the body never reaches an answer that travels without it. Leave the marker out and the lifted answer reads as though the product shipped.
A caveat belongs in the same place. An answer saying the outcome depends on the platform still has to say what depends on it and which decision was taken in which piece of work. Put the caveat in place of the answer and the question stays unanswered.
Not every answer carries all four steps: a closed-list answer needs no evidence, and in a blog FAQ the evidence step is optional. The rule is “declare the missing step and add no filler”.
Does the claim sit on the page beside its source
An outside fact is written with the source’s own words in quotation marks, the institution’s name and a link to the page where the quote lives, all in the same paragraph, so the sentence carries its source when lifted out. The link points at the page holding the claim, not at the institution’s front door, and one source gets one link per article.
That is how it is built in the article where a Board announcement is quoted with the institution named and linked in the same paragraph. Quotation marks are a boundary rather than decoration: they separate what the source said from what we concluded. The unsourced version of the same fact sits in the last row of the table below.
One note about language: where the source is Turkish and the article is in English, we report its wording instead of quoting it, because a translation is our sentence and not the source’s. The institution is still named and the announcement still linked.
For an internal fact the source is your own record. At Günlük Hava Durumu a climate summary is not fed in from a separate source; it is computed from the day records in the archive, and every past day sits at a permanent address of its own. How the summary is computed out of the archive was written up separately. What matters here is that a derived summary stays attached, at its own address on the same site, to the records it derives from.
Where no source can be shown, the sentence does not get written. In the table mentioned above, the cell where we could not give a definite answer says exactly that, and the case column beside it was left blank.
The reader’s side of this has its own article: where a question goes when nobody stands behind the answer.
Demanding a source for every sentence turns an article into footnotes. The rule covers factual claims: a decision, a figure, an institution’s words.
The same fact, built to stand and built to fall
A sentence that falls when lifted is not a wrong fact but a flawed construction of a right one. Every row below puts two constructions of the same fact side by side. The left column holds no invented example; it is the right-hand fact, broken.
| Construction that falls when lifted | The flaw | Construction that stands when lifted | Where the fact comes from |
|---|---|---|---|
| A weather site bringing 15-day forecasts, dated historical records and climate summaries together. | The subject is unnamed: lifted out, it is unclear whose site holds what. In its own slot, a search result card, the sentence is correct. | gunlukhavadurumu.com.tr is a weather site publishing 15-day forecasts, dated historical records and climate summaries computed from that archive, per date and location. | The Günlük Hava Durumu project page |
| On this platform new and used listings share one list. | Backward reference: lifted out, “this platform” no longer says which platform. | At Plastik ve Makina new and used listings share one list, and a buyer switches condition on to compare both options inside a single search. | The Plastik ve Makina project page |
| AlgoKlima offers a powerful offline experience. | An adjective with no fact behind it; ask what happens and the sentence has no answer. | At AlgoKlima the record lands in the phone’s own queue first, and the app keeps working without waiting for a signal. | The AlgoKlima project page |
| Çatalca Merkez Kasap’s QR menu gives every section its own address and needs no reprint when a price changes, because the page loads once. | Three ideas in one sentence, and “because” builds a link the record never claims. Cut in half, what remains is false. | Çatalca Merkez Kasap’s menu is text on a page, so a price change is an edit rather than a print run and the same QR code stays valid. | The Çatalca Merkez Kasap QR menu project page |
| At Secretum image generation runs in a server-side job queue and the user keeps moving through the feed. | The tense has dropped: a product in development reads as live, an intended target as a current behaviour. | Secretum is being built so that an image is generated in its own job while the person keeps moving through the feed; that is the arrangement we are putting together now. | An FAQ answer in our article on queued AI generation |
| The Board banned fingerprint-based timekeeping. | No source, and no way to tell which board; “banned” is a certainty the source does not express. | The Personal Data Protection Board’s announcement on processing biometric data for timekeeping purposes states that, since there is no explicit statutory provision foreseeing attendance tracking through biometric identification systems, carrying out timekeeping by processing biometric data may constitute a breach of the law. | The Board’s public announcement, as reported in our article on employee monitoring |
The flaw sits inside the sentence rather than in wherever the sentence ends up. The left cell of the first row is a correct sentence in its own slot, because earning a click is what a meta description is for; printed as a definition line, the same sentence falls.
The right cell of the third row comes from the project page where a record is written to the phone’s local queue first, with its subject moved to the front. One construction tells you where a record is written; the other does not.
The fourth row is the one-idea condition at work. In the project record where the unit of address is the section rather than the item, two facts are kept as separate entries: sections carry their own addresses, and a price change needs no reprint. The left cell welds them with a “because” the record never claims.
Read the table against your own text: find the row whose flaw matches your sentence and rebuild it the way the right column does.
One address answers one question
At Plastik ve Makina, where every listing sits at its own address as real text, the list is prepared on the server. Open a listing from a forwarded message or from a search result and the page is not waiting for the browser to run anything before the text is there.
The second fact sits on the record side: whether a machine is new or used is a field of the same listing record rather than a second catalogue, so the seller writes one text and no duplicate record is created. A listing page then answers which machine and which condition on its own, without the list around it.
The same platform’s category tree and condition filter were opened up separately; the question here is what the unit of an address is.
The second face of the same question shows up at Günlük Hava Durumu, where every date and location combination is published at its own lasting address, so the site gets deeper as the archive fills instead of being restructured. The unit of an address is the unit of the question, not the unit of the record.
The counter-case is in the same portfolio. In Çatalca Merkez Kasap’s menu the unit of address is the section rather than the item: sections are separate addresses on one page, and different QR codes land on different parts of it. Which scope was deliberately left out when a printed menu moved into software is a separate write-up.
Saying that browser-built lists are bad would contradict our own evidence: the interface of our own quote form runs in the browser. What matters is whether the question held by the person opening the link is answered at one address.
Running the test by hand
To run the test by hand, copy the field, close the page and ask four questions of what is left: which system, what happens, where is the source, and what tense. Those four map onto the four conditions in order.
Which fields you read depends on what your site has. A first sentence exists everywhere; card summaries and definition lines get reprinted in listings and in files; an FAQ answer, if you have one, is read in the schema node with no body around it. With no FAQ block the test gets smaller without disappearing: the first sentence and the card summary are still there.
The formal trace of the same test is wired into our own build. The definition line is required and cannot be identical to the meta description. The length of an FAQ answer, the length of its first sentence, whether a project name opens its sentence and whether the answer addresses the reader are all measured. So is the rule that two pages may not target the same key phrase, which is the mechanical form of “one page, one question”.
What the checks measure is shape: identical strings, lengths, sentence openings, phrase uniqueness. They measure neither a difference in meaning nor whether a fact is true. The manual test replaces the check; the check does not replace the manual test.
One more detail, because it happened to us. Every check reports how many units it scanned, so a check that scanned nothing cannot print “clean”; it prints “not inspected” and the build exits with an error. The rule was broken once: the old parser behind the key phrase check shifted a column, never read the English one at all, and went on reporting clean. “Zero errors” and “zero checks” look identical from the outside; the counter is what separates them.
Correct these four fields and what you keep is sentences that read true and complete without their page; beyond that we cannot measure anything and we promise nothing. How each question gets its own address and each field reads correctly without its page is what our web development work is built around. If a system is already running and these fields have to be fixed without disturbing the parts that work, working inside an inherited codebase is a job of its own.
