Outdated SEO practices: what to stop doing in 2026, and what replaced each one

Keyword density — repeating an exact phrase a set number of times — stopped being reliable once engines began scoring pages on whether they satisfy a query rather than whether they contain it. Repetition now reads as low quality. The replacement is intent matching: answer the question the query implies, and cover the sub-questions a reader would ask next.

The meta keywords tag belongs in the same bin: Google has said publicly that it ignores the tag, and its only remaining reader is a competitor inspecting your source. Remove it, then read the page as the person who typed the query. Someone searching for a service in a city wants to know whether you operate there and how to start, not the phrase restated in every heading.

Publishing a page for every keyword variant — one for "SEO agency", one for "SEO company", one for "SEO services" — stopped working because engines resolve those to the same intent and rank one of your pages against the others. The replacement is one page per intent that covers the variants, with near-duplicates consolidated and redirected.

The on-page chapter of this series gives the test for whether two pages are competing. Our comparison of consolidating against publishing makes the full case; the rule it lands on is to consolidate when several pages chase one query and publish only when a real query has no page at all.

Word-count targets — "top-ranking pages are long, so write long" — stopped being useful because length was only ever a by-product of covering a topic properly, and padding to a number produces the preamble engines skip. The replacement is depth measured by questions: list what a reader would ask, answer each in a standalone passage, and stop.

The answer-first pattern we describe elsewhere is the practical form. Each heading is a question phrased the way people ask it; the paragraph beneath is a complete answer of roughly forty to sixty words that stands alone; the detail follows for those who want it. A page built this way is as long as its questions require, and offers more discrete passages for an engine to select.

Mass-generating thin AI copy — pages of generic prose with nothing a reader could not get elsewhere — stopped working because engines act against scaled content produced for ranking rather than people, and assistants have no reason to cite a page that restates what they know. The replacement is fewer pages carrying original judgement, method and first-hand detail.

The tool is not the problem; having nothing to say is. Generated prose over a thin dataset produces pages that differ in wording but not in substance, which our comparison of programmatic against hand-written pages identifies as the pattern engines classify as spam. What survives removing the template is usually material a team holds and has never written down: the constraint a client hit, the trade-off a method makes.

Bought links, link networks and mass directory submissions stopped being reliable because engines discount links they can identify as manufactured, and a manual action for link schemes remains a risk. The replacement is earning references from the sources engines and assistants already cite for your category: comparison posts, review platforms, trade press, community threads and original research.

Chasing authority scores belongs in the same bin: the DA and DR style numbers are a vendor's estimate of link strength, not a ranking factor engines use, useful only for discarding obvious junk. Targets come from asking your buyers' commercial questions of Google, ChatGPT, Perplexity and Claude and recording which sources the answers cite; our comparison of citation acquisition against link building explains why that list is unlike a conventional one.

Fake review aggregates, invented ratings and Review or Offer markup that the visible page does not support stopped being safe because engines compare structured data with the page and treat mismatches as spam, and a fabricated rating is a false claim in machine-readable form. The replacement is markup that mirrors what a reader can see, or none.

Google's structured data guidelines require marked-up content to be visible on the page and exclude self-serving reviews from an organisation's own markup. The rule that follows is short: emit a node only where the page shows the thing it describes, and delete an unmeasured figure rather than relabel it, least of all in a script tag a machine takes at face value. The technical chapter of this series covers which nodes that rules out.

Doorway pages — one location page per city, differing only by place name, for cities with no office — stopped working because engines treat near-duplicate sets built to funnel traffic as spam. The replacement is honest presence: a city page earns its URL with local substance, and a page for a city with no office says so plainly.

Our comparison of local SEO against programmatic city pages draws the line: local SEO where a staffed address can earn a map listing, substantive city pages where an area is served remotely. On this site a page for a remotely served city fails the build unless its copy says there is no office there, and local business markup exists only for the office we occupy, because anything else is a false location claim.

Blocking AI crawlers by default and treating llms.txt as a visibility strategy are both outdated. The block removes you from the answers your buyers read on ChatGPT, Perplexity and Google, and llms.txt grants and blocks nothing. The replacement is a deliberate robots.txt decision, allowing unless there is a commercial reason not to, with llms.txt kept as housekeeping.

A site that disallows GPTBot and PerplexityBot while maintaining a careful llms.txt has locked the door and posted directions to it. Decide robots.txt on purpose and write down why; publishers whose content is the product have a genuine case, most businesses do not. Publish llms.txt if it is cheap to maintain, expect no measurable visibility from it, and spend the effort on retrievable pages and presence in the sources models cite.

Reporting rankings alone — a table of keyword positions — stopped being sufficient because a first position beneath an answer box that resolves the query can produce fewer enquiries than a third position above nothing. The replacement is a report that ties visibility to pipeline, and measures answer-engine presence with a repeatable prompt panel rather than a screenshot.

A screenshot of an assistant naming you is one sample from a distribution nobody has measured; it demonstrates possibility, not position. Our methodology piece on measuring who gets cited describes the alternative: a fixed panel of commercial prompts drawn from how buyers speak, run cold on a schedule across named engines, scored against rules written before the first run, and paired with position, impressions and clicks for the same queries.