How to engineer attention with AI: what AI marketing really takes

Meta wants a business to supply only a product photo and a budget by the end of 2026. The AI then makes the ad, with images, video and text, chooses who sees it and suggests the budget1Reuters (via The Star), 2025Meta aims to fully automate advertising with AI by 2026, WSJ reportsOpen the source. Google is doing the same in search. Advertisers who switch on AI Max typically get 14% more conversions at a similar cost per conversion, according to Google2Google, 2025Unlock next-level performance with AI Max for Search campaignsOpen the source.

So what’s left for a marketing partner? Look at what’s being automated: making the variations, choosing the audience, spreading the budget. That’s the execution, and handing it over is a gain. What no platform can automate is what its AI learns from. Meta sees what happens in the browser, as far as cookies and ad blockers allow. It doesn’t see the quote that was accepted a week later, the booking that came in by phone or WhatsApp, the customer who cancelled, or what a customer was worth over a year. Those facts sit in your booking system, your till, your CRM and your bank, and they don’t reach the algorithm by themselves.

The AI optimizes for whatever you feed it. Count every click on a button as a conversion, and it will very efficiently find you more people who click buttons. Send it the real orders, with their value, and it looks for more people like your actual customers. The more of the execution the platforms take over, the more the input decides who wins: a message that lands, a page that convinces and data you can trust.

Together, those three numbers sum up the craft well. The creative does most of the work, a specialist who builds AI in properly makes the difference, and whoever measures well pays less for the same customer.

Every platform reports on itself. Meta and Google each count results by their own rules, and an order that came after both an Instagram ad and a Google search can show up in both reports. Meanwhile, the bookings are in one system, the payments in another and the enquiries in your inbox. We bring it together in one place: spend from Meta and Google, visits from Analytics and Search Console, enquiries from the website, and bookings, orders and payments from your own systems. Day by day and campaign by campaign, so you see what a real customer costs per channel, not what each platform credits to itself.

Then we tell the algorithm what really happened, from the server instead of the browser. Paid orders and bookings go to Meta through the Conversions API, and to Google with enhanced conversions, which uses hashed customer details to make measurement more accurate and bidding stronger6Google Ads Help, 2026About enhanced conversionsOpen the source. Sales that happen later or offline go back too: a quote that’s signed, a booking by phone, a payment at the counter. Meta recommends the Conversions API for offline and in-store sales, uploaded in real time or daily7Meta for Developers, 2026Sending Offline Events Using the Conversions APIOpen the source, and Google matches enquiries that later become customers through enhanced conversions for leads8Google Ads Help, 2026About enhanced conversions for leadsOpen the source. Always with consent, and with customer details made unreadable first.

That’s how Meta’s and Google’s automation starts working for your business instead of for the average advertiser. The same buttons, fed with better data, find better customers. It’s the core of what Forwarddd builds, and of forwarddd signal, which puts spend and real results side by side every morning and sends the real orders back to Meta and Google.

AI marketing is marketing in which AI does part of the work. Think of making variations of ads, analysing data, predicting who will buy and preparing answers to customer questions. Marketing automation is older and simpler: fixed rules that take the next step based on behaviour, such as an email when someone leaves their cart.

The strength is in the combination. Fixed rules where things are predictable, AI where understanding is needed, and a person who decides where it should go. How to make that distinction in practice is covered in the article on AI agents for business.

A marketer runs campaigns. An AI marketing engineer builds the machine around them: the pages, the tracking, the connections between ads, website, customer data and email, and the AI steps that prepare the work. The craft sits where marketing psychology, data and software meet, three things that used to sit with three different agencies. That’s exactly what Forwarddd does: one partner that builds your website, your tracking, your integrations and your campaigns, and adjusts them every week.

People recognize AI ads, and they don’t always like them. NielsenIQ showed more than 2,000 people AI-generated ads and also measured brain activity for some of them. Most AI ads were recognized, and people found them more annoying, boring and confusing than conventional ads. Even high-quality AI ads were remembered less well9NielsenIQ, 2024NIQ research uncovers hidden consumer attitudes toward AI-generated adsOpen the source.

The MIT research that’s often quoted as proof that AI doesn’t work actually says something else. Around 95% of AI pilots deliver no measurable result, mainly because the tools don’t learn from the way people work and don’t fit the rest of the business. The projects that do succeed more often run through a specialist partner: those succeed in about two out of three cases, three times as often as building it yourself. And the biggest returns came from automating back-office work4Fortune, on MIT NANDA's The GenAI Divide: State of AI in Business 2025, 2025MIT report: 95% of generative AI pilots at companies are failingOpen the source. Gartner sees the same pattern with AI agents: projects get canceled because of high costs, unclear value and poor control of the risks10Gartner, 2025Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027Open the source.

And every part has its own rules. Google warns that using AI to generate many pages without adding value may violate its spam policy11Google Search Central, 2026Google Search's guidance on using generative AI content on your websiteOpen the source. For personalized ads and measurement in Europe, you need to collect visitors’ consent and pass it on to Google12Google Ads Help, 2026Updates to consent mode for traffic in European Economic Area (EEA)Open the source. Influencer advertising has to be recognizable as advertising, whatever the number of followers13Ondernemersplein (RVO), 2026Reclameregels voor social media (influencer marketing)Open the source. A mistake in one step carries through to every step after it.

