Publishing articles that get found: what to do after the site is built

A site nobody arrives at is a brochure in a drawer. An article gets found when it answers a question a real buyer types, in its first two sentences, on a page a crawler can reach and a person can scan. It flips in one case: if you already own the single page your market searches for, and it already brings orders, twenty near-identical variations will not add reach — they compete for the same query.

What does one page per real question mean?

The unit of publishing is a question, not a keyword. "Cash on delivery" is a keyword. "Who pays the courier when a parcel comes back refused" is a question, and the answer changes what a shop does on Monday morning. The second earns a visit from someone with a parcel to argue about; the first earns a visit from someone who is not yet buying.

So keep two lists: what customers actually asked — the last fifty enquiries, in their own words, including the voice notes — and what you have published. The gap between them is your calendar. Three to six real questions make one article.

Why must the first paragraph answer before it introduces?

Because the part of a page most likely to be lifted and shown elsewhere is a self-contained passage, and a passage that opens with "in today's fast-moving market" contains nothing to lift.

Bad: "Choosing how to publish content is an important decision that many businesses face."

Good: "A sitemap is a claim that a page exists. A link is a recommendation that it matters."

The second sentence can be quoted; the first is throat-clearing. Write the answer, then the condition under which it changes, then the evidence.

Do AI Overviews and AI Mode need a special file?

No. Google documents that a page needs to be indexed and eligible to be shown as a snippet; no additional markup, feed or file is involved. The details are on the page about AI features and your website. Anything sold to you as AI optimisation that consists of adding a file is effort taken away from the page.

What Google does list there is unglamorous: crawlable content, internal links, page experience, text that is really text, and structured data that matches what the visitor sees.

How do internal links decide what gets read?

Google documents links as one of the ways it discovers pages and works out what a page is about: how Google uses links. A page nothing links to exists for the sitemap and for nobody else.

On a bilingual site rendered from Markdown, that recommendation usually lives in frontmatter rather than in the body, because the same body appears under both languages. Three to five siblings per page, in both directions: the piece about refused parcels points at the piece about delivery cost, and the delivery piece points back. A hub names its children; every child names the hub. That is a cluster, and it is the part of discovery you fully control.

What do the sitemap and hreflang actually assert?

The sitemap asserts that these URLs exist and when they last changed. hreflang asserts that one URL is the Arabic version of another. Both are machine-readable claims, and both fail quietly: a sitemap entry for a page you unpublished, or a language pair whose other side was never written, sends a reader to a page that is not there.

Two rules survive every redesign. Arabic gets its own URL carrying Arabic text, never an English page with a language tag. And the canonical of the Arabic page is the Arabic page.

What should you measure, and where?

Search Console, and mostly one report: queries. A query that collects impressions and no clicks is a page that appeared on a results screen and did not earn the visit. That is your next article, or your next title. A high position with few clicks usually means the description summarises the page instead of answering the question.

Average position is a poor target. It averages queries with different intent, and it can improve while the traffic that pays for the site gets worse.

Why do twelve pages beat a hundred?

A hundred thin posts produce near-duplicate titles, split the links that would have made one page authoritative, and give a reader nothing to trust. Scaled content whose main purpose is to manipulate rankings is a spam policy violation, and it applies whether a person or a model wrote it — that is in Google's guidance on creating helpful content.

Twelve pages that each answer something are also twelve pages you can keep correct, and correctness is what makes a page worth citing.

Which shape should the page take?

Approach	What it buys	What it costs
One page per question	A page you can keep correct	Research before writing
One page per keyword	Faster to produce	Two pages, one query
Volume on a calendar	An index that grows	Trust you cannot rebuild

In what order do you do this?

1. Collect the questions in the customer's wording, before deciding any title.
2. Group them: three to six per page, never more.
3. Write the answer as two sentences at the top, before the first heading.
4. Turn each remaining sub-question into a ## heading, phrased as asked.
5. Give each page three siblings in related, both directions.
6. Check the machine-readable claims: title, description, canonical, hreflang, sitemap entry.
7. Wait a quarter, then open the queries report and let it choose the next five pages.

Nothing in that list is a trick, which is why it survives an algorithm update.

Where this goes wrong

• The calendar wins. Pages get written to fill a week, the question list is never opened, and the corpus becomes a record of good intentions.
• The description is the first line of the body. Two pages end up sharing a summary, and the one that appears is not the one that answers.
• Two pages answer the same question. Both sit in the sitemap, neither is the canonical answer, and each makes the other look thin.
• hreflang is added once and never checked. A deleted translation leaves a language pair pointing at nothing.
• The date moves and the text does not. Refreshing a publication date without changing the content is a documented negative signal.

What this does not cover

It does not tell you which language your buyers search in — that is a market decision before it is a publishing one. It says nothing useful about earning links from other sites, which is slow work with no reliable shortcut. And it cannot predict whether a topic will collect impressions; that answer arrives after publication, which is why the procedure ends there.

We build the publishing side of this: the content model, the article templates and the frontmatter wiring that keeps two languages answering the same question.