State of the Art SEO: What Actually Matters in 2026

There are two ways to talk about SEO in 2026. One declares the discipline dead because answer engines intercept the clicks. The other sells the same checklists as in 2019, only with “AI” in the title. Both are convenient and both are misleading.

What has actually happened: the channel has split. In front of the classic results list sits a second surface where machines assemble answers and cite sources. The craft underneath – be findable, be unambiguous, be verifiable – has not changed. Master it and you serve both surfaces with the same work. Fail at it and you lose on both.

This article sorts out what genuinely matters today. Throughout, it separates what is established from what is plausible and from what is merely being sold.

Key takeaways

• Search is splitting into two surfaces: the results list and the answer engine. The technical and editorial foundation is identical for both.
• The most expensive mistakes are almost never editorial but structural: indexability, canonicals, duplicate titles, broken internal linking.
• Core Web Vitals are a weak ranking lever but a strong conversion lever. Anyone selling it differently has not read the sources.
• With structured data, machine readability now matters more than the rich result – with one clear exception: merchant fields for shops.
• GEO is not a second discipline but the question of whether your facts are available in quotable form. Much about it is plausible, little of it proven.
• Prioritise by effort × impact, not by novelty. The boring items pay out most reliably.

What has changed – and what has not

The most honest description of the situation: the number of searches has not fallen, the number of clicks per search has. When Google displays a summarised answer at the top, some users no longer need the results list. This does not hit all content equally.

Vulnerable is content whose entire value is a single piece of information: definitions, conversions, opening hours, “what does X mean?”. Those clicks are not coming back, and no trick brings them back. Robust is content that supports a decision involving money, effort or risk. Nobody budgeting 3,000 euros for a shop system or two months for a migration reads a summary and stops – they want the reasoning, the counter-calculation, the experience behind it.

From this follows the first strategic decision, and it is more uncomfortable than any technical one: if your content portfolio consists mostly of answerable questions, your problem is not SEO but the portfolio.

What has not changed: both surfaces draw on the same index. Even a generated answer needs a source that was crawled, understood and judged trustworthy. There is no side entrance that skips the foundation.

The foundation: findable, unambiguous, fast

Technical SEO has a poor reputation because it is unspectacular. It is still where most visibility is lost – not gradually, but in a single deploy.

Indexability is level zero. A noindex from the test environment, a robots.txt carrying staging behaviour, a canonical pointing at the development domain: such mistakes do not cost positions, they cost the entire page. They often survive for weeks because nobody looks – in day-to-day work people check rankings, not indexability.

Unambiguity is the second level and the most expensive one in e-commerce. Filters, sorting, pagination and tracking parameters quickly turn one category page into dozens of variants. Without a clean canonical they compete with each other, and Google decides for itself which variant it considers the right one – not always the one you meant. The same family includes the classic that survives on a surprising number of shops: www and apex domain both serving content instead of one 301-redirecting to the other. Two hosts, two pools of signals, half the effect.

Just as common and just as silent: duplicate titles. Category pages that differ only by a filter yet carry the same title. Pagination pages without a page number. In the results list those pages are indistinguishable – for Google as much as for people.

Internal linking is the most underrated lever because it costs nothing but thought. It is the only ranking you assign yourself: whatever gets linked often and from relevant context is visibly what you consider important. Most shops link at random – through navigation and footer, but rarely from body copy. Link deliberately from strong pages to the pages that are supposed to earn money and you shift weight without a single backlink.

Speed deserves a clear assessment, because a great deal of nonsense is sold here. Page experience is one signal among many at Google – it can tip the balance between otherwise comparable pages, but it carries no weak content upward. Anyone selling load-time work as a ranking miracle has not read the sources.

The real reason to take it seriously is commercial: load time acts directly on bounces, cart abandonment and conversion. The lever is real, it simply is not where it is advertised – covered in detail in our article on page performance as a double lever. One thing matters there: field data from real users counts, not the lab score of a test run on your development machine.

