AI visibility vs SEO: what transfers and what does not

A page can sit at position one for a commercial query, hold that position for two years, and be absent from every answer four different assistants give to the same question in plain language. That is not a contradiction or a bug. The two systems are being asked different things and they produce different shapes of output.

Plenty of people have responded to this by declaring search engine optimisation finished, which is both wrong and easy to spot as hype. The useful version of the question is narrower. Which parts of the work you have already done carry over, and which parts have to be done differently because the result itself has changed shape.

What changed about the result

A search engine hands back a list. Ten results, plus ads, plus whatever panels are in fashion. The list is ranked, the ranking is observable, and every entry is a link you can be clicked through.

An assistant hands back a paragraph. Inside it are two or three brand names, sometimes five, occasionally one. There is no list to be tenth on. You are either named or you are not, and if you are named, what matters is whether you came first or arrived in a clause near the end after the options the model actually preferred.

Three consequences follow from that single structural difference, and most of the confusion about AI visibility comes from not having worked them through.

Position is inside a sentence, not on a page

There is no rank to track, because there are no slots. What exists is order of mention and prominence within an answer. Being named first, as the primary recommendation, is a materially different outcome from being named fourth after three competitors have already been described. This is why serious measurement weights position rather than counting mentions: a brand mentioned in every answer but always last is in a much weaker place than the numbers alone suggest.

There is often no click to count

Somebody asks an assistant which tool to use, gets three names, recognises one, and goes directly to that brand's site by typing the name. In your analytics that is direct traffic or a branded search. The assistant that put you on the shortlist is invisible in the attribution. You cannot measure this channel by waiting for it to appear in your referral report, because most of the time it never will. You measure it by asking the questions yourself and reading the answers, which is a different discipline from reading logs.

The competitive set is decided inside one answer

On a results page, you and your competitors coexist. Ten links, ten brands, the user chooses. In an answer, naming you often means not naming someone else. The question is no longer whether you can be found. It is whether you make the cut in a set of three.

What transfers from SEO

More than the "SEO is dead" crowd will admit.

Being crawlable and rendering server side still matters, because retrieval reads pages the same way crawlers do. Content that only exists after client side JavaScript executes is content some retrievers will never see.

Topical authority transfers in substance if not in mechanism. The work of becoming the site that thoroughly covers a subject is the same work that gets you quoted, because the pages that models retrieve are, unsurprisingly, often the pages that rank.

Off site reputation transfers strongly. The link building instinct, in its legitimate form of earning mentions on other people's sites, maps almost directly onto what changes an assistant's answer, because third party pages are what assistants summarise. The difference is that for AI visibility an unlinked mention on a review site or in a roundup still counts, since the model reads the sentence rather than following the hyperlink.

Clear structure transfers. Headings that say what the section is about, plain answers stated near the top, specific facts written as sentences rather than implied by a graphic. All of this was already good practice for featured snippets and it is straightforwardly good for extraction.

What does not transfer

Keyword targeting in the traditional sense loses much of its grip. People type fragments into a search box and whole sentences into an assistant, and the assistant then rewrites that sentence into its own internal queries before retrieving anything. Optimising a page for an exact phrase is optimising for a step that no longer happens in the same form.

Ranking for your own brand name buys you very little. In search, owning your branded queries is a real asset. In an assistant, the user rarely types your name at all. They describe a problem, and you are either part of the answer or you are not. Brand defence and category presence are separate problems here.

Your own pages carry far less weight relative to everyone else's. Every vendor site in a category makes the same claim in the same register, so as a body of evidence it separates nobody. Comparison articles, review platform category pages, directories and community threads do the separating instead.

Freshness works on different clocks per model. Pages that get retrieved can change an answer within weeks. What a model absorbed in training only changes when a new model ships. One tactic does not cover both.

And the feedback loop is slower and noisier. An assistant asked the same question twice can answer differently, so a single check is a reading rather than a fact. You are looking at direction over weeks, not at a number that moves the day you publish.

Measuring a channel with no traffic in it

This is the part that tends to stall teams. There is no console to open. The only reliable method is to ask the questions your buyers ask, on the models they use, in the language they use, and record what comes back: whether you were named, how early, which competitors were named instead, and which sources the answer leaned on.

Done by hand that is an afternoon per round, which is fine for a baseline and impractical as a habit. Spegla exists to run exactly that loop, and the free check needs no account, but the method matters more than the tool. If you take one thing from this, take the shape of the measurement rather than the vendor.

Two numbers are worth separating when you do it. A plain mention rate tells you how often you appear at all. A position weighted score tells you how often you appear in a way that would actually influence the decision. A brand that is always mentioned last has a healthy looking mention rate and a weak score, and the gap between the two is the most actionable thing on the page.

What to do next

Keep doing your SEO. It buys you the pages that get retrieved and the authority that gets you quoted, and none of that has stopped working.

Then add the part it does not cover. Write down five questions a buyer would actually type, with no brand names in them. Ask them on more than one assistant. Note who gets named and which sources the answers cite. Then go and work on those sources, because they are the pages deciding your category, and they are almost never your own.