Sean Summers begins by qualifying the premise of the question. His company is not leveraging generative AI much, because it is still finding opportunities in the traditional machine learning AI it has used across every area of the business for more than ten years.
He is not worried about being late. The reason he gives is that starting from a solid base in traditional AI, with known processes and methodologies, makes scaling easier when generative AI does become central.
Key findings
- The first generative AI use case was text: email, push notifications and search advertising. It produced about a 10 percent improvement in open rates and similar improvements in conversion rate.
- The point of that work is volume of iteration rather than better single executions: producing copy at scale so five versions can be tested instead of one, and continuously iterating.
- The change he wants is who asks the question. Previously a human revisited campaign creative once or twice a year. He wants the system constantly asking how to improve the creative and running the tests itself.
- The company is not asking generative AI to originate. The process does not start with it, and briefing a campaign from scratch is something he considers they are not ready for.
- Media planning was taken away from agency tools more than a year ago and given to the data science team using traditional AI, with the stated goal of taking full control of the entire planning process.
Refusing the black box
The pattern he describes is consistent across tools. Start with the platforms of the large players, then stop trusting them, because everyone has their own self-interest and the company has the scale to build its own view.
His objection is explicit. The platforms want all the data so they can optimize, and they like their black boxes, which he does not.
The company built its own metric to replace reach and frequency, calling it an effective contact, because the number of impressions needed to shift consideration differs by platform.
The other principle is incrementality. Roughly half of spend goes to branding and half to performance, and all of the performance half