Content that is an answer

The content side has changed less than the debate suggests. What has disappeared are the levers that never were any: keyword density, word count as an end in itself, synonym sprinkling.

What matters can be said in four points.

Search intent before search term. The same phrase can carry three different intentions. Someone typing “Shopware vs. Shopify” wants a decision aid, not a feature table and not a product page. The fastest check is the most honest one: look at what actually ranks for your target query. If the page one is all comparisons, you will not win with a sales page.

Answer first, reasoning second. The question belongs in the heading, the answer in the first two sentences, the derivation below. That helps impatient readers, it helps snippets, and it is simultaneously the structure answer engines can quote from cleanly. One format, three benefits.

Entities instead of word lists. Search engines work with relationships between things, not with character strings. In practice that means: name things consistently, link thematically related content to each other, and make your brand tangible as an entity – with an organisation markup, a logo and sameAs references to the profiles that genuinely belong to you. It is tedious and invisible. It is also why some brands with ambiguous names are recognised reliably and others are not.

E-E-A-T is craft, not a trust bonus. Experience, expertise, authoritativeness and trustworthiness are not a score you optimise. In practice it means: a real author with a name, a profile and a traceable background instead of “the editorial team”; verifiable statements instead of assertions; an imprint people can find. For topics involving money or health this is not polish, it is the entry ticket.

And a word on freshness: a changed date is not an update. What counts is whether the content still holds. Date cosmetics without editorial work becomes obvious the moment somebody reads the text.

Links: the part nobody likes to be honest about

No topic has burned more budget than backlinks, and none is discussed soberly less often. Three observations that hold up in practice.

Links still work, but not linearly. A reference from a site that holds authority in your field moves something. Fifty references from directories, guest-post networks and comment sections move nothing – at best. The number in the tool rises anyway, and part of the industry lives off exactly that.

Bought links are a risk with compound interest. They work sometimes, for a while. The problem is not detection tomorrow but that you are building a foundation you do not own: when the effect disappears or is devalued, the part of your visibility you paid most for collapses. For a shop meant to be run long-term, that is the wrong bet.

The realistic route is less comfortable and cheaper. It consists of things you should be doing anyway: publishing your own data that someone can cite; building tools someone links because they are useful; writing professional pieces that take a position instead of dodging it. That is slow. It is also the only link building that cannot be undone.

And the most important qualification: for most shops, what stands between today and clearly better visibility is not a link problem but a structure problem. Anyone with duplicate titles across forty category pages who complains about too few backlinks is optimising in the wrong place – and in the expensive one.

Structured data – the honest balance

Few areas are recommended so broadly and assessed so rarely. The state of play after several rollbacks by Google:

What was rolled back. FAQ rich results have been essentially reserved for authoritative government and health sites since 2023. The sitelinks search box was discontinued. Anyone adding FAQ markup to get a snippet is adding it for a reason that no longer exists.

What still delivers. For shops the merchant fields are the most concrete lever in the entire field: shipping cost, delivery time, return window and product identifiers such as GTIN or MPN appear directly in the result. Leave them out and the space stays empty while a competitor fills it. That is not a ranking question but a click-through one – and it is settled with one set of fields in the offer object. Equally reliable: breadcrumbs, which replace the bare URL in the result with a readable path.

What has shifted. The second purpose of structured data now matters more than the first: it is the machine-readable version of your facts. Price, availability, author, date, affiliation – in JSON-LD they are stated unambiguously instead of being inferred from prose. That is exactly what any machine summarising your page needs.

Which is why FAQ markup remains sensible – just with a corrected expectation: not for the snippet, but for quotability.

One practical note to close this chapter: validity is the lowest bar. Google’s testing tool says “valid” and stays silent about half the useful fields being absent. That gap is exactly what our Structured Data Check covers.

GEO: visibility in generated answers

Generative engine optimization is the term for the question of whether your content appears in a machine-generated answer. The most useful contribution here is a clean separation – because in few areas is so much currently being claimed.

