SEO and search visibility

Search optimisation: what it is and what it is not

In short: Search optimisation is the work that makes a page findable and selectable among the results an engine has not sold. It is neither advertising, which buys a slot, nor technical compliance scored by a tool, nor a position anyone could guarantee.

In everyday use the word covers two realities that engines handle separately. Being indexed, strictly speaking, means being known to the engine: the page has been discovered, read, judged worth keeping, and it sits in the index from which answers are built. Ranking well is something else: among all indexed pages that could answer a query, this one is selected and shown ahead of the others. A page can be perfectly indexed and never appear, either because no query points at it or because other pages answer better. The distinction looks formal; it in fact governs every diagnosis, since the causes and the remedies have nothing in common. An indexing failure is an access problem, which can be observed and fixed. A ranking failure is a problem of comparative merit, which has to be worked at and cannot be repaired by a setting.

Advertising inside engines answers the same query, at the same instant, in a neighbouring slot, through an entirely different mechanism. The advertiser chooses the queries on which they wish to appear and agrees to pay when someone clicks; the slot is contested by auction and obtained as soon as the campaign is live. Search optimisation buys nothing: no sum paid to an engine places a page among its unpaid results, and the slot is not allocated but estimated, in the sense that the engine alone judges that this page serves the query better. Three practical consequences follow. Advertising switches on and off, optimisation does not: what has been built keeps producing an effect once you stop working on it, and what has not been built cannot be obtained because you suddenly need it. Advertising is steered finely — you choose the query, the message, the moment — whereas optimisation is subject to the engine's interpretation. Finally the two fail differently: a poorly targeted campaign spends without producing anything, while misdirected optimisation consumes time and leaves a site full of pages nobody asks for. Setting them against each other is an analytical mistake: they play distinct roles, one immediate and reversible, the other slow and cumulative.

Several received ideas deserve to be cleared away at once, because they misdirect decisions. Optimisation is not a registration: you do not submit a site to engines for their approval, they discover pages through the links pointing at them and through the files a site publishes for their benefit. It is not a secret or a trick: the principles are publicly documented by the engines themselves, and what is not documented belongs to shared observation rather than to a formula someone would hold. It is not a score: the tools that grade a page measure conformity to a list of recommendations, which helps spot defects, but a high grade ranks no better than a low grade demotes. Nor is it a matter of publishing volume: adding texts that answer no query dilutes a site instead of strengthening it. It is finally not an assignment that ends, since queries move, competitors publish and engines change their interpretation.

The phrase "being first" lost its precision at the same time as the results page stopped being a list. Depending on the query, an engine displays an extract that answers directly, a block of images or videos, a series of expandable questions, business listings, reviews, forum discussions, and increasingly a written answer that synthesises several sources. First place in the historical sense — the first conventional link — can sit visually far down the screen, or receive nothing at all if the answer was already given above it. Add to that the fact that results are not identical for everyone: the device, the language, the context of the query and stored preferences alter what appears, so a position is not a single value but an average of observations. The consequence for whoever decides fits into a reformulation: the useful question is not "where do I rank", but "for which queries do I appear, in what form, and does that match what I actually do".

What remains is knowing what exactly is being worked on. The relevant unit is neither the site nor the keyword, but the pair formed by an intent and the page that serves it. An intent is what a person genuinely seeks when they phrase a query: understanding a term, comparing options, finding a supplier, checking a detail before deciding, or simply returning to a brand they already know. Two different phrasings can carry the same intent and then call for a single page; two similar phrasings can carry opposite intents and call for two. That division avoids a costly mistake, which is creating one page per searched expression and ending up with a site where several pages compete for the same query without any of them prevailing. Above all it recalls what the work is for: a visit is not a result in itself, it becomes one only if the person who arrives finds what they were looking for and what they were looking for is what the business does.

Crawling, indexing, ranking, and what generative engines change

In short: An engine discovers addresses, decides which to keep, then chooses which to display for a query it has interpreted. Technique, content and reputation multiply rather than add up. Systems that compose written answers remove none of those three factors, but move the selected unit from the page to the quotable passage.

