AI in digital marketing 2026: generative AI, AI search, agents and how marketers actually use them

AI in digital marketing has stopped being a pilot and become plumbing. Salesforce’s ninth State of Marketing survey found 76% of marketers now use some form of AI, but only 13% use agents, and the split matters: the money is in ad automation and measurement, while the risk sits in customer-facing content. This guide covers what the tools do, how AI search changes demand, and what to buy.

What marketers actually use AI for

The honest answer in 2026 is: drafting, sorting and buying. Not strategy. Salesforce’s ninth State of Marketing report, a double-blind survey of 4,450 marketing professionals in 26 countries published on 25 February 2026, found 76% using at least one form of AI, whether predictive, generative or agentic. Only 13% had agents running. And 61% said adoption was high while full integration was still a work in progress, which is a polite way of saying the tools sit beside the workflow rather than inside it.

That 98% data figure is the one worth staring at. Nearly every team using AI for personalisation hits a data wall, and 46% said they lack enough customer preference data to act on. The average marketing organisation is stitching together seven data sources. AI does not fix that. It makes the gap more obvious, because a model asked to personalise an email with nothing but an order date will happily invent a reason to write.

The tasks where AI has genuinely stuck are unglamorous. First drafts and variants. Summarising research calls and survey verbatims. Translating and adapting creative for markets. Classifying support tickets and reviews into themes. Writing the SQL or the spreadsheet formula nobody wants to write. Producing 40 ad headline variations in a minute so a human can throw away 36. On the buying side, bidding and budget allocation have been machine-run for years and almost nobody argues about it now.

AI search and where demand now shows up

The second-order effect of AI is bigger than the tooling. People are asking machines questions they used to type into a search box, and the machines answer without a click. Google’s Gemini app passed one billion monthly users on 11 August 2026, generating more than 150 million images a day, with 63% of users talking to it rather than typing. OpenAI said ChatGPT had 900 million weekly active users in February 2026.

Inside Google itself, BrightEdge tracking reported by Search Engine Journal on 1 March 2026 showed AI Overview coverage grew 58% in twelve months to reach close to half of all tracked queries, with healthcare at 88% and B2B technology at 82% by December 2025. Prevalence estimates vary a lot by panel, because a tool that tracks keyword-research terms sees a different Google than a tool tracking long informational questions. The direction is not in dispute.

What that means for traffic is covered in depth in our SEO in 2026 guide. What it means for AI as a channel is this: the referrals are small but growing fast and converting well. Similarweb’s Generative AI Landscape report, published on 3 September 2026, put worldwide AI referral traffic at 770.7 million average monthly visits for June 2025 to May 2026, up 117.4% year on year. Marketplaces took 46.8 million average monthly visits, news and travel 44.5 million each, finance 19.1 million and consumer electronics 17.1 million. Growth was fastest where it started smallest.

Then there is quality. Adobe Digital Insights, cited in the same Similarweb analysis, found AI-sourced traffic to US retail sites converting 54% better than non-AI sources by May 2026. That premium has a boring explanation. Someone who has spent four turns describing what they want to an assistant arrives on your page already qualified. A broad organic searcher has not. If you sell online, the practical follow-through sits in our ecommerce marketing guide.

Generative engine optimisation: getting cited, not just ranked

Generative engine optimisation, or GEO, is the work of getting your pages quoted inside AI answers. It is not a separate discipline with its own secret levers. It is mostly good technical SEO plus a writing style that machines can lift cleanly. But there are real differences from classic ranking work, and three of them are worth acting on.

First, the unit of retrieval is the passage, not the page. An assistant breaks a question into sub-questions, retrieves candidate passages for each, and assembles an answer. A 3,000-word page with one buried answer loses to a 900-word page with the answer in the first two sentences under a clear heading. Front-load. Put the number and the date in the same sentence as the claim.

Second, engines disagree with each other. Perplexity cites far more sources per answer than ChatGPT does, and the overlap between the domains each one cites is small. Optimising for one and assuming the others follow is how teams end up invisible in three places at once.

Third, being quotable means being checkable. Named authors with real credentials, visible publication and update dates, inline links to primary sources, and specific figures rather than ranges. Models are trained to prefer content that looks sourced, and human editors reviewing AI output are too.

