What AI changes in marketing (and what it does not)
Three kinds of AI show up in marketing work. Generative models produce text, images, video, audio and code from a prompt and whatever material is supplied with it. Predictive models estimate a likelihood or a value, which is what automated bidding, audience expansion, propensity scores and forecasts rest on. Classification and retrieval models sort things into categories and find the relevant item in a pile, which is what powers tagging, routing, search and the matching of a query to a document. The first kind is new at this scale; the other two have been part of the platforms marketers buy from for a long time, which is why the AI in ad platforms guide treats automation as the default rather than as a novelty.
What changes is cost and speed at the first-draft stage. An outline, a set of headline variations, a resized banner, a summary of last month's report or a query against a data warehouse can be produced in minutes rather than hours, and produced again with a different brief. That moves the bottleneck. Production is no longer the slow part; checking is. A team that adopts AI without changing how it reviews ends up publishing more, faster, with the same number of eyes on each piece, which is how confident errors reach customers under a brand's name.
What does not change is everything that decides whether marketing works. The audience still has needs the team has to understand. The offer still has to be clear and the claim still has to be true. The economics of a channel still depend on margin, on the cost of acquiring a customer and on what that customer is worth. The qualities that search engines and readers reward in content, summarised as E-E-A-TExperience, Expertise, Authoritativeness, Trust., still come from people with experience of the subject. A model can frame those things; it cannot supply them. And accountability stays where it was: a published error is the organisation's, whichever tool drafted it.
One more thing is worth stating plainly. Tools change quarter by quarter, and what any of them produces today will differ next year. The guides in this hub are therefore written around tasks, failure modes and review habits rather than around particular products, because those are what carry over.
How AI works, in plain terms
The foundations section explains the mechanics a marketer needs before relying on a model's output: how language models generate text and why they sometimes state things that are not true, how groundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. and embeddingsA list of numbers that represents a piece of text so that texts with similar meaning sit close together, used for search, retrieval and clustering. work, and prompt patterns for marketing tasks that make output easier to check. A tutorial on preparing a document library for an AI assistant puts grounding into practice: one current source per fact, owners and dates, self-contained sections and tests. The section also compares rule-based automation with AI agents, and retrieval with fine-tuningTraining an existing language model further on a set of examples so that a style, format or task pattern is built into the model itself..
Applications
The applications section sorts AI use by the marketing task it supports. The core guides sit beneath it, with task guides alongside them, and each covers where a tool helps, where its output tends to fail and what review the task needs before anything goes live.
AI for content covers research and outlining, drafting from briefs, editing, repurposing and localisation, together with Google's published guidance on AI-generated content. It is the application most teams meet first and the one where plausible-looking output is most dangerous, because a draft that reads well is reviewed less carefully than one that reads badly. The AI-assisted content workflow in the content marketing section owns the step-by-step production process; the guide here takes the application view.
AI for creative covers image and video generation, variations built for testing, adaptation across sizes and placements, brand consistency, and the rights and disclosure questions that come with generated media. It also describes the creative tools the ad platforms document inside their own interfaces.
AI for analysis and reporting covers querying data in plain language, anomaly detection, summarising and narrating reports, and forecasting with the caveats it needs. Its central argument is that every AI-produced number has to be verifiable against the source before it is acted on, and it leans on the analytics and BI section for the definitions and metrics layer that make verification possible.
AI in ad platforms describes the automation that already runs inside Google Ads and Meta Ads as the help centres document it: bidding, audience expansion, asset assembly and the campaign types that bundle them. It sets out what the advertiser still controls and how to measure what has been handed over.
AI for marketing automation covers models inside workflows: classification and routing, personalisation at send time, agents that draft and act, and triggers from predictions, with the guardrails and approvals each needs.
Beyond the core guides, the applications section holds task guides for brand voice, keyword research, ad copy and image testing, customer service, research synthesis, SQL checking, product descriptions, synthetic research, AI in search results, and accessibility text. Further tutorials cover reviewing the AI enhancements and generated assets in Google Ads and Meta accounts, with an owner, a record and a re-check routine; checking what an AI tool writes before it is used, from tracking code proved in preview against the tracking plan and consent to structured data checked against Google's guidelines, the visible page and the feed and spreadsheet formulas and calculated fields; and writing the standing instructions an AI assistant follows in every conversation.
Governance
Using AI well is mostly a governance problem, and the governance section treats it as one. The question is not whether a tool is allowed but which uses are low risk enough to need only the usual editorial checks, which need a second person, and which need sign-off from someone who answers for the claim, the data or the customer relationship.
The core guides carry the detail. The accuracy and review workflows guide sets out the failure modes of generated output, how to ground a tool in sources the team supplies, how to run fact and claim checks, and how to tier review by risk so that a social caption and a pricing page are not treated the same. The disclosure and data privacy guide covers what to tell customers and platforms about AI use, what customer data may and may not go into a prompt, and how to read a vendor's data terms.