Crawling is the stage where automated programs travel the web by following links and reading the files sites publish for them. Two discovery mechanisms coexist: links, internal as well as external, which cause an unknown page to be reached from a known one, and the sitemap, a file listing the addresses the owner wants crawled. A page that no link points to and no sitemap declares exists for its direct visitors and practically not for an engine. A second file, placed at the root of the site, tells crawlers what they must not travel through; it governs access and not presence in the index, a distinction that, as we shall see below, traps people regularly. Crawling finally has a cost for the engine, and therefore an implicit limit: a site exposing thousands of interchangeable addresses — combinable filters, tracking parameters, printable versions, results of its own internal search — keeps crawlers rereading variants instead of discovering what matters. Reducing that waste improves nothing by itself; it puts the available attention back where it is useful.

Indexing is the engine's decision to keep a page and make it eligible for answers. It does not automatically follow crawling: a page can be read and then set aside because it repeats what another already says, because it is empty of substance of its own, or because the site as a whole inspires little confidence. Several mechanisms operate here and are often confused. The server response code first: an empty section that answers like a normal page instead of signalling its absence clogs the index with contentless results. The canonical address next: when the same content is reachable through several addresses, an indication placed in the page designates the authoritative one, failing which the engine decides for itself, sometimes badly. The no-index instruction last, which explicitly asks for a page to be discarded; it has to be readable, which produces the trap announced above — forbidding the crawl of a page you have also asked not to be indexed prevents the crawler from reading the instruction, and the page can remain in the index on the strength of the links pointing at it alone. One further point deserves attention: when a page only composes its content once displayed in the browser, that content is taken into account only after an additional rendering stage, which is neither immediate nor guaranteed for every crawler.

Ranking is the choice, for a given query, of the order in which indexed pages are presented. It rests on an interpretation: the engine does not look for pages containing exactly the words typed, it tries to establish what the person wants to obtain, then compares the pages able to provide it. That is why repeating an expression in a text produces no mechanical effect, and why a page can rank on phrasings it uses nowhere. The criteria are numerous, shifting, and no engine publishes their weighted list; their nature, however, is known and comes down to three questions. Does the page genuinely deal with the subject asked about, and deal with it completely enough that the person does not go looking elsewhere. Is the site a reasonable source on that subject, as attested notably by the links other sites spontaneously devote to it and by the coherence of what it publishes elsewhere. Is the page consultable without obstacle on the device in use. Inside a site, the linking between pages plays an underestimated role: it tells the engine which pages count and on which subjects, and it circulates towards deep pages the standing accumulated by entry pages.

The field is often summed up in three factors — technique, content, reputation — and the summary is accurate provided you understand that they multiply rather than add. Excellent content placed on a page no crawler reaches is worth nothing; a faultless technical architecture serving pages that say nothing is worth no more; and a site nobody cites stays invisible on contested queries whatever its own qualities. That multiplicative structure explains the order in which problems are handled: you first fix what nullifies, that is, what prevents, before improving what is merely improvable. It also explains why the three factors are not steered in the same place — technique belongs to those who build the site, content to those who know the trade, reputation to what the business does and says away from its site — and why optimisation work carried out by one person isolated from the rest of the organisation quickly stalls.

Systems that compose an answer instead of listing links modify the selection mechanism without changing its foundations. A conventional engine compares pages and leaves you to choose; a generative system composes a text and cites only a small number of sources in support of what it states. Three differences matter. The first is the unit retained: no longer the page but the passage. The system retrieves fragments that answer a precise point, and it will more readily retain a paragraph that stands on its own than a development whose meaning depends on what precedes it. The second is reformulation: a question asked in ordinary language is broken down into several searches, often phrased differently from how the person would have put it, so that targeting an exact expression loses its point in favour of covering a question and its neighbours. The third is citation without a visit: being retained as a source does not necessarily produce traffic, and a business can gain presence while watching its curves stay flat. It is that third difference which most destabilises measurement habits, because it separates two things that had always moved together.

