Best SEO practices for 2026: the ten that still decide rankings

The first SEO best practice is one page per intent. A page that tries to rank for a category, a comparison and a how-to at once is the best answer to none of them, and the title tag should name the single question the page exists to settle, in the words a searcher would use.

Intent is discovered rather than decided. Read the pages that already rank for the query and note what shape they take — a list, a definition, a comparison, a tool — because the engine has already voted on what satisfies that search. Writing a different shape because it suits you is a bet against the evidence, and it rarely pays.

The test at the end is simple. Cover the page's body and read only the title and headings. If they describe one question and its parts, the page has an intent. If they describe a subject area, it has a topic, and topics do not rank; answers do.

Answer-first structure means the question is the heading and a complete, standalone answer of roughly forty to sixty words sits directly beneath it, before any context or caveat. Google's snippets, People Also Ask and every answer engine select passages, not pages, and a passage that needs the paragraph above it to make sense is never selected.

This is the cheapest practice on the list, because it usually needs no new material. Most pages that rank without winning any answer surface already contain the answer; it is three paragraphs down, after an introduction that establishes why the question matters. Moving it up changes nothing about the substance and everything about whether it is extractable. Our piece on the answer-first rewrite covers the pattern in detail.

Repeat it per section, not once per page. Each heading is a separate opportunity to be the selected passage for a related question, and a page built this way reads as a sequence of settled points rather than an essay that eventually arrives somewhere.

Technical SEO best practice reduces to one condition: a crawler can fetch the page, see its content without running a browser, and find one canonical URL for it. Robots.txt should allow by default, content should be present in the server's HTML, and the sitemap should list only URLs that exist and return a successful response.

The mistake we see most often is not a missing feature but a contradiction between two of them: a page blocked in robots.txt that also carries a noindex tag it can no longer show, a canonical pointing at a URL that redirects, a sitemap listing pages the site removed last quarter. None of these is difficult to fix. They persist because nobody owns the interaction between the parts.

Topical depth means covering the questions a buyer actually has about one subject, from definition to comparison to failure modes to how-to, in linked pages, before starting on the next subject. Engines and models both infer expertise from coverage of a topic, and a site with one page on each of thirty topics demonstrates none.

Depth is not length. A two-thousand-word page that circles one question is padding; five linked pages of four hundred words that each settle a different question are depth. The measure is how many of the buyer's real questions the cluster answers, and the raw material for that list is sales calls and support tickets, not a keyword tool's suggestions.

The order matters commercially as well. A cluster that is complete on one service can be cited for that service today, while a site that is half-finished on ten services is cited for nothing. Finish, then move.

Off-page SEO best practice in 2026 is earning references, not acquiring links. A mention in an industry publication, a listing on a directory buyers actually use, an answer on a forum a model reads, or a citation in a comparison article all count, and the ones that also carry a link are a bonus rather than the goal.

The shift is that language models weigh what a source says about you, not only whether it links. A page that names you as a credible option for a specific situation is worth more than ten links from pages that say nothing. That changes what to ask for: a description, a placement in the right list, a corrected entry, rather than an anchor text.

Bought links, networks and mass directory submissions do not just fail to help; they are the pattern every engine's spam system is trained on, and recovering from them is slower than earning the references would have been. The comparison of citation acquisition against link building on this site sets out where each still applies.

When two pages on a site target the same intent, the best practice is to merge them into one and redirect the weaker URL, not to publish a third that tries to rank for the overlap. Near-duplicate pages split whatever authority the topic has earned, and engines then pick one arbitrarily, or neither.

This runs against the instinct that more pages mean more chances. It is the instinct behind a page per keyword variant, a page per city the business does not serve, and a page per product colour, and each produces a site that is large and thin. The consolidate-versus-publish comparison on this site argues the case; the short version is that a smaller site whose pages each own an intent outperforms a larger one whose pages share them.

The exception is genuine difference. Two pages that answer different questions about the same subject belong apart. The test is whether a reader who landed on the wrong one would be poorly served, not whether the keywords differ.

Structured data best practice is that every claim in the markup is visible on the page and true. FAQPage should carry the questions the page displays, Organization should state the facts about the business that appear elsewhere, and there should be no Review, AggregateRating or Offer node for anything the site cannot substantiate.

Schema is read as a hint about what the page says, and it is checked against the page. Markup that adds ratings the page does not show, or an address for an office that does not exist, is not a shortcut; it is the specific behaviour manual actions are issued for. The technical chapter of this series sets out which nodes that rules out, and how to validate and connect the ones that remain.

Core Web Vitals best practice is to treat page speed and layout stability as a conversion project that happens to carry a small ranking signal. Work from field data collected from real visitors rather than a single lab run, fix the template-level causes first, and expect the commercial return to show up in conversion long before rankings move.

We say the signal is weak because it is, and an agency that sells performance work as a ranking lever is selling the smaller of its two benefits. The larger one is that a page which loads and settles quickly is used, and a page that shifts under a thumb is abandoned. That is a reason to do the work regardless of what any engine rewards.

Third-party scripts, unsized media and fonts that block rendering account for most of what we find. The fix is usually a decision about what to stop loading rather than an engineering feat, which is why the conversation belongs with whoever owns the tag manager as much as with the developers.

Answer engines and generative engines such as ChatGPT, Claude, Perplexity and Google's AI Overviews select passages from pages that are crawlable, standalone and consistent with what other sources say about the organisation. Classic SEO gets a page retrieved; these additional practices get it quoted, and in 2026 both are part of the same job.

The practices are not exotic. Allow the AI crawlers by name in robots.txt. Write each passage so that it names its subject and survives being lifted out of context. State the same facts about the business everywhere it is mentioned, on your site and off it, so a model has no contradictions to resolve. Be present in the third-party sources models cite for your category.

What is different is measurement. A single assistant answer is one sample from a distribution, and screenshots of a favourable one prove nothing. The AI search chapter of this series covers the practices in depth, and our methodology piece explains how to measure whether they are working.

The last SEO best practice governs all the others: report what was measured, and nothing else. Rankings are a diagnostic, not an outcome, so tie the work to pipeline or revenue in the client's own analytics, fix the competitor set and the prompt panel before the first reading, and never publish a figure that cannot be reproduced.

This is the practice most often missing from lists of best practices, because it is the one that constrains the person writing the list. Every invented statistic, every sample case study with a plausible percentage, every review aggregate nobody collected, is a liability the moment a model summarises the page and a buyer repeats the number back. This site removed all of its own rather than relabelling them.

What changed is where answers are read: increasingly inside an AI Overview, a chat response or a Perplexity citation rather than on a results page. What did not change is what earns selection. A crawlable page with one intent, a standalone answer and corroboration elsewhere wins on the new surfaces for the same reasons it won on the old.

The practical consequence is that a site does not need a separate strategy for AI search so much as a stricter version of the one it should already have. The pages that were merely adequate for Google, ranking on the strength of a domain, are the ones that disappear when a model has to choose a passage rather than a URL.

The practices that stopped working are covered in their own piece, because knowing what to stop is half the list. Keyword density, bought links, a page per keyword variant and fabricated review markup were never best practice; they were merely tolerated, and the tolerance has ended.