Two external references anchor the section. Google's guidance on AI-generated content, cited in this page's sources, says that its systems aim to reward helpful content however it is produced, that automation used primarily to manipulate rankings is covered by its spam policies, and that disclosure can be appropriate where readers would reasonably want to know how content was made. The NIST AI Risk Management Framework, written for organisations deploying AI of any kind, organises the work into four functions, govern, map, measure and manage, and describes the characteristics of trustworthy systems, among them validity, accountability, transparency and the management of harmful bias. A marketing team does not need the whole framework; it needs the habit the framework describes, which is to name the risks of each use before relying on it. Regulation in the European Union and elsewhere is moving in the same direction, towards transparency obligations for AI-generated content, and the governance guides describe that direction without offering legal advice.
The governance section also covers writing an AI use policy, measuring AI's impact on team output and building AI skills in a marketing team, along with claims, copyright questions, data minimisation, fact-checking, incidents and tool evaluation, and a tutorial on writing a sampling plan for one automated use. Further guides cover setting AI rules for the agencies and contractors working on a team's marketing, and answering customer questions about the business's AI use when customers ask whether they spoke to a bot, whether an image or review is real, or what happened to their data in AI tools. When something has already gone wrong, such as invented facts in a draft, an assistant giving customers wrong answers or a report figure nobody can reproduce, the AI in marketing troubleshooting section starts from the symptom and traces it to the layer that failed.
Skills for marketers
The skills that AI rewards are mostly ones good marketers already had, applied more deliberately.
Briefing. A prompt is a brief, and a vague brief produces generic output whoever receives it. The audience, the single question, the angle, the proof the piece may use and the voice all transfer from a content brief into a prompt; the ability to write them down precisely is the skill.
Evaluation. Judging output against a source, a brand standard and a reader's need is the work that replaces production. It includes knowing which failure modes to look for in which task, which the application guides set out one by one.
Data literacy. Plain-language querying lowers the barrier to asking questions of data and raises the importance of knowing whether the answer is right. Understanding how a metric is defined, where its data comes from and what a KPIA metric tied directly to an objective. is supposed to represent is the check; the SQL for marketers guide is the fastest route into it.
Platform literacy. Automated bidding and audiences decide most of what an ad account does. Knowing what the platform documents about each setting, what it treats as a suggestion and what as a constraint, is now a core paid media skill.
Experiment design. Generated variations are only useful if they are tested properly. The discipline of the A/B TestControlled comparison of two variants., with a hypothesis, a sample and a stopping rule, is covered in the CRO testing section, and it applies to AI output exactly as it applies to anything else.
Stack awareness. AI features arrive inside tools the team already pays for, and the martech hub covers how to judge them as part of the stack rather than as separate purchases.
A first AI adoption plan
A sensible first plan is small, time-boxed and built around review rather than around tools. The table sets out one shape it can take; the phases are sequential because each one earns the trust the next one needs.
| Phase | What the team does | What it decides | What it measures |
|---|---|---|---|
| Set the rules | Writes a one-page policy: permitted data in prompts, review tiers, disclosure, ownership of errors | Who approves new uses; which uses are off limits for now | Whether everyone can state the rules without looking |
| Pilot low-risk tasks | Outlines, headline variations, report summaries, data tidy-ups, with a named reviewer per task | Which tasks genuinely save effort after review time is counted | Time per task including review; defects caught in review |
| Add grounded tasks | Drafts from supplied sources, plain-language queries checked against the warehouse, creative adaptation | Which outputs need a second person; what the fact-check procedure is | Error rate found at review; how often a draft is rewritten |
| Hand over bounded automation | Platform bidding and audience features with exclusions and goals set; classification in workflows with approvals | What the automation may not do; which decisions stay manual | Outcome against the pre-automation baseline, by holdout where possible |
| Review and widen | Quarterly review of what was used, what failed and what was learned | Which pilots become standard, which are dropped | Share of output that passed review first time; incidents |
A retailer's first quarter with AI tools
A homeware retailer with a marketing team of six writes a one-page policy in week 1: no customer records in prompts, every published piece reviewed by someone other than the person who prompted it, and AI-drafted text disclosed in the editorial log. In weeks 2 to 6 the team pilots three tasks: category page outlines, headline variations for existing search ads, and a plain-language summary of the weekly trading report. It logs 48 outputs; 11 are rejected at review, mostly outlines that answered a broader question than the brief. Over weeks 7 to 12 it adds product description drafts grounded in the supplier specification sheets, with a fact-check against the sheet before publication. Of 120 descriptions, 9 are returned for invented features. At the quarterly review the team keeps the outlines, the headline variations and the grounded descriptions, drops the report summaries because reading the summary took as long as reading the report, and sets the error rate on descriptions as the figure to watch next quarter.
The plan deliberately leaves the ad platforms to the fourth phase, not because their automation is riskier than a drafting tool but because the measurement to judge it has to exist first. The AI in ad platforms guide explains what to put in place before handing over bidding and audiences, and the accuracy and review workflows guide explains the review tiers the earlier phases depend on.
