Where AI helps in the content process
A content process runs from strategy to measurement, and AI is useful at some stages, marginal at others and a liability at a few. The table takes the stages in order. "Fit" is a judgement about the kind of task, not about any tool; "what to watch" is the failure that review has to catch at that stage.
| Stage | Where AI fits | What stays with the team | What to watch |
|---|---|---|---|
| Strategy and goals | Marginal; can summarise research the team supplies | Audience, positioning, the content goals and the KPIA metric tied directly to an objective. that measures them | Strategy written in generalities that fit any company |
| Research | Good for mapping a topic's questions and sub-topics | Primary research, interviews, the team's own experience | Invented sources; stale information presented as current |
| Outlining | Strong; outlines are quick to judge | The angle and the single question the piece answers | Structures that answer a broader question than the brief |
| Drafting | Good with supplied sources; weak without | Facts, examples, the argument, calls to action | Invented facts, quotes and figures; smooth generic prose |
| Editing | Strong for length, level, format and consistency | Voice, judgement about what to cut | Loss of specific detail; flattening of the author's voice |
| Repurposing | Strong when the source is a reviewed piece | Which formats suit which audience | New claims introduced in the derivative; drift from the source |
| Localisation | Good as a first pass | Market knowledge, legal and cultural adaptation | Literal translation; wrong register; references that do not travel |
| Publishing and measurement | Marginal; metadata suggestions and summaries | Decisions about what to publish and what the numbers mean | Descriptions that promise what the page does not deliver |
The pattern across the rows is that AI is strongest where the input is specific and the output is easy to compare with it, and weakest where it is asked to know things or to decide things. A content programme built on that pattern hands the tool outlines, rewrites and adaptations, and never hands it the facts.
Research and outlining
Research with AI is most useful as a way of mapping a subject rather than learning it. Asked for the questions a particular audience has about a topic, the sub-topics a thorough treatment would cover, or the angles that competing pages take, a tool produces a map that would take a person an hour or two to assemble from a results page and a few forums. The map is a starting point. Each item on it is a hypothesis about what readers want, and the search intent guide describes the check: read the results page for the target query and see what actually ranks.
What research with AI cannot do is supply facts the team can rely on. Tools present invented references with the same confidence as real ones, describe studies that do not exist, and report figures that are plausible and wrong. Anything the research turns up that will be stated in the piece has to be traced to a source the writer has opened. A workable rule is that AI research produces questions and leads, and people produce citations.
Outlining is the strongest single application. Given a brief with the audience, the question the piece answers, the angle and the required format, a tool returns an outline in seconds, and an outline is cheap to judge: either it follows the brief or it does not. The common failure is an outline that answers the general question in the title rather than the specific question in the brief, with the same sections any page on the subject would have. The fix is in the brief, not the prompt: a brief that states the angle and names what the piece will not cover produces an outline worth keeping. The content-strategy framework this guide pairs with puts the brief at the centre for exactly this reason.
Drafting from briefs
Drafting is where the saving looks largest and is most often illusory. A tool will produce a full draft from a one-line instruction, and the draft will read fluently. Fluency is the problem. A draft that reads well is reviewed less carefully than one that reads badly, and the errors in a fluent draft are the kind that survive a skim: a feature the product does not have, a figure that sounds right, a quote attributed to someone who never said it, a regulation described as it was before it changed.
The difference between a draft worth reviewing and one not worth the review time is what goes in. A draft from a detailed outline, with the relevant source passages pasted in and an instruction to use only those and to mark anything that goes beyond them, is checkable: each claim can be matched to a passage. Tools that assemble those passages automatically from a document library work the same way, as how grounding and retrieval work explains, along with the points where they still miss. A draft from a topic and a word count is not checkable except by researching the subject from scratch. The AI-assisted content with review guide covers how to turn a brief into a prompt and how to run the claim check; the point here is that drafting is only an application worth having when the inputs make the output verifiable. Drafting at catalogue scale from structured product attributes is a special case with its own risks, covered in the guide to generating product descriptions from catalogue data.
Two kinds of content should not be drafted by a tool at all, whatever the inputs. The first is anything whose value is the author's own experience: a founder's account of a decision, a practitioner's method, a review of something the team used. A tool can only produce a generic version of experience it does not have, and readers and search engines both notice. The second is anything where the organisation would be liable for an error: pricing, legal terms, medical or financial claims, product specifications. Those are written by people and checked by people, and AI's role is at most to tidy the prose afterwards.
Editing, repurposing and localisation
Editing is the application with the best ratio of value to risk, because the source is already true and the tool is only changing its form. Cutting a piece to a target length, raising or lowering the reading level, converting prose to a list or a table, standardising spelling and terminology, and checking a piece against a style guide are all tasks where the output is compared directly with the input. The failure mode is loss: a rewrite that drops the specific detail that made a paragraph useful, or smooths a distinctive voice into the same even tone every tool produces. Editing with AI works best as a suggestion a person accepts or rejects paragraph by paragraph, not as a wholesale replacement. Keeping drafts in the brand's own register is easier with a voice pack built for the tool, which the brand voice guide for AI drafting walks through step by step.
Repurposing takes a reviewed piece and produces derivatives: a summary for email, a thread for social, a script for a short video, a slide outline, a set of excerpts for a newsletter. Because the source has already passed review, the derivatives inherit its accuracy, provided the tool does not add anything. That proviso is the whole review: read each derivative against the source and strike any claim, figure or example that is not in it. The repurposing content guide covers which formats suit which audiences and how to plan derivatives from the start; AI makes the production step fast and leaves the planning where it was.
