In field experiments, generative AI for marketing produced measured sales gains in a pattern consistent with removing friction a customer already feels. In randomised field experiments, answering pre-sale questions, interpreting search queries and filling in thin product descriptions all produced measured sales gains. Two areas are weaker. It is risky where an untuned model writes for a channel with its own rules, such as paid-search ad titles, and the evidence so far says nothing about long-term trust or loyalty.
The useful move is to stop treating “GenAI for marketing” as one budget line. A marketing plan usually mixes three kinds of work: tasks current models can automate, tasks that look automatable but quietly need super-human capability, and questions that need investigation before anyone commits to a build. Our GenAI use case feasibility framework sets out that sorting logic in full. Below, it is applied to marketing workflows one at a time, using measured results where they exist.
Generative AI marketing use cases with measured results
Evidence on individual marketing workflows comes from seven randomised field experiments. A large cross-border e-commerce platform ran them between September 2023 and June 2024. In each one, the treatment group differed from the control only by the addition of generative AI, while prices and the rest of the workflow stayed fixed. Samples ranged from 30,000 to 13 million subjects.
| Workflow | Measured effect | Reading |
|---|---|---|
| Pre-sale service chatbot | Sales +16.3%, conversion rate +21.7% | Ready |
| Search query refinement (Arabic, Japanese, Polish searches) | Sales +2.9% | Ready where baseline search handles intent poorly |
| Product description generation | Sales +2.1% overall; +6.5% for products whose original description had fewer than 50 words; no significant gain above 50 words | Ready for thin catalogue content |
| Marketing push messages | Sales +1.6%, not statistically significant; clicks +3.1%, orders +2.8% | Test first |
| Google Shopping ad titles, model not fine-tuned | Negative and statistically insignificant; fewer views and clicks than human titles | Risky without domain tuning |
| Banner ad imagery (separate study) | Up to 50% higher click-through than a human-made stock image, field study of over 173,000 impressions | Promising; test in your own channel |
The study found no evidence the gains came at the expense of measured customer satisfaction. Product return rates and customer ratings did not deteriorate after adoption. Returns actually fell by 3.9% for AI-written product descriptions and by 11.4% for AI-written push messages.
The image result comes from a different line of work. Hartmann, Exner and Domdey (International Journal of Research in Marketing, 2025) compared images from seven text-to-image models with human-made marketing images, using 254,400 human evaluations, and found the AI imagery could surpass the human-made images on quality, realism and aesthetics. We would treat it as evidence that the use case is worth a controlled test. It does not settle what effect size to expect.
Why the ad titles did not work, and the rule we draw
The study’s authors read the search-refinement and live-chat translation effects mainly as friction reduction. There, generative AI improved semantic understanding, matching or communication, in a multilingual setting where such frictions are high. The gains came through higher conversion, not larger baskets. Newer, less active and lower-spending consumers benefited disproportionately.
The authors offer a specific explanation for the ad-title result. The model was not fine-tuned for e-commerce advertising, and the result is consistent with it omitting commercially salient keywords ad titles usually carry. That may have lowered Google quality scores and with them the chance of winning auctions at a given bid. The authors also suggest that Google’s algorithm may identify and deprioritise AI-generated titles.
From that contrast we draw a working rule. A use case is ready when the existing baseline is thin or missing and nothing downstream scores the output by rules a general model does not know. It is risky where, as with the ad titles, a platform’s ranking may grade output on keywords the model was not tuned for.
There are four limits on how far these results generalise:
- Short horizon. The experiments lasted days to months, so they measure short-run effects only, with no data on retention, trust, loyalty or engagement over time.
- One platform. All seven experiments ran on one cross-border e-commerce platform, a multilingual setting with high search and communication frictions.
- Non-random choice of workflows. Managers picked the workflows they judged most promising on feasibility, cost and value, so the set is not a random sample of marketing tasks.
- Top-line value. The estimated $4.6 to $5.2 annual value per consumer measures sales, not profitability or full return on investment.
So the honest answer on readiness is “ready for short-run conversion gains in friction-heavy workflows”. The evidence does not support “ready to run your brand”.
Is this use case feasible with our data, team and tools?
Use case feasibility depends on three things you can check before building: a defined benchmark, first-party data the task actually needs, and the people to review and approve output. Marketing teams already use the tools. Basis’s 2026 survey of marketing and advertising professionals found 98.4% use generative or agentic AI in their work. The same survey found that only 6.5% evaluate AI investments against defined metrics and benchmarks, and 65.1% either do not measure AI ROI or do not know whether anyone does. Only 13.0% described their organisation’s first-party data as extensive and well-structured.
Taken together, those figures make the measurement gap the main problem for generative AI for businesses in marketing. Adoption is close to universal already. Each of the field experiments above had a randomised control group. The push-message result was positive but not significant, which the authors say likely reflects a 0.16% baseline conversion rate and high variance in spending.
A pre-build feasibility checklist
- Benchmark first. Name the metric, the current baseline and the holdout design before choosing a model.
- Data the task needs. Product attributes for descriptions, brand proof and performance history for copy. If those inputs are missing or messy, the work is research, not automation.
- Domain vocabulary. Ask whether the channel rewards terms a general model lacks. In the ad-title test an untuned model produced no gain, which the authors link to missing ad keywords.
- Review capacity. Someone has to read, judge and refine the output, and that decision work becomes a job of its own.
