Skip to main content
These are starting defaults, not laws. Fill in the right-hand column with what your team agrees, and edit any row that does not fit how you publish. Each section has a concept review dropdown explaining what the standard is for.
Numbers on this page come from two places: published research on how content structure affects citation in AI answers, and ordinary publishing convention. Where research exists it is named in the concept review. Where it does not, the number is a convention worth having, not a finding.
Greyed text marked E.g. is an example. Examples use Doughnut Labs, a company that sells Doughnut Technology.

​
Heading hierarchy

A heading structure serves a reader scanning for the part they came for, a screen reader announcing document structure, and a retrieval system deciding where one passage stops and the next starts. All three read the same markup, which is why a hierarchy chosen for visual reasons tends to break something invisible.Two H1s on a page is the most common version of this. It leaves both the page and any system parsing it without a single clear subject.On depth, the structural citation research found three to five levels of hierarchy outperformed both flatter and deeper structures, with too little offering no organizational signal and too much diluting attention across structural markers. In practice most articles land at H1, H2, and some H3, and a section reaching for H4 is often a section that wants to be its own page.

​
Article length

Length follows the question. These are the ranges a type of article usually needs to answer its question properly, not word counts to hit.
There is no length that ranks better, and the correlation people cite between long articles and good performance mostly reflects that thorough treatments of complex questions tend to be long, not that length causes anything.What length is useful for is agreeing expectations before someone writes. A brief that says 1,000 words and a draft that arrives at 4,000 represent a disagreement about scope that nobody had out loud, and it is far cheaper to have that conversation at brief stage.The failure mode in both directions is worth naming. Padding a thin answer to reach a number produces exactly the general, unsourced prose that neither readers nor answer engines have any use for. Cutting a genuinely complex answer to fit a range produces something that answers the question incompletely, which is worse. A range is a signal about expected depth, and an article that has a good reason to sit outside it should sit outside it.

​
Section and structure ratios

These come from research on how content structure, independent of what the content says, affects whether a generative engine uses a passage. Sections of roughly 150 to 300 words performed best, with longer sections showing measurable attention loss across their middle and much shorter ones splitting a single idea across too many fragments. Lists and tables were extracted substantially more accurately than the same information in prose, with content around a quarter to a third structured outperforming either extreme.The reason the high end has a limit is editorial rather than technical. An article that is nearly all bullets has usually stopped arguing, because the connective reasoning between points is the first thing lost when everything becomes a list. Prose carries reasoning, lists carry parallel items, and the ratio takes care of itself when each is used for what it does.Treat all four numbers as direction. They come from a young field studying systems that change frequently, and an article written to satisfy a ratio rather than a reader will read like one.

​
Slugs

A slug is read by people deciding whether to click and by systems working out what a page is about, and it is the part of a page that is most awkward to change later. Every inbound link, citation, and bookmark points at it.The date rule is the one that costs teams most often. A slug containing a year commits you to either publishing a new URL every year, which splits the authority the old one accumulated, or maintaining a page whose URL contradicts its content. An article meant to be updated indefinitely is easier to run from a URL that never had a date in it.

​
Metadata limits

These are display limits, not rules: search engines truncate at a pixel width, not a character count, and they rewrite descriptions when they judge a different snippet more relevant. Writing to the limit means the sentence you wrote is the sentence most people see.The meta description is worth more attention than its lack of ranking influence suggests. It is the pitch that decides a click, and answer engines sometimes use it as a ready-made summary of the page. Writing it after the article is finished, rather than as a placeholder at the start, tends to produce a real summary instead of a guess.

​
Internal linking

Internal links do the structural work that makes a set of separate articles read as coverage of a subject rather than a pile of pages. They also route a reader onward, and they distribute whatever authority the site has accumulated.The inbound rule is the one most often skipped, because it is work on pages other than the one being published. An article with nothing pointing at it is isolated regardless of its quality, and the pages that should point at it are usually obvious at brief stage and forgotten by launch day.Descriptive anchor text matters for a plainer reason than most link advice: it is what a screen reader announces when a user tabs through the links on a page, and a list of “click here” tells them nothing.

​
Images and media

Alt text exists first for people using screen readers, and the fact that it is also machine-readable text about the image is a secondary benefit rather than its purpose. Written as a description of what the image shows, it serves both. Written as a list of keywords, it serves neither and makes the page worse for the people it was meant for.Stating a chart’s key numbers in the body text as well is the practical version of the same idea. A finding that exists only inside an image cannot be read by a screen reader, quoted by an answer engine, or found by anyone searching for the number.

​
Schema and structured data

Schema is a machine-readable statement of what a page is and what it contains, sitting alongside the human-readable version. It removes guesswork from systems trying to work out whether a block of text is a step, a question, an author, or a date.The Organization markup with sameAs links is the piece most directly connected to entity authority. It states, in a form a machine can act on, that the company on this site is the same company as the one on your other profiles, which is what lets separate mentions accumulate into one recognized entity instead of several unconnected ones.Schema describes what is on the page. Markup claiming a review, an author, or an FAQ that does not exist in the visible content is a misrepresentation that search engines act against, so the safe version is to mark up what is there.

​
Dates and freshness

Freshness has become one of the clearer signals in how AI answers select sources. Published analyses of citations on commercial queries put the large majority on pages updated within the past year and well over half on pages updated within six months. The figures vary by who is measuring, but the direction is consistent enough to plan around.The distinction between an update and a date bump is what keeps the practice honest. Changing a timestamp without changing anything else is visible to readers who have seen the page before, and it removes the one signal that would otherwise tell your own team which articles genuinely need attention.Cadence is where teams reasonably differ. A quarterly cycle on a large hero set is real recurring work, and a team that can only sustain twice-yearly reviews is better served by an honest twice-yearly cadence than by a quarterly one it silently misses. The article maintenance process covers how the cycle runs.
Last modified on August 6, 2026