When an assistant answers a question about your industry, it is not reading your website the way a customer does. It is looking for sentences it can lift: short, self-contained statements that carry a fact, attach it to a named thing, and survive being pulled out of the paragraph around them.
Most business websites contain almost none of these. They are full of sentences that mean something only in context, hedge every number, and refer to the company as "we" on a page that never says who "we" is. That writing is not bad. It is just unusable by a system that reads one page and has to produce one sentence.
This is a writing problem before it is a technical one, and it is fixable in an afternoon per page. Here is what to change.
The unit is the sentence, not the page
The mental model to adopt: someone is going to quote exactly one sentence from this page, out of context, to a person who has never heard of you. Which sentence would you want it to be? Is that sentence actually on the page?
Most of the time it is not, because the information is distributed. The price is in a table, the delivery time is in a different section, the thing you do is implied by the page title. A human assembles those. An extractor takes what is complete.
So write the complete version somewhere on the page, even if it repeats what the layout already communicates.
Seven rules
1. Make every important sentence self-contained
Weak: "Our standard turnaround is two weeks." Quotable: "Odilient delivers a landing page in three to five days."
The first sentence needs the page to explain who "our" is and what is being turned around. The second survives extraction. This is the single highest-value change and it applies to every claim on your site.
The habit to break is the pronoun. "We", "our", "this", "it" — each one is a dependency on surrounding text.
2. Put the number in the sentence with the thing it describes
Weak: "Pricing starts low and scales with complexity." Quotable: "Landing pages start at $99. Corporate websites start at $299."
Ranges and qualifiers are what make a number unquotable. If your real answer is "it depends", give the entry point as a fact and let the dependency be a separate sentence. "From $99" is a fact. "Affordable" is not.
This is the practical extension of an argument we have made before: a price you do not publish is a price a model will quote from your competitor. See publish your prices.
3. Name yourself, not "we"
At least once per page, and always in the first paragraph, use the actual name. "Odilient is an engineering studio based in Uzbekistan."
This sounds unnatural to write and it is how an extractor learns that this page is about a specific entity. It also matters for consistency: the same name, spelled the same way, on every page, is how a model assembles one entity instead of several partial ones.
4. Answer the question in the first sentence under the heading
Not after a paragraph of context. If the heading is "How long does a website take?", the sentence below it should contain a duration.
Journalism calls this not burying the lede. For machine extraction it is stronger than a preference: content that answers late frequently does not get read that far, because the extractor has already found a candidate on a competitor's page that answered immediately.
5. Use headings that are the question
Weak: "Our Process" Better: "What happens after you get in touch"
People search and ask in questions. A heading that matches the question is a strong signal that the text below it is the answer, and it also makes the page better for the human skimming it — which is the reliable test for all of this.
6. Give units, dates and qualifiers explicitly
"Two weeks" not "a couple of weeks". "$299" not "three hundred dollars". "As of September 2026" on anything that could age.
An unqualified claim is one a careful system will decline to use, and a stale claim is worse than a missing one. Where a fact has a shelf life, date it in the sentence.
7. Admit the limits in the same voice
A page that says what it cannot do is more useful to an extractor and more credible to a reader.
We do this on our own work and it costs us. Our AI receptionist is described on its own page as working end to end as a browser demo only, with the production voice flag off, no telephony leg and no phone number. That is not modesty; it is the difference between a page that can be quoted accurately and one that will eventually be quoted wrongly.
Structures that extract well
Beyond the sentence, four shapes carry information cleanly.
A definition line. "X is Y." Early on the page, in plain terms. Your homepage should contain a sentence that would be a reasonable dictionary entry for your business.
A comparison table. Rows and columns are unambiguous about what belongs with what. A price table with plan, price, delivery time and what is included does more for machine legibility than three paragraphs saying the same thing.
A numbered process. Steps in order, each with a name and a duration.
A genuine FAQ. Question as a heading, answer immediately below, one topic each. This is the highest-density quotable structure available, and it is usually written badly — we go into how in your FAQ page is doing two jobs badly.
Two of our own products publish structured versions of exactly this kind of content because extraction was the point: BiliMind emits schema.org/Course on every public course, and RecipeReel exports schema.org/Recipe JSON-LD. The structured data and the prose should say the same thing — see structured data is not an SEO trick.
The habits that make a page unusable
Four things we see constantly, each of which makes a page invisible to extraction.
Marketing abstraction. "Solutions that drive results." There is no fact in that sentence. There is nothing to quote, nothing to verify, and nothing that distinguishes you from anybody.
Text inside images. A price in a graphic, a comparison in a screenshot, a list rendered as a picture. Unreadable to extractors, unreadable to screen readers, unsearchable, and it does not reflow on a phone. If it matters, it goes in the page as text.
Contradiction across pages. Your homepage says five days, your services page says a week, your FAQ says "usually under ten days". A reader tolerates this. A system assembling an entity has to pick one, and it may not pick yours.
Content that exists only to rank. Long pages padded to a word count, saying the same thing six ways. An extractor working over this finds no dense factual sentence to take, because there is not one.
The consistency audit
This is the exercise worth doing, and it takes an hour.
- List the ten facts a customer most needs: what you do, where, what it costs, how long it takes, what is included, how to reach you, what you will not do.
- For each, find every page on your site that states it.
- Write down each version, verbatim.
- Look at the differences.
The differences are what a model has to reconcile, and where they are irreconcilable it will either pick one or decline to answer with any of them. The output of this exercise is a single canonical phrasing per fact, used everywhere.
We run this on our own site with a rule that is stricter than most people need: every fact the blog states about the studio — a price, a timeline, a case study number — has to come from one source document, and anything not in it does not get written. It is unglamorous. It is also why the same delivery time appears in the same words in four places.
The test
Take any important page. Copy the text. Ask an assistant: "Based only on this, what does this company do, what does it cost, and how long does it take?"
If the answer is vague, the page is vague. If the answer is wrong, the page contains a contradiction. If the assistant says the text does not specify, then it does not specify — and no amount of technical SEO will fix a page that does not contain the fact.
That test takes two minutes and it is more useful than most audits.
We build conversion-first websites from $99, and copy that persuades is a named deliverable rather than something left to the client to supply. Writing for a reader and writing for an extractor turn out to be nearly the same discipline: say the specific thing, in a complete sentence, near the heading that asked for it.
Tell us what you need. You get a fixed price and a delivery date back — free, and with no obligation to proceed.
Frequently asked questions
Does writing for machines make the page worse for people?
No, and this is the happy part. Every rule above — specific numbers, complete sentences, answers near headings, questions as headings, stated limits — makes a page easier for a human to skim. The writing that fails extraction is also the writing readers dislike: abstract, hedged and slow to the point.
Should I repeat facts on multiple pages?
Yes, in the same words. Repetition across pages is not duplicate content in any harmful sense when it is a factual statement, and consistency is worth more than variety here. Vary the surrounding prose, not the fact.
How long should a page be?
As long as it takes to answer the question it is about. Padding to a word count is the specific failure described above. A short, dense page that answers completely beats a long one that circles.
What about writing in more than one language?
Say the same facts in each, and do not let one language drift into being the detailed version. On Bronjoy we treat English leaking onto a Russian or Uzbek page as a bug, and directory vocabulary comes from one shared dictionary — which is the multilingual version of the consistency audit above.
Does any of this help with ordinary search too?
It helps with both, for the same underlying reason: clear, specific, consistent statements about a named entity are what every system reading your site is trying to find. The techniques diverge at the margins. The core does not.