- Approval path. Generative AI shortens production time but does nothing to shorten client approval time.
- Setup time. Unilever’s semi-automated production line took over a year to build. The setup included legal consulting, choosing suitable LLMs and turning brand identity and past output into a brief a model can use.
Brand voice and sameness: where generative AI content falls short
Generative AI marketing copy tends to sound like every other brand’s because a language model is trained to produce likely text. Given a thin brief, it returns the phrasing it has seen most often for that category. A model cannot add a distinction its input lacks, so a brief written from category norms produces generic output however capable the model is.
An online experiment found a related effect. Doshi and Hauser (Science Advances, 2024) found that access to AI story ideas made individual stories more creative and better written, but also made them more similar to each other: writers were better off individually while the group produced a narrower range of novel content. In that study, stories using one AI idea were 10.7% more similar to each other than stories written without AI.
The push-message experiment might look like a counterexample, since it produced nearly 2.7 million unique AI message variants against about 2,000 human ones. In our reading it is not. Variety across customers within one brand is a different thing from distinctiveness against competitors: millions of variants can all sit at the same category average.
The fix is to change what the model conditions on. When the model conditions on competitor messaging, evidenced audience tensions, the brand’s own product truths and its past winners and losers, the most probable output becomes the most probable answer for this brand rather than for the category. A cheap test follows from this: swap your brand name for a competitor’s in your last ten ads, and if most still read naturally, the copy is generic to the category.
Checking output is accurate and on-brand before it goes out
Every generative AI marketing pipeline needs a human review step before publication, because hallucination can be reduced but not eliminated. UC Berkeley’s Sutardja Center cites research concluding that it is impossible to eliminate hallucination in LLMs. The same review notes that Air Canada was found financially liable by Canadian courts for statements its chatbot made.
Mitigations help, but none of them removes the review. Fine-tuning improves factual accuracy, but it depends on training-data quality and cannot see information newer than its training. Retrieval and knowledge graphs are only as objective as the data fed into them. Automated consistency checks such as SelfCheckGPT may not scale to real-time use. We would map the review process against NIST’s 2024 Generative AI Profile of its AI Risk Management Framework (NIST AI 600-1).
Responsibility does not move to the tool. The UK Advertising Standards Authority states that the CAP Code is media-neutral and that the advertiser remains responsible for any AI output used in its ads. It held an advertiser primarily responsible for an irresponsible ad even though Google’s automated tools produced it. In the EU, Article 50 of the AI Act requires providers of systems that generate synthetic content to mark outputs in a machine-readable, detectable format. Deployers must also disclose image, audio or video deepfakes and AI-generated text published to inform the public on matters of public interest. People talking to an AI system must be told so unless that is obvious. Whether a given campaign meets these duties is a question for your legal counsel.
A pre-publication review covers, in order:
- facts checked against the product data or another source of truth, not against the model’s own answer;
- imagery and endorsements free of misleading AI likenesses;
- social responsibility and freedom from harmful or discriminatory stereotypes;
- the brand-distinctiveness test above;
- disclosure and machine-readable marking where Article 50 applies.
For customer-facing bots, see also our piece on ToxicChat toxicity benchmarking.
What generative AI for marketing costs at scale
The cost of generative AI for marketing goes well beyond the model: integration, tuning, infrastructure and human time. ITRex Group, a development vendor, estimates implementation costs from $0.0005 per 1,000 text units for an off-the-shelf commercial tool to over $190,000 for a fully custom solution on a fine-tuned open-source model. A fine-tuned closed model runs $10,000 to $50,000, an open-source model used as-is $20,000 to $50,000, and a fine-tuned open-source model $80,000 to $190,000 or more. Beyond model usage, it lists cloud compute or GPUs, integration, developer time, storage, MLOps and model retraining as ongoing costs.
Two cautions apply to those figures. The per-token prices are tied to named model versions such as GPT-3.5 and GPT-4. More usefully, they show where the step change sits. The ad-title finding says some channels need domain tuning, and the vendor prices tuned open-source builds in five- or six-figure ranges.
Some operational costs stay hidden. Usage adds up with volume: Coca-Cola’s Christmas ad reportedly used 70,000 prompts, and undisciplined testing with premium models can quietly inflate spend. Human selection and refinement of the many options a prompt returns becomes the real work. One practitioner quoted by Digiday says the gap between minutes of generation and approvals taking weeks or months is becoming the most expensive part of the pipeline. A cost model that budgets for generation and leaves out review and approval hours will understate the total.
Go or no-go on a marketing use case
Before any go or no-go, first ask whether the use case is technically feasible at current model capability. “Do this repeatable task instead of us” is far more likely to pay back than “be smarter than our best marketer”, and we avoid committing to outcomes that would need super-human capability. In practice, a chatbot answering pre-sale questions from catalogue data, refined search and descriptions for thin product pages pass that test. A model asked to find the next breakthrough campaign idea does not.
For a team with no measured baseline, the right start is one friction-heavy workflow, a randomised holdout and a review step already costed. A strategy deck that spreads one budget across every channel is the wrong starting point. Apply the full GenAI use case feasibility framework before any other money is spent.
One question remains open, and every team should settle it for itself before scaling: how will you know if the short-run lift is costing you long-run trust? The field study above has no data on it. A programme that cannot detect it is running on an assumption, however good its conversion numbers look.