Measurement is the weak spot. Google Search Console reports impressions from AI features but not clicks from them, and ChatGPT and Perplexity report nothing to you at all beyond referrer data in analytics. So AI visibility tools exist to fill the gap by running prompt panels and recording which brands and domains get named. Ahrefs Brand Radar sits on paid plans from around $129 a month, Semrush bundles AI visibility into tiers from roughly $199, and dedicated platforms such as Profound start several hundred dollars higher. Buy one only after you have a reason to check a specific set of prompts weekly. Our content marketing guide covers how to build the pages those prompts should find.

AI inside the ad platforms

This is where AI has quietly earned the most money, and where marketers have the least choice. Google and Meta are both removing manual levers and replacing them with models. You can argue about it, but you cannot opt out for long.

Google’s AI Max for Search, announced on 6 May 2025, bundles keywordless search term matching, text customisation and final URL expansion into existing Search campaigns. Google reports 14% more conversions or conversion value at similar CPA or ROAS for advertisers who turn it on, rising to 27% for campaigns that were still mostly running exact and phrase keywords. On 15 April 2026 Google confirmed Dynamic Search Ads are being folded into AI Max, with campaigns using automatically created assets and campaign-level broad match auto-upgrading from September 2026 and the wider DSA sunset pushed to February 2027.

Meta’s numbers are bigger in adoption terms. In its Q1 2026 results, reported by ppc.land on 29 April 2026, Meta said 8 million advertisers were using at least one AI ad creative tool, double the 4 million at the end of 2024, against $55.02 billion in quarterly ad revenue. The performance claims are smaller than the marketing suggests: an adaptive ranking model expansion delivering a 1.6% conversion rate increase across Facebook and Instagram, and advertisers using video generation seeing a 3% conversion rate gain in large-scale tests.

My read: turn the automation on, but keep the inputs and the guardrails manual. Feed the models clean conversion data, exclusion lists and creative that reflects an actual brief. Broad-match plus Smart Bidding plus generated assets with no negative keywords and no brand rules is not a strategy. Full detail on account structure and bidding lives in the PPC and paid search guide, and the current month’s changes are in our AI marketing news for September 2026.

AI for measurement and reporting

Measurement is the least discussed and most useful AI application in marketing. Not because a model writes your dashboard commentary, though it can, but because statistical methods that used to need a consultancy are now free code.

Google open-sourced Meridian, its Bayesian marketing mix modelling library, and has since added Meridian GeoX for geo-based incrementality experiments, announced in May 2026 and covered by MarTech. Meta maintains Robyn on the same principle. Mix modelling used to be a six-figure engagement with a nine-month cycle. It is now a library, a dataset and someone on your team who can run Python. That does not make it easy, and a badly specified model produces confident nonsense, but the gate is gone.

The pairing that works is mix modelling for budget allocation, geo holdout tests for causal proof, and platform attribution only for in-channel optimisation. If you want the full picture including GA4, server-side tagging and consent, read the marketing analytics guide.

Agents, MCP and what they should be allowed to touch

Agentic AI is where the gap between demo and deployment is widest. Salesforce found 13% of marketers using agents, while 82% of those using or planning them expect meaningful ROI improvement. Expectation is running well ahead of practice.

The technical reason agents got real in 2026 is the Model Context Protocol, an open standard Anthropic published in November 2024 that lets an assistant connect to external tools and data through one authenticated interface. Through 2026 the ad platforms adopted it. Amazon Ads shipped MCP support in February 2026, with Google Ads, Meta Ads and TikTok Ads following between April and May, and HubSpot in public beta. That turns “pull last week’s spend by campaign and flag anything above target CPA” into a question rather than a report request.

Where it stops being sensible is write access. An agent that can change bids, pause campaigns or publish pages can also do all of that wrongly at three in the morning with no one watching. The pattern that works today is read, draft, approve. Let the agent read accounts and data, let it draft the change or the copy, and keep a person on the button. Revisit that in six months, not before.

Hallucination, disclosure, copyright and the backlash

Four risks are worth managing properly. The rest is noise.

Accuracy. A BBC study of 100 news questions put to ChatGPT, Copilot, Gemini and Perplexity, reported by Search Engine Journal, found 51% of answers had significant problems, 19% of responses citing BBC content contained factual errors such as wrong dates and figures, and 13% of quotes attributed to the BBC were altered or did not exist. Models have improved since, but the failure mode has not changed: they are most wrong when the question contains a false premise, which is exactly what a busy marketer’s prompt often does.

