The Workflows: Content, Personalisation, and Search

The difference between a content operation that scales and one that produces liabilities is structural rather than a matter of care.

The shape that works has four stages.

Brief. A human decides what the piece is for, who it addresses, what it must say and what it must not claim. This is the stage that determines quality and it is not automatable, because it is the strategy.

Ground. Supply the model with the approved material: specifications, approved claims, prior copy that established the position, and the style examples. The model works from this rather than from its own parameters, which is the grounding pattern from the hallucinations cursus applied here.

Generate. Produce several variants rather than one, since selection by a human from a set is faster and better than editing a single draft.

Review, with two distinct passes. One for quality and voice, which is what everyone does. One for claims specifically, checking every factual assertion against the supplied source, which is what almost nobody does and which is where the exposure sits.

The two-pass review is the load-bearing part. A single reviewer asked to assess a piece holistically will read for flow and pass a plausible invented statistic, because the two tasks use different attention and the fluent one dominates.

And the brief is where the leverage is. A generative pipeline with a weak brief produces polished content that says nothing, faster.

The clearest measurable win, and it works because the test itself supplies the evidence.

The old constraint: producing variants was expensive, so tests ran with few arms, which meant low power and slow learning. A team testing three headlines learns slowly and may never find the good one because it was never written.

What changes: generating thirty variants costs minutes, so tests can be wider. More arms means a better chance the winning variant exists in the set at all.

What does not change, and this is the part teams get wrong: the statistics. More arms means each receives less traffic, so per-arm power falls unless total traffic rises. Testing thirty variants on traffic sized for three produces noise that looks like a winner, and the existing A/B testing lessons cover why this fails and how multiple comparisons inflate false positives.

So the discipline is to widen the candidate set and then narrow it before testing. Generate thirty, have a human select the six most distinct, test those properly.

That sequence gets the benefit of cheap generation without abandoning the statistics that make a test mean anything.

Personalisation is the promise generative AI seems to deliver most directly, and the returns are real and shallower than expected.

What works. Segment-level adaptation, where the same substance is framed for different industries, roles or use cases. This was previously uneconomic below a certain segment size, and now the long tail of small segments becomes addressable.

What works less well than claimed. Individual-level generated content. The returns diminish sharply once the segment is small, because the information you hold about an individual is usually thin and the model fills the gap with plausible assumption. Personalised often means the model guessed, and a wrong guess about someone reads worse than a generic message.

Where it goes wrong. The creepiness threshold is real and it is crossed by demonstrating knowledge the recipient did not expect you to have. A message referencing a person's recent behaviour in detail produces discomfort rather than relevance, and the reaction is asymmetric: mild benefit when it lands, substantial harm when it does not.

And the data protection dimension. Personalisation is profiling, which engages transparency obligations and, at scale, the rules on automated processing. The lawful basis needs identifying rather than assuming.

The practical position: personalise at segment level, where the inference is safer and the gain is most of what is available. Individual generation is a smaller prize with a sharper downside, and the honest test is whether you would be comfortable if the recipient saw what you inferred about them.

The channel change with the largest strategic consequence, and the one most marketing plans have not absorbed.

When a search engine answers a question directly rather than returning links, the traffic that used to arrive from that query does not. The user got what they needed, and the site that would have supplied it was read rather than visited. The existing lessons on when the answer replaces the link and on generative engine optimisation cover the mechanics.

Three consequences for a marketing function.

Informational content loses traffic value fastest. Pages answering a factual question are exactly what an answer engine replaces. Content whose value is transactional, comparative, or requires the reader to be on your site retains more.

Being cited becomes the objective where being clicked was. That shifts optimisation toward being the source an answer engine draws on and attributes, which favours clear structure, factual density, and being the recognisable authority on a specific claim.

And measurement breaks. Impressions and clicks no longer capture influence, because a brand mentioned in an answer received value that appears in no analytics property. Tracking citation presence across answer engines is a different measurement discipline, covered in the existing GEO measurement lesson.

The uncomfortable strategic implication: a content strategy built entirely on ranking for informational queries was building an asset whose value is falling, and generating more of that content with AI is accelerating investment in a declining channel.