According to McKinsey, revenue gains from AI are most often reported in marketing and sales. The businesses that get the most out of it don’t put AI next to their existing work; they redesign the work around it. Nearly three quarters of them did so, against about a quarter of the rest. They also decide in advance when a person needs to check what an AI model produces14McKinsey & Company, 2026The state of AI in 2026: On the road to ROIOpen the source.

BCG reaches the same lesson. 74% of companies don’t yet get tangible value from AI. The leaders put 10% of their effort into the algorithms, 20% into technology and data, and 70% into people and processes15Boston Consulting Group, 2024AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale ValueOpen the source. The AI model is the smallest part of the work.

That’s how Forwarddd works. An example from practice: at a hospitality business, AI scans the supplier invoices and puts them in Exact Online after one click to check them. Exactly the kind of back-office work where MIT saw the biggest returns, built into the systems the owner already used. More examples are in Why modern businesses need connected digital systems.

The top of your marketing funnel. This is about getting the right people to stop scrolling, and being found where they search, including in ChatGPT.

AI turns one recording into dozens of variations: different opening seconds, different formats for Reels, Shorts and the feed, subtitles and translations. That makes testing cheap. The base still has to be real. Footage of your own business, product or team, and an opening line taken from conversations with customers. Given the NielsenIQ research, AI works best as a tool behind the scenes, not as the face of your brand9NielsenIQ, 2024NIQ research uncovers hidden consumer attitudes toward AI-generated adsOpen the source.

The Meta Ad Library shows every ad running in the EU, with its run dates, its targeting and who saw it. Meta keeps those ads for a year16Meta, 2023New Features and Additional Transparency Measures as the Digital Services Act Comes Into EffectOpen the source. An ad that has been running for months is usually paying off. AI helps sort hundreds of ads by promise, offer and format. The thinking is in asking why something works for their customer, and what you can honestly offer instead. Learning from someone is different from copying them.

More and more people ask an AI instead of Google. Research from Princeton and Georgia Tech showed that texts with sources, quotes and concrete numbers end up in AI answers more often, and that repeating keywords actually backfires17Princeton University & Georgia Tech (KDD 2024), 2023GEO: Generative Engine OptimizationOpen the source. How to approach it is covered in How to get found in ChatGPT.

With programmatic SEO you make hundreds of pages from one template and a data source, for example a page per service and per city. That can bring a lot of traffic, but it’s also exactly what Google warns about11Google Search Central, 2026Google Search's guidance on using generative AI content on your websiteOpen the source. A page should only exist if it holds something you can’t find elsewhere: real prices, availability, local examples or your own numbers. Collecting and checking the data is the work; generating the pages is the easy part.

AI helps find creators whose audience looks like your buyers, instead of choosing by follower count. Give each creator their own link or code, and you see what each one brings in and pay for results. In the Netherlands, the advertising code for social media and influencer marketing applies to everyone, including small accounts, so agree up front how the partnership is made recognizable13Ondernemersplein (RVO), 2026Reclameregels voor social media (influencer marketing)Open the source.

Attention is only worth something when it turns into something. This is where most businesses lose the most money, and it’s also the least visible part of the work.

Google rates every search ad partly on landing page experience: how relevant and useful the page is to people who click the ad18Google Ads Help, 20265 ways to use Quality Score to improve your performanceOpen the source. A page that delivers what the ad promised, with the same offer and the same image, pays less per click and loses fewer visitors. AI can make a version of the page for each ad group. Deciding which offer and which proof win people over is still the real work.

Tracking in the browser misses more and more: because of ad blockers, cookie choices and privacy settings on phones. With server-side tracking, you send enquiries, bookings and orders straight from the server. According to Meta, advertisers who also sent their web conversions through the Conversions API had, on average, a 17.8% lower cost per result5Meta, 2026Eliminamos barreras técnicas para ayudar a empresas de todos los tamaños a aprovechar más sus anunciosOpen the source. In Europe, properly set-up consent is part of that12Google Ads Help, 2026Updates to consent mode for traffic in European Economic Area (EEA)Open the source.

An email that goes out the moment someone does something works far better than a newsletter on a fixed day. In Omnisend’s data for 2025, automated emails were 2% of all emails sent, but brought in 30% of email revenue. Per email sent they earned $2.87 on average, against $0.18 for a regular campaign19Omnisend, 2026Omnisend's 2026 Ecommerce Marketing ReportOpen the source.

Think of a welcome series, a reminder for an abandoned cart, a message before an appointment, a request for a review and an email to customers who haven’t bought in a while. The thinking is in timing, tone and restraint. Send too often and you lose the list, and every email and text needs consent.

An ads agent reads your ad accounts, Search Console and Analytics, puts the spend next to your real results and works out what an enquiry, booking or order costs. Then it proposes changes. In forwarddd signal it works like this: Claude, an AI model, analyses and proposes, a person approves every change, and only within set limits can a change happen automatically. Every change is logged, and paid orders and bookings go back to Google and Meta, with consent and with customer details made unreadable first.

The same data produces a report every week: what happened, what probably caused it and what the next step is. AI writes the draft. A person checks whether the story holds, because a dip after a holiday calls for something different than a dip after a new ad.

This is the third part of the series. Is SaaS dead? explains why software increasingly adapts to the business, and Why modern businesses need connected digital systems shows how to get those systems working together. This part shows what you do with them: turn attention into customers.

Want to know what this could bring your business? See Forwarddd’s marketing packages, what automation can do for you, or how to get found with SEO and GEO.