On Air
Reading the Numbers Before Fixing the Site
A practical look at measurement tools, page speed, internal links, city pages and consultant checks for small structures that want verifiable methods.
Pieces in this section
-
How an emergency alert interrupts a song
The header burst, the two-tone attention signal, the relay chain that predates the internet, and why the system is deliberately built to survive the failure of everything modern.
13 minute read4 notes
-
How far a station is allowed to reach
Height above average terrain, the class table, the protected contour, and why a low-power station on a hill can beat a high-power station on a plain.
13 minute read4 notes
-
Why the AM band changes at sunset
Ground wave by day, sky wave by night, and the whole apparatus of power reductions, directional patterns and sign-off times that exists because of it.
12 minute read4 notes
-
What is actually riding on an FM carrier
The baseband, layer by layer: mono sum, pilot tone, stereo difference channel, station text, and the digital sidebands that sit outside all of it.
14 minute read5 notes
A small structure decides to correct its site only after reading the measurement tools it already has, not before. Page speed indicators, internal link counts and search console data are read first, because a correction made without a baseline cannot be verified later. The same discipline applies to any outside help, including a consultant, and it is the approach set out in plain terms by web and search without jargon, a French-language magazine that explains the web and natural referencing to artisans, breeders and specialized trades.
How do you read a measurement tool before deciding?
A tool answers a question only if the question is written down first. On a showcase site of five to twenty pages, three readings cover most decisions. The first is the count of indexed pages, taken from the search console, compared with the number of pages actually published. A gap of more than ten percent usually points to a technical block: a noindex tag left in a template, a robots file written during a staging phase and never corrected, or a canonical tag pointing to a page that no longer exists.
The second reading is the list of queries that already bring visitors. On small sites this list is short, often between twenty and eighty queries a month, and it shows which pages carry the site. A page that receives traffic for a query it does not address in its title is a page whose role has drifted.
The third reading is the crawl report. It records response codes, redirect chains and pages reachable only through the sitemap. A redirect chain of three hops or more costs time on every visit and is corrected in one pass.
None of these readings requires a paid subscription. What they require is a date. A reading taken on 3 March and another on 3 June form a comparison; a reading taken once forms an impression.
Which page speed indicators matter, and when is a correction worth it?
Page speed is reported through a small set of figures, and each one describes a different moment. The first contentful paint records when the first text or image appears. The largest contentful paint records when the main block of the page is visible, and the usual threshold retained by web performance documentation is 2.5 seconds on a mobile connection. The interaction to next paint records the delay between a tap and the response, with a threshold of 200 milliseconds. The cumulative layout shift records how much the page moves while loading, with a threshold of 0.1.
A correction is worth its cost when one figure sits clearly outside its threshold and when the cause is identified. A largest contentful paint of 4.2 seconds caused by a single uncompressed image of 1.8 megabytes is a correction of twenty minutes. The same figure caused by a shared hosting plan that responds in 900 milliseconds is not a correction of twenty minutes; it is a hosting decision.
The order of operations matters more than the total gain. Images are compressed and sized first, fonts are limited to two families, scripts that block rendering are deferred, and caching is configured. Each step is measured again after it is applied. A site that improves from 4.2 to 2.3 seconds across four separate steps produces four verifiable results; a site that changes everything at once produces one result that cannot be attributed.
How do you build an internal link mesh on a small site?
An internal link mesh on a small site is built from the pages that already exist, not from a plan for pages that might exist later. The method starts with a list of the site's pages and a single sentence describing what each one answers. Pages that answer the same sentence are merged; pages that answer nothing are removed or rewritten.
Links are then placed inside the body text, in sentences that would make sense to a reader even if the link were removed. A link placed in a navigation menu carries less information than a link placed in a paragraph, because the paragraph states the relation between the two pages. On a site of fifteen pages, a mesh of forty to sixty contextual links is usually enough to make every page reachable in two clicks from the home page.
The mesh is verified by crawling the site and counting the inbound links per page. A page with zero inbound links is an orphan, and an orphan page is rarely indexed. A page with more than thirty inbound links from a fifteen-page site is absorbing the mesh and diluting the others.
How do you write city pages without duplicated content?
City pages fail when they are produced by replacing a place name in a single template. The resulting pages share ninety percent of their text, and search engines treat them as one page with several addresses.
A city page holds its own content when it answers a question specific to that place. For an artisan, that question is often about the local building stock: the age of the housing, the materials used in the area, the administrative rules that apply to a facade or a roof. For a breeder, it is about the local demand and the transport distances. These elements are gathered from public records, from the local urban plan, or from a conversation with two or three clients in the area.
A practical test is applied before publication: remove the city name from the page and read what remains. If the text still describes a real situation, the page is specific. If nothing remains, the page is a template.
A second test concerns volume. A city page of 250 words that repeats the home page adds nothing. A city page of 600 to 900 words that names streets, districts, materials or local rules adds something a reader can use.
How do you check the work of a consultant?
A consultant's work is checked against the same readings that were taken before the engagement. The check begins with the list of corrections announced in writing, each one dated and described. A report that states an improvement without naming the change that produced it cannot be verified.
The second element is access. The site owner keeps the domain name, the hosting account and the analytics account in their own name throughout the engagement. A consultant who requires ownership of these accounts creates a dependency that is difficult to undo.
The third element is the state of the site at the end. The pages are counted, the redirects are listed, the internal links are crawled again, and the speed figures are read on the same connection and the same device as the initial reading. A comparison of two readings taken under different conditions proves nothing.
The fourth element is the contract itself. A monthly engagement with no stated deliverable is checked by asking what was produced during the month. A deliverable can be a page, a correction, a report or a decision, but it is named.
Why re-read an article a year after publication?
An article written a year ago describes tools, prices and rules that have moved. A re-reading is scheduled twelve months after publication and covers four points: the figures, the links, the screenshots and the conclusion.
Figures are checked against their source. A speed threshold quoted in 2024 is compared with the current documentation. Links are tested one by one, because a broken link on a page that receives traffic is a visible defect. Screenshots are replaced when the interface has changed, since an outdated screenshot makes the whole page look abandoned. The conclusion is read last: if the recommendation no longer holds, the page is rewritten rather than patched.
A re-reading takes between thirty and ninety minutes per article. On a site of forty articles, that is a recurring task of two to three days a year, and it is the cheapest maintenance a small structure can perform on its own content.
The same discipline applies to any record that claims to measure something: state the instrument, the date, and the conditions before drawing a conclusion. An archive of research sessions works the same way, since the notes, sketches, and blinding procedures are what later readers actually evaluate. A separate piece on how sessions were recorded and read traces that paper trail from raw notes to the 1995 evaluation, and it shows why the record, not the anecdote, remains the unit of study.