What GEO shares with SEO. Almost everything. An answer engine can only quote what it finds, understands and considers reliable. Crawlability, clear structure, unambiguous entities, verifiable facts: identical to the list above. Anyone who believes GEO is a separate discipline with its own tricks has usually done neither properly.

Where it genuinely differs. In three points:

• The unit is the passage, not the page. What gets quoted is a paragraph that stands on its own. Text whose point only emerges across five paragraphs is harder to use than text that delivers its core statement in two sentences.
• There is no position you can look up. Answers vary by phrasing, user and moment. “Position 3” does not exist.
• Success is often click-free. A mention without a visit can still be valuable – just not in the analytics column you normally check.

The honest three-way split. What can reasonably be treated as established today: what gets quoted is preferentially what already ranks well, answers a question directly and delivers its facts extractably. Plausible but unproven: that machine readability in the narrow sense – markdown delivery, MCP endpoints, clean file twins – helps additionally. The effort is small, the risk zero, the benefit unmeasured. And simply unproven: that an llms.txt increases visibility. Google states explicitly that it is not needed for its own AI surfaces. It costs nothing, it harms nothing, it replaces nothing.

Anyone who does not keep these three categories apart eventually buys a “GEO audit” that is an opinion with a score attached.

The concrete implementation for shops – product feed, freshness signals, measuring mentions – is covered in detail in our article on LLM optimisation for stores. If agents are supposed to buy rather than merely quote, that is another field again: see agent-ready commerce.

Measure instead of believe

SEO rarely fails for lack of data, often for misreading it.

Search Console is the only source Google itself provides – and the most misunderstood. Three things pay off. First, taking the indexing report more seriously than the performance report; “alternate page with proper canonical tag” and “crawled – currently not indexed” say more about your structure than any position curve. Second, thinking in query groups rather than individual terms: the position of a single keyword is noise, the trajectory of a topic cluster is a signal. Third, comparing periods of equal length and accounting for seasonality before diagnosing a crash.

Field data beats lab scores. A performance test run shows what would be possible under ideal conditions. What counts are the values of real users on real devices and real networks. They are sluggish – changes take weeks to show – and precisely for that reason dependable.

And then what cannot be measured. How often you appear in AI answers, how often a brand is mentioned without anyone clicking, what share of your direct traffic stems from a recommendation: there are tools with samples for this, but no full census. Anyone promising precision here is selling a feeling. The honest way to handle it is to say so and still check by sampling – ask the same twenty buying questions regularly and record whether you show up.

The order in which it pays

Most projects fail not from missing knowledge but from the wrong order. As a rough prioritisation:

Measure	Effort	Impact	When
Check indexability and canonicals	low	very high	now, then quarterly
Eliminate duplicate titles and descriptions	low	high	now
Set internal links deliberately	medium	high	ongoing
Add merchant fields to Product/Offer	low	high (shops)	now
Answer-first structure in core articles	medium	high	top 20 first
Improve load time based on field data	high	medium for SEO, high for CR	after the quick wins
Mark up entities and authors properly	medium	medium, long-term	once, then maintain
Machine readability for AI surfaces	low	unproven	once the foundation holds
Earn references through your own content	high	medium to high	permanently, never as a campaign

The order is deliberately unglamorous. The biggest mistake in practice is not doing the wrong thing – it is starting with the newest thing while a noindex from the last relaunch keeps half a category invisible.

The bottom line

State of the art in 2026 does not mean everything is new. It means two surfaces are served from the same substance: a results list that ranks pages, and an answer engine that quotes passages. Both reward the same things – findability, unambiguity, verifiable facts, clear structure.

The difference between good and bad work today lies less in tactics than in honesty: knowing what is established, what is plausible and what is merely claimed – and allocating your time accordingly. Keep the foundation clean and you win in both channels. Leave it be and you lose in both, however modern the measure you start with instead.