Localisation is a first pass, and the quality of the first pass depends on the language pair, the subject and how idiomatic the source is. A fluent reviewer who knows the target market is not optional. Beyond correcting the translation, the reviewer adapts: examples and references that mean nothing locally, units and formats, regulatory statements that apply in one market and not another, tone that reads as rude or stiff when carried across literally. The brief for a localisation should say which passages must stay exact, such as legal wording and product names, and which the reviewer may rewrite freely. A termbase sent with every job keeps those names and terms fixed, and the tutorial on building a termbase and translation memory sets one up with the in-country reviewers.
One guide, five derivatives, one afternoon
A B2B software team publishes a 2,400-word guide on procurement approval workflows after full review. The content lead asks a tool for five derivatives: a 150-word email summary, a 6-post social thread, a 90-second video script, a 10-slide outline and 3 pull quotes. Production takes under 20 minutes. Review takes about 2 hours: the thread has introduced a statistic about approval times that is not in the guide and is struck; the video script has invented a customer name for the example, which is replaced with the guide's anonymised version; the slide outline is fine; the email summary promises a template the guide does not contain and is rewritten. Four of the five derivatives ship that afternoon. The team's log records that review took roughly six times as long as production, which it treats as normal for derivatives rather than as a failure.
Where quality fails
Quality fails in a few recognisable ways, and naming them makes review faster.
Invention. Facts, figures, quotes, sources, product features and examples that the tool produced because the prompt asked for a piece and the piece needed them. This is the most dangerous failure because invented material is written with the same confidence as the rest. It is caught only by checking claims against sources. When inventions keep reaching drafts or published pages, the diagnosis for AI-written content that contains invented facts traces the pattern to the sources, the prompt, the review step, the retrieval library or an untested change.
Staleness. Information that was true when the model's training data was gathered and is not true now: prices, policies, features, people in roles, the state of a regulation. Any piece that touches current facts needs its currency checked by a person who knows the field.
Genericness. Prose that could sit on any competitor's site, because the prompt gave the tool nothing that only this organisation knows. It is not an error in the sense that invention is, but it is the failure search engines and readers punish most reliably, since the page adds nothing to what already ranks. The E-E-A-T guide explains what first-hand experience and demonstrable expertise look like on a page; generic drafts have neither. The cure for the prompt side of the problem is a fuller brief, and prompt patterns for marketing sets out what any prompt has to carry.
Drift. A derivative, a rewrite or a translation that has moved away from its source in substance while keeping its shape. Caught only by comparing to the source.
Over-promising. Titles, descriptions and introductions that promise what the piece does not deliver, because tools optimise for appeal. Harmful for search click-through and for trust.
Sameness at scale. A programme that publishes many AI-assisted pieces from similar prompts produces a body of content with one voice and one structure, which readers notice before any single piece looks wrong. The remedy is variety in briefs and authorship, not in prompts.
The accuracy and review workflows guide turns these failure modes into a review procedure with tiers by risk; this guide's job is to make them recognisable.
Search guidance on AI content
Google's published guidance on AI-generated content, cited in this page's sources, is short and worth reading in the original. Its position is that Google's ranking systems aim to reward high-quality content however it is produced, and that automation has long been used legitimately for things like sports scores and weather reports. It states that using automation, including AI, to generate content with the primary purpose of manipulating search rankings is a violation of its spam policies. It points to the qualities its systems look for, summarised as E-E-A-TExperience, Expertise, Authoritativeness, Trust., and suggests that publishers consider who created the content, how it was created and why. On disclosure, it says that AI or automation disclosures are useful where readers might reasonably wonder how the content was produced.
What the guidance does not say matters as much. It does not describe a penalty for AI use as such, and it does not publish how, if at all, Google distinguishes AI-assisted text from human text. Google has not documented a detection method, and this guide does not speculate about one. The practical reading for a content team is that the familiar SEOImproving visibility in unpaid search results. principles apply unchanged: a page earns its place by being more useful, more specific and more trustworthy than what already ranks, and the production method is neither a shortcut to that nor a bar to it. A team that uses AI to publish large volumes of generic pages is in the territory the spam policies describe; a team that uses it to draft, edit and adapt pieces that people with expertise have shaped and checked is doing what the guidance describes as acceptable.
That shaping and checking is the human reviewThe step where a named person checks AI-assisted work for accuracy, sources, originality, privacy and disclosure before it is published or sent. step, and the AI-assisted content review checklist lists what a reviewer confirms before an AI-assisted piece is published.
Common mistakes
- Treating the draft as the deliverable. The draft is raw material; the deliverable is a checked, edited piece that someone stands behind.
- Drafting without sources. A draft from general knowledge cannot be verified without redoing the research.
- Reviewing for fluency. Reading a draft to see whether it sounds right instead of checking whether each claim is right.
- Using AI for experience pieces. Asking a tool to produce the first-hand account that is the only reason the piece would be worth reading.
- One prompt for every topic. The sameness readers notice across a programme comes from reused prompts and absent briefs.
- Letting derivatives add claims. A summary or thread that introduces a figure the source never contained.
- Shipping machine translation unreviewed. Literal translations that read as foreign and carry references that do not apply.
- Reading the search guidance as permission or prohibition. It is neither; it is a description of what is rewarded, and that has not changed.
The applications section sets content beside the other applications and gives a way to judge any of them by value, review cost and risk.