The practical consequences show up in the way a page is written. A page meant to serve as a source answers before it develops: the statement comes first, the qualification after, the reverse of the usual rhetorical construction. Its headings state a content rather than a theme, so that a statement can be attached to a precise place. Every term used is defined where it appears, because the fragment retained travels alone and loses the context of the page. Enumerable elements — steps, conditions, categories — gain from being presented as such rather than buried in a paragraph, and frequent questions from being asked and then answered explicitly, that form matching exactly what the system looks for. Two less visible requirements complete the picture: the page must be attributable, that is, it must say clearly what it deals with and when it was updated, since a system facing two contradictory sources needs grounds to decide between them; and the content must be present in the page rather than revealed by an interaction, a tab to open or an endless scroll not necessarily being travelled through. One last point often escapes notice: the crawlers of those systems announce themselves under identifiers distinct from those of indexing crawlers, and their permissions are set separately — a site can appear in search results while being excluded from written answers, and the decision to allow them or not is taken explicitly rather than by default. What does not change deserves saying at the end: those systems rely on pages that had to be written, that a crawler must be able to read, and that there must be a reason to prefer over others. The three factors remain; only the unit of selection has moved.

How search optimisation work unfolds

In short: Establish what is actually indexed before producing anything, map intents rather than words, give each intent one page and only one, write in order to answer, then measure in sequence: indexing, appearances, clicks, enquiries.

The starting point is not production but observation. A technical audit consists in looking at a site the way a crawler sees it, which almost always gives a different picture from the one its owners hold. You first compare what exists with what is effectively indexed: pages missing from the index signal an obstacle, pages present that nobody would have wanted published signal an overflow. You then crawl the site automatically to record what blocks or wastes: addresses returning an error, redirect chains, content reachable through several addresses with no canonical indication, a sitemap declaring pages that have gone, crawl instructions inherited from a construction phase, content that appears only after execution in the browser. The result is sorted by a hierarchy that is not the one tools use: what prevents discovery comes before what hinders it, and what hinders it before what amounts to tidying. Many automated reports instead present uniform lists in which a missing tag weighs as much as an entire section made unreachable, which leads to working at length without unblocking anything.

Intent research replaces what used to be called keyword research, and the change of name is not cosmetic. You start from the questions the people concerned actually ask — those that come up on the phone, in messages received, in commercial exchanges — then observe how they are phrased in an engine, which suggestions and related queries make legible. Each phrasing is then classified by what it aims at: understanding, comparing, choosing a provider, or returning to a brand already known. A simple and often neglected test consists in looking at what the engine already displays for that query, because that display reveals the intent it has settled on: if the results are definitions, a commercial page has no place there, and conversely. You finish with an honest sort between the intents you can serve better than the pages already in place and those where the gap is too wide to close. That sort is the most useful part of the exercise and the least often carried out.

Structuring turns that map of intents into an architecture. The principle fits in one rule: one intent, one page, and one page for each intent retained. Breaking it one way produces twin pages that compete for the same query and weaken each other; breaking it the other way produces catch-all pages that answer several queries badly at once. The work then consists in making that division legible: an address that describes the subject, a main heading that states it, a hierarchy of subheadings that follows the reasoning rather than the layout, and internal links that connect neighbouring pages by meaning and not through a menu identical everywhere. A multilingual site adds a constraint of its own: each version must have its own address, be tied to its equivalents by the declarations provided for that purpose, and exist in a language genuinely written. Structured data completes the whole; it does not improve ranking by itself, it describes in a format machines read without ambiguity what the page already contains, which helps display it correctly and refer to it.

Content production is the longest part and the least delegable, because the only thing that durably distinguishes a page is what its author knows and others do not. A useful page answers the question it announces, from the start, then develops what deserves it: the particular cases, the conditions, what usually goes wrong, what is done when the standard answer does not apply. It uses the words of those searching rather than the internal vocabulary of the business, a gap that sometimes suffices to explain complete invisibility. Two reflexes are worth more than a race to publish. The first is to revisit what exists before adding: a page already indexed that gets appearances without clicks says precisely what it lacks, and improving it ordinarily produces more effect than creating a new one that will start from nothing. The second is to appoint someone responsible for the editorial side, failing which production stops at the first busy stretch, and a site whose pages contradict each other or age unwatched loses the confidence it had earned.