The highest-value use in most marketing functions is not content generation, and it receives a fraction of the attention.

The problem it solves. Organisations collect large quantities of qualitative customer input, support tickets, sales call notes, survey free-text, review sites, community discussion, interview transcripts, and read almost none of it systematically. It is too much for a person and too unstructured for traditional analysis, so it sits unused while the team makes positioning decisions on impression.

What AI does well here. Clustering large volumes of free text into themes. Surfacing the language customers actually use, which is frequently different from the language the company uses and is directly usable in copy. Identifying objections that recur. And tracing how the pattern shifts over time.

Why it is safer than content generation. The output is an input to human judgement rather than something a customer sees, so a wrong cluster is a hypothesis to check rather than a published claim. The failure mode is a misleading summary, and the mitigation is reading a sample of the underlying items, which is cheap.

The caution that matters. A thematic summary is a compression, and it will under-represent views held by few people, which are sometimes the important ones. So read the outliers deliberately rather than trusting the clusters, and treat the volume of a theme as one signal rather than as importance.

Most functions would gain more from doing this properly than from any amount of faster copy.

AI is genuinely useful for describing what happened in marketing data and genuinely unreliable for saying why, and the distinction determines whether it helps or misleads.

What it does well. Summarising performance across channels into readable prose. Surfacing anomalies worth investigating. Answering descriptive questions about data it can query. Drafting the recurring report that consumes an analyst's Monday.

Where it fails, and fails confidently. Causal claims. Asked why conversions fell, a model will produce a fluent, plausible explanation constructed from the correlations available and the shape of such explanations in its training data. It has no access to the counterfactual, and the answer will read exactly as authoritative as a correct one.

This is the hallucination problem in its most commercially dangerous form, because the output is a recommendation that redirects budget.

Three practical rules.

Use it for description and route causal questions to the experimental methods in the existing lessons on knowing what actually works. A causal claim needs a design, not a summary.

Require every stated number to be traceable to a query, so a fabricated figure is detectable.

And treat any generated explanation as a hypothesis with a named test, rather than as a finding. The useful output of asking why conversions fell is a list of things to check, not an answer.

A report that says conversions fell twelve percent, driven by mobile, is describing. One that says the checkout redesign caused it is claiming, and the model cannot know that.

Generated imagery, video and voice raise obligations that text mostly does not, and the timeline is nearer than the high-risk deadlines marketers have heard about.

Article 50 transparency duties apply from 2 August 2026. Deployers generating or manipulating image, audio or video content constituting a deepfake must disclose that it is artificially generated or manipulated, with an adjustment where the content is evidently artistic or satirical. Providers of generative systems must mark outputs in a machine-readable format, with systems already on the market having until 2 December 2026 for that marking requirement.

Beyond the Act, three practical considerations.

Synthetic people in advertising. Generated models and voices raise questions about consent and likeness, particularly where the output resembles an identifiable person, and several jurisdictions are legislating on voice and likeness specifically.

Generated imagery of products. An image showing a product with features it does not have is a misleading claim in visual form, and the consumer protection analysis is the same as for text.

And audience reaction, which is a commercial risk rather than a legal one. Disclosure requirements aside, audiences have responded badly to undisclosed synthetic media in campaigns, and discovering it later is worse than disclosing it initially.

The practical posture: disclose synthetic media as a default rather than at the legal minimum, and keep a record of what was generated and from what, because the question will be asked.

Pulling the workflows into an order.

Start with research synthesis. It is the highest-value use, the risk is low because the output is internal, and it improves every subsequent decision including what content to produce at all. Most functions skip it because it is not what AI is marketed for.

Then creative testing with a widened candidate set and a human narrowing step, which is the clearest measurable win and is self-evidencing through the test.

Then the content pipeline with grounding and a two-pass review, applied first to lower-stakes content while the claims discipline is established.

Then segment-level personalisation, stopping well short of individual generation.

Then analytics for description, with an explicit rule that causal questions route to experiments.

And revisit the content strategy in light of the search change, because generating more informational content is investing in a channel whose value is falling.

The thread across all of it: the uses where a human decides and the model produces volume are safe and valuable. The uses where the model's output is a claim or a cause are where the function gets into trouble, and both look identical in a draft.