Disclosure. The EU AI Act’s Article 50 transparency obligations became enforceable on 2 August 2026. Providers must mark synthetic audio, image, video and text in a machine-readable format, and deployers must label deepfakes and certain AI-generated text published to inform the public. Penalties run to EUR 15 million or 3% of worldwide annual turnover. In the US, the FTC applies existing deception and endorsement rules to AI content, and on 1 July 2026 it sought comment on a policy statement covering distorted AI outputs. If you run creator campaigns, the disclosure stacking question is handled in our influencer and creator marketing guide.

Copyright. On 20 July 2026 a US federal judge approved Anthropic’s $1.5 billion settlement with authors, at roughly $3,000 per work across about 500,000 works. The earlier ruling held that training on lawfully acquired text was fair use while obtaining it from piracy sites was not. For marketers the practical takeaway is narrower than the headline: your exposure is usually in outputs, not training. Generated images that reproduce a protected character, or copy that lifts a competitor’s phrasing wholesale, are the things that get letters.

Audience reaction. This is the underrated one. IAB and Sonata Insights research reported by Digiday on 12 February 2026 found 82% of ad executives believed Gen Z and millennials feel positive about AI-generated ads, while only 45% of those consumers did. Gartner expects around 20% of brands to position on the absence of AI. Yet the same reporting cites Taboola and Columbia University research where AI ads averaged a 0.76% click-through rate against 0.65% for human-made ads, and a VML study where only 21% would like a campaign less on learning AI was involved. Performance and sentiment are pointing in different directions, which is uncomfortable and true.

“The closer it gets to defining the meaning of the brand itself, the more cautious leaders become.” Sunny Bonnell, CEO of Motto, speaking to Digiday, May 2026.

A practical AI stack by team size

Buy fewer tools than you think. Most marketing teams get 80% of the value from one general assistant, the AI already inside the platforms they pay for, and one specialist tool for whichever job is their bottleneck. Prices below are published list prices as of September 2026 and change often.

Two spending rules. Do not buy a tool for a job you are not already doing manually, because AI accelerates a process, it does not invent one. And check what you already own before you buy: Gemini is bundled into Google Workspace tiers, Copilot into Microsoft 365, and both ad platforms give their generative features away free inside the ad product.

A 30-day plan to put AI to work

Week one: write the rules. One page, agreed by whoever signs off marketing. What AI may draft, what it may never publish unreviewed, which tools are approved, what customer data may go into a prompt, and how AI-generated creative gets labelled for EU delivery. Teams that skip this end up with a policy written after an incident.

Week two: pick three repeating tasks and time them. Something like weekly performance commentary, product description drafting, and turning webinars into social posts. Record how long each takes now. Run each through your assistant with a proper prompt that includes brand voice, examples and constraints, then time it again. Keep what saves real hours and drop the rest without sentiment.

Week three: fix the answer-first paragraphs on your top ten commercial pages and add author bylines and update dates. Then set a baseline for AI visibility by running 20 real customer questions through ChatGPT, Gemini and Perplexity and recording who gets named. Do it manually before you buy a tracker. Twenty prompts in a spreadsheet tells you whether you have a problem.

Week four: turn on one ad automation you have been avoiding, with proper guardrails, and give it a fixed test window and a single success metric. Connect one read-only data source to your assistant. Then write down what you learned and what you will not repeat, because the fastest way to waste a year on AI is to run twelve unrecorded pilots. If you are building a channel plan from scratch, start with the complete digital marketing guide and the weekly digital marketing news roundup to keep up.

Where AI does not help yet

Positioning. A model can summarise what your competitors say and can generate fifty taglines, but it cannot decide what your company is for or which customers you are prepared to lose. That decision is where most marketing performance actually comes from, and it is made by people with commercial context the model does not have.

Original research is another gap. AI cannot survey your customers, run your pricing test or sit in on sales calls. It can help you design the survey and analyse the answers, which is genuinely useful, but the data has to exist first. In a market where every competitor is generating summaries of the same public sources, proprietary data is the cheapest way to be different.

Then there is judgement on the edge cases. Which complaint escalates, which partnership is off-brand, when a campaign should be pulled after a news event. Salesforce’s finding that 61% of marketers describe integration as unfinished is not a tooling problem waiting for the next model. It is the ordinary work of deciding which decisions you are willing to hand over. Testing what actually converts remains a human-designed process, covered in our conversion rate optimisation guide, and platform-specific tactics sit in the social media marketing guide.

Last researched and updated: 7 September 2026.