Monitoring only makes sense measured in the right order, because each stage conditions the next. You first check indexing: a page absent from the index can produce neither appearance nor click, and looking further is pointless. You then observe appearances, that is, the queries for which the page is offered, which indicates whether the engine has understood its subject. Then come clicks, which depend on the title and summary displayed as much as on the position. Then enquiries, the only measure that interests a business, and the only one that requires connecting what was asked of the engine with what happened next. Three precautions keep that monitoring honest. Look by group of intents rather than as one global curve, since a general rise can mask the collapse of the queries that matter. Keep a dated log of changes, without which no effect can be attached to a cause. And accept that the effect of a change is observed over cycles rather than immediately, which forbids concluding in the heat of the moment and argues against altering everything at once.

The method failures that cost the longest

In short: Targeting queries with no commercial intent, replicating the same content across languages or pages, taking traffic for a result, buying links, and handing your judgement to a tool's score.

A common trap is choosing queries by their volume rather than by what they reveal. A heavily searched expression attracts people who, for the most part, are trying to understand a term, repair something themselves or write an assignment; they will not become customers, not out of ill will, but because that was not their purpose. The symptom is recognisable: rising traffic and no incoming enquiries. The remedy is not to give up explanatory content, which has its role — becoming known to someone who will not buy today makes sense — but to avoid confusing it with the pages meant to produce enquiries, and to avoid judging one by the criteria of the other. The question to ask of every targeted query is this: is the person phrasing it in a position to choose a provider, and if not, what is the next step the page should let them take.

The second failure is replicating the same content without rethinking it, across languages or across pages. Machine-translating a site without review produces texts nobody would phrase that way in the target language, and therefore pages matching no real query: people do not search for the literal translation of what they would search for in their own language, they use their own turns of phrase, and selection turns precisely on those words. To this is added a matter of signalling: several language versions must declare themselves as such, failing which the engine may retain only one or offer the wrong one. The same failure occurs within a single language when a site multiplies near-identical pages to cover phrasing variants or per-locality declensions: they compete, none prevails, and the whole gives the impression of a site that fills rather than informs. The rule avoiding both cases is the same: a page exists only if it has something to say that no other already says.

The third failure is taking traffic for a result. A visits curve is pleasant, available and reassuring, which explains its career; it says nothing, however, about who came, what the person was looking for, or what became of them. A site can see its traffic climb sharply without a single additional enquiry arriving, and the reverse is also observed: a drop in visits accompanied by a rise in enquiries simply signals that you have stopped attracting people who had no business there. The remedy calls for some discipline: identify what genuinely counts — a written enquiry, a call, a quotation request, an appointment —, know through which page and which query the person arrived, then judge pages on that complete chain. This requirement takes on new importance with answers written by generative systems, since part of the visibility stops translating into visits: measuring traffic alone now amounts to ignoring part of what is happening.

The fourth failure is buying links. The reasoning looks unanswerable — links count, so you must get some — and it ignores what gives the signal its value. A link weighs because it is supposed to express an editorial recommendation: someone judged a page useful enough to send their own readers to it. A purchased link is not that signal, it is an imitation of it, and engines devote considerable resources to telling the two apart. The consequences run from least to most serious: the link is ignored and the spending is lost; the link is identified as part of a scheme and the site loses the benefit of other legitimate links; the site becomes the object of a penalty that will have to be lifted. There is also a rarely anticipated fragility: a rented link disappears the moment you stop paying, which builds a visibility you do not own. What works is slower and less convenient to buy: producing something that deserves citing, being present where the subject is discussed, earning mentions because you did or said something. One criterion settles the matter without hesitation: if a link would have no reason to exist without the consideration paid, it does not carry the value you think you are paying for.

