AI in marketing: what actually works

The useful question is not whether to use these tools. It is which parts of the work get better when a machine does them, and which parts get quietly worse.

Two years of enthusiasm have produced a reasonably clear picture. The gains are real and they are concentrated in a narrower band than the sales material suggests. The losses are real too, and they show up later, which is why they are underreported.

Where it genuinely helps

• Volume work with a verifiable output. Resizing and versioning, transcription, first-pass translation, alt text, product descriptions from structured data, cut-downs from a long edit.
• Research compression. Reading a hundred reviews, clustering open-ended survey answers, summarising a category, pulling the recurring objection out of sales call transcripts.
• Drafting, where someone who knows the subject then rewrites it. The draft is scaffolding, not output.
• Analysis you would not otherwise do. Anomaly checks across campaigns, pattern finding in performance data, the second look nobody has time for.

Where it quietly degrades the work

Anything that needs a point of view. These models produce the consensus of what has been written. In a crowded category the consensus is precisely what everybody else is already saying, which is why AI-written thought leadership reads as competent and changes nothing.

Facts you do not check. Confident wrong numbers are the characteristic failure, and in regulated categories a wrong number is a compliance event rather than an embarrassment.

Volume for its own sake. Publishing forty thin pages a month because it is now cheap produces a site search engines treat as low value and buyers ignore. The economics of content did not change; only the cost of the first draft did. The standard is still the one in content that earns links.

Disclosure and risk

Synthetic faces and voices in advertising carry real exposure: likeness rights, platform policy and the reputational cost of being caught presenting a generated person as a customer. Do not present generated imagery as documentary. Do not use a synthetic voice resembling a real person without permission. Keep a record of what was generated and with which tool, because you will eventually be asked.

The other side of the coin

AI is also now a distribution channel rather than only a production tool. A growing share of buyers ask an assistant before they search, and what it says about you is assembled from sources you do not control. That is a marketing problem with a marketing answer, which we set out in AEO and GEO and in how answer engines choose sources.

How to judge a tool

Ask what it replaces and what it adds to the review queue. A tool that produces ten times the output and requires a human to check every item has moved the bottleneck rather than removed it. Ask where the data goes, which matters for client confidentiality. And run it against work you have already done, where you know what good looks like, before you run it against work you cannot yet evaluate.