Some of what’s sold as AI in marketing is genuinely new, but plenty of it isn’t. The real shifts have one thing in common: they use models that learn from data to predict behaviour or understand language, rather than following rules someone wrote.

Send-time optimisation, predictive audiences, machines that read in context and content briefing at scale all pass that test. A workflow that sends an email when a form is filled in does not, even when it’s sold as AI.

The distinction matters because teams are committing budgets and roadmaps to tools on the promise of AI rather than on what they actually do. Here’s how we separate the two.

Why is it so hard to tell real AI from rebranded automation?

“AI” has become a default label rather than a description. When the UK’s Advertising Standards Authority (ASA) reviewed online ads using the term between July and September 2024, it found around 16,000. Roughly two-thirds were for B2B products and services, from CRM and accounting software to social media tools.

Marketing software is no exception. Features that have existed for a decade, such as triggered workflows, A/B tests and lead scores built on points you assign yourself, now regularly appear under an AI banner.

The AI label changes how teams use what they’ve bought. Rules-based output gets trusted as if a model had learned it, when it’s only as good as the assumptions someone configured. Teams stop refining processes by hand, expecting the tool to improve, but it keeps doing exactly what it did on day one. Comparing vendors becomes guesswork when every product claims the same thing. The ASA advises advertisers not to exaggerate what AI features can do and buyers should hold vendors to that standard.

What has genuinely changed?

Four capabilities stand out because each does something rules-based software couldn’t do on its own. Two predict behaviour and the other two understand language, reading content in context and doing the groundwork behind a content brief.

Send-time optimisation

Traditional scheduling sends one campaign to everyone at once, at a time chosen by gut feel or a test across the whole list. Send-time optimisation uses a model trained on each contact’s past engagement to predict when that person is most likely to engage. It then holds their message until that moment.

For example, Salesforce learns from 90 days of engagement data, updates weekly and falls back to a general model when a contact has too little history. For marketing teams, a prediction for each person replaces a single best guess for the whole list. That removes a round of send-time testing.

The caveat is measurement. Apple’s Mail Privacy Protection downloads email content in the background when a message arrives rather than when it’s read, so open data is a weaker signal than it looks. Judge results on clicks and conversions.

Predictive analytics in GA4

Google Analytics 4 (GA4) is Google’s current analytics platform for websites and apps, which replaced Universal Analytics. Its predictive metrics apply machine learning to your event data. For each user, they estimate how likely that person is to purchase or churn and how much revenue they could to bring in.

The change is that audiences can look forward rather than back. Instead of segmenting by what people have already done, such as everyone who bought last quarter, teams can target users who are likely to buy soon or lapse. Retention effort can then reach people before they drift away and spend can lean towards those already close to converting.

It only works with enough data. A property needs at least 1,000 returning users who triggered the relevant condition and 1,000 who didn’t within the last 28 days. Google also stops updating predictions if model quality drops.

Machines that read content in context

Language models interpret meaning rather than matching keywords, so they can tell that “took ages to hear back” and “slow response times” describe the same problem. That changes two things for marketers:

• Internally, customer insight no longer depends on someone tagging call transcripts, reviews and support tickets by hand. Models can summarise and group them into themes in hours rather than weeks, so feedback can shape messaging while it’s still current.
• Externally, AI search systems read your pages and draw answers from them, so your content increasingly reaches buyers through someone else’s summary. Google says there are no extra requirements to appear in AI Overviews or AI Mode beyond normal SEO and no special schema.org markup to add.

Structured data still matters as long as it matches the visible content on the page. What counts is clear, well-organised writing that a model can understand and quote.

Content briefing at scale

A good brief used to mean hours of groundwork: reading what already ranks, mapping the questions an audience asks and shaping it all into a structure a writer can follow. Generative AI now handles that first pass in minutes. That leaves strategists more time for what it can’t do, such as choosing the angle and adding original insight.

Google’s own guidance supports using generative AI for research and to add structure to original content. It also warns that generating many pages without adding value can breach its scaled content abuse policy. Briefing at scale is the shift; publishing at scale isn’t.

What’s mostly noise?

Most of the noise comes from familiar features that have simply been rebranded as AI. Here are the ones we see most often:

• Triggered workflows: These are automated sequences that run when a contact takes a set action, such as “if someone downloads a guide, send email two”. They follow rules a person wrote and run them the same way every time. Nothing in them learns, however useful they are.
• Points-based lead scoring: This model gives each lead points for actions you choose, such as ten for a pricing-page visit, and flags leads that pass a threshold. The weightings are your judgement, so the intelligence is yours. Predictive scoring, which learns which behaviours actually come before a sale, is a different thing.
• Standard A/B testing: This means sending two versions, such as two subject lines, to split audiences and keeping whichever performs better. Picking the winner is statistics, not learning. It becomes something more only when a system shifts traffic as results come in or generates the variants itself.
• Bolt-on writing assistants: These are text boxes inside a platform that send your prompt to a general-purpose language model. The model is real AI, but the feature has no access to your data or brand guidelines, so it adds little you couldn’t get elsewhere.
• Threshold alerts sold as insights: These are notifications that fire when a metric moves past a percentage you’ve set. A fixed trigger can’t tell a real problem from a normal Monday dip. Anomaly detection, which accounts for seasonality and normal variation, can.

None of these features are bad. The problem is paying an AI premium for them or assuming they’ll deliver what genuine prediction does.

How can you test an AI claim before you buy?

Before purchasing a subscription to an AI platform, these are the five questions you should ask yourself to really understand the benefit and the return on investment: 

1. What does it predict or understand that a rule couldn’t? 

This is the core test. A genuine answer names a specific output, such as the likely send time for each contact or the probability that a user will churn, and explains that the system works it out from data. If the answer describes a sequence of steps someone has set up, it’s automation, however it’s packaged.

2. What data does it learn from and how much does it need? 

Models need volume to find reliable patterns. That’s why GA4 won’t produce predictions below its thresholds. A vendor who can name the data sources and the minimum amount needed has thought about where the product works and where it doesn’t. If they promise accurate predictions from any list size, either the product isn’t learning much or the predictions will be unreliable for smaller businesses.

3. Does it learn from your data or rely on a generic model? 

A generic model is trained on data from many businesses. It works from day one, but it reflects average behaviour rather than your customers’. A model trained on your own data can pick up what’s distinctive about your audience, but it needs time and volume first. Neither is wrong. Knowing which you’re getting tells you whether to expect results straight away or after a few months.

4. What happens when there isn’t enough data?

Every model meets gaps, such as new contacts, small segments or quiet periods. Good systems switch to a sensible default and say so. If a vendor can’t explain this, you won’t know which results come from the model and which come from a guess.

5. How will you measure lift? 

Ask for a holdout test. A share of your audience stays on your current approach, so you can compare results directly. Judge it on clicks, conversions or revenue rather than opens. This is the question that turns an AI claim into a business case. If the vendor resists testing, treat that as an answer.

If a vendor can answer all five clearly, the AI label is probably earned.

Is your content ready for AI?

Most of this piece has been about the AI you choose to buy. One change doesn’t wait for a buying decision. AI search systems already read your pages and summarise them for your buyers, whether or not you’ve added AI to your own stack.

The same rule applies here as with any vendor claim: look for evidence, not assumptions. Our GEO audit shows how AI search engines currently see your brand and where your content could be easier for them to understand and quote. Talk to us about where yours stands.