The fifth failure is handing your judgement to a tool. Analysis software produces scores, lists of defects and recommendations, and it renders real service in spotting what a human would not see on a large site; it is unaware, on the other hand, of what the business sells, which queries matter to it, and what its competitors do better. Optimising a score then amounts to working on what is easy to measure rather than on what decides, and one meets sites that are impeccable by their own reports and unfindable in the results. The same analytical laziness shows in the idea that optimisation is a project one completes: queries move, competitors publish in turn, engines change their interpretation, and a position acquired without upkeep is lost without anything having been done wrong. Optimisation is better held as a regular, modest activity than as an operation: decisions taken at regular intervals, informed by the measurements, are worth more than an overhaul decided on an impression.

When search optimisation is not the right answer

In short: A site with no content and no history does not rank on technique alone; some markets are not played out in engines; some queries are out of reach; an immediate need calls for another lever; and visibility does not repair what breaks after the visit.

The first limit is structural: technique produces no ranking in the absence of content and history. A new site, correctly built, fast, perfectly crawled and indexed, remains a site nobody cites and one that says nothing yet that cannot be read elsewhere. Fixing technical defects lifts obstacles; it does not create the substance that is missing. That is why an isolated technical audit disappoints so often: it produces a list of achievable corrections and a null result, because the limiting factor was elsewhere. The honest order is this: make sure nothing prevents, then observe that the remaining work consists in writing something worth finding and in getting others to notice it. An organisation with neither the time nor the appetite to produce that content should know it before committing, because no setting substitutes for it.

The second limit lies in the market itself: not every activity is played out in engines. Some decisions are taken by tender, by listing with a principal contractor, through professional relationships or by referral, and search intervenes only at the margin, to check that a shortlisted supplier exists and looks serious. Other needs are not phrased at all: nobody searches for a solution they cannot name, and a genuinely new offer runs into zero search volume, not because it interests nobody, but because the demand has no word yet. There are finally activities where the decision is taken inside a closed space — a marketplace, a matching platform, a professional network — whose display rules have nothing to do with those of a general engine. In those situations a site remains useful as proof of existence and as material for conviction, but the visibility effort must go where the decision is taken. The test fits in a question put to recent customers: how did you find us, and how often does an engine figure in the answer.

The third limit is disproportion. Some queries are occupied by sites whose history, quantity of content and number of references exceed what a modest organisation can catch up with; attacking them head-on consumes resources without shifting anything. Recognising that situation does not mean giving up, but changing target: more precise, narrower queries expressing a particular need are at once reachable and better qualified, because whoever phrases them already knows what they are after. It must still be accepted that the total of appearances obtained will be modest. The difficulty is as much psychological as technical: a broad query flatters, a narrow one looks insignificant, and yet it is the second that produces enquiries. When even the narrow queries are saturated, the honest conclusion is that this channel will not be the main one, and it is better to admit it when deciding where to place the effort.

The fourth limit is one of tempo. Search optimisation is not steered like a switch: you do not set its level for the period ahead, and it cannot be made up under pressure. A business that must fill an order book now, launch an offer on a fixed date or offset a sudden loss of customers needs a lever that can be activated, which a paid campaign, a direct approach or a partnership allow and optimisation does not. Saying so does not devalue the underlying work: the two are run together, the immediate lever handling the urgency while the cumulative work reduces future dependence on that lever. It is the inversion that causes trouble — starting slow work in the expectation that it will settle a pressing problem, then concluding that it does not work.

The fifth limit sits after the visit. Bringing the right person to the right page serves no purpose if what they find does not answer, if the way to reach you is nowhere to be seen, if the message sent goes unanswered, or if the offer does not match what the query led them to expect. This case is frequent and easily diagnosed: normal appearances and clicks, no enquiries. The problem is then neither crawling, nor ranking, nor reputation, and devoting more effort to those makes the waste worse. One last case deserves flagging for the same reason: when a site is due to be rebuilt shortly, part of the work invested in the current pages will be lost if the architecture changes, and it is more economical to handle structure and intents during the rebuild than to come back to it afterwards. Search optimisation work is judged as much by what it declines to undertake as by what it takes on.