Content strategy spent a decade answering one question: how do you get Google to send people to your site. That question is no longer enough. A growing share of your readers are not people arriving through a search result, but models that read the content, condense it and hand the answer to the user — usually with no click at all.
This does not mean content is dead. It means its job has changed: from pulling in traffic to teaching the systems that produce the answer. This page sets out how we build a content strategy that works in both worlds at once.
What changed: from traffic to citation
When someone puts a business question to a model, the model assembles an answer from several sources and sometimes names them. You can be one of those sources, or not. If you are, you have gained something stronger than a click: a mention in the context of a recommendation. If you are not, you are simply absent from the conversation, even if your site ranks well in organic search.
This changes the unit of planning. It is no longer “an article built around a keyword” but “a clear answer to a question people actually ask”, written so that a self-contained passage can be lifted out of it without losing its meaning. The metrics that show whether this is happening — mention frequency, citation rate, fidelity of representation and referral volume — are set out in ourguide to measuring visibility in AI engines.
The precondition: content a machine can read at all
Before we talk about quality, there is a technical question most sites fail. A model or an agent that reaches your page does not run JavaScript the way a browser does. If the text loads dynamically, if the important information is hidden behind an accordion or a tab, and if the heading hierarchy is missing, what the agent sees is an empty shell.
This problem is completely invisible in an ordinary check, because to you the page looks fine. We went into it inYour site was built around a gravity that no longer exists, and into the wider design implications inThe screen is dead, the agent is in: from UX to AX. Every content project we take on starts with this readability check, because without it everything else is wasted effort.
The four layers of a content strategy that works in 2026
Layer 1: Mapping the real questions
Not keywords — questions. What the customer actually asks before buying, in their own words. The best sources are transcribed sales calls, support enquiries, and the questions people genuinely put to answer engines. That mapping produces a topic list driven by buying intent rather than search volume.
Layer 2: Cluster structure
One flagship asset per subject area, surrounded by articles that each answer a single sub-question, linked in a clear hierarchy. This structure does two things: it stops your own pages competing for the same intent, and it gives the model context — it sees a body of knowledge rather than an isolated article. We described how we work with a properly structured knowledge base inan enterprise knowledge base for RAG.
Layer 3: Depth you cannot generate with a prompt
Generic content can be produced today in seconds, in unlimited quantity, for nothing. It will not set you apart from anyone. What does set you apart: numbers from your own work, case studies with a measurable result, decisions you took and the ones that turned out to be wrong. That is also what a model chooses to cite — a source with specific information it cannot find anywhere else.
Layer 4: Maintenance and refresh
Content that has not been updated in two years loses value in the eyes of Google and of a model alike. We work to a quarterly review cycle: what has lost accuracy, what has slipped in ranking, and what should be merged because two pages are competing for the same question.
Page structure: how to write something worth citing
A model assembling an answer looks for a passage that stands on its own — a paragraph you can lift out, without what comes before or after it, and still understand. That sounds technical, but it is a writing decision: answer the question in the opening paragraph, and only then expand. The opposite habit, building a long run-up towards a conclusion, works against you.
Three practical rules we apply to every asset. First, subheadings phrased as a question or a full statement rather than a label — “How long does it take to see results” and not “Timelines”. Second, numbers and ranges in the body text and not only in a chart or an image, because whatever sits inside an image is not read. Third, explicit definitions — if you use a technical term, define it in one sentence at the point where it first appears.
There is a credibility dimension here too. A model tends to cite a source it can verify: a visible update date, a named author, and linked sources for claims that are not your own. These are not decoration but signals that influence which source is chosen for the answer. We gathered the wider background on how answer engines work inRecommended AI engines.
Where AI helps in content production — and where it does damage
Helps: early research and question mapping; drafting an outline and structure; language editing and adjusting length; translation and adaptation across languages; and producing derivatives — condensing an article into a post, pulling key points into a video script. We wrote about the tools on the video side inAI tools for video creation.
Does damage: end-to-end content production with no content specialist reviewing it. The result reads well, looks correct, and contains nothing that could not have been produced without you — which is to say it lacks exactly what would make a model cite you. This failure follows the same pattern as the other adoption failures we described inwhy most AI tools end up on the shelf.
There is a further consideration: what may be fed into a model at all. Content that draws on customer data or internal information needs to be classified before it touches an external tool — the subject is covered in detail inclassifying corporate information for LLMs.
Content, marketing and sales: a single loop
A content strategy disconnected from sales produces traffic that never converts. What connects the two is a shared definition of what counts as a quality enquiry, and feeding what is learned in sales conversations back into the content. We set out the full framework inMarketing and sales are not talking — and it is costing you money, and the sales side inDigital sales.
A parallel shift is under way in paid media, because advertising is moving into answer engines too. We gathered what is known so far inPPC in AI engines.
What the engagement looks like
Diagnosis (2–3 weeks). An audit of existing content — what ranks, what gets cited, what competes with itself, and what agents cannot read at all. The stage ends with a list of actions ranked by return.
Structure (2–4 weeks). Defining the clusters, choosing the flagship assets, and planning internal linking. Deletions and merges are decided here too — usually the highest-return action in the whole project.
Ongoing production. A regular cycle of writing, professional editing and publication, with complete meta fields and schema on every asset.
Quarterly measurement. Organic visibility, visibility in answer engines, and conversion in practice. Metrics not connected to revenue stay out of the report.
Frequently asked questions about content strategy
How long before we see results?
A change in organic visibility is measured over three to six months. Visibility in answer engines can move faster — sometimes within weeks — because the models refresh more often than organic ranking does.
How much content do we need to produce?
Less than people tend to think, and deeper. Ten outstanding assets beat fifty mediocre ones, because the mediocre ones compete with each other for the same intent and dilute one another.
Is AI-written content a problem for Google?
Google states that it measures quality, not method of production. In practice generic content struggles to break through regardless of who wrote it, so the practical question is not who wrote it but whether there is something inside that exists nowhere else.
What do we do with an archive of hundreds of old articles?
You do not delete them in one go. You sort them into three groups — refresh, merge, remove from the index — and work through them in order of contribution. Merging two competing pages usually produces an immediate lift.
Do we need schema on every page?
Yes, and beyond the basic markup. Schema is the most explicit way to tell a machine what is on the page, and when it is done properly it shortens the model’s route to the answer.
Should an in-house team write, or an external supplier?
The knowledge has to come from inside; the writing does not. The model that works best is a short interview with your internal expert, from which an asset is drafted and edited externally and then comes back for approval. That keeps the content distinctive without tying the pace of production to the availability of a busy person.
What do we do when a competitor copies our content?
It does happen, and most legal action is not worth the cost. The best protection is content built on your own data and your own case studies — wording is easy to copy, numbers the copier does not have are not.
Start with a content audit
The first step is knowing what you already have and which of it works — including what machines can read at all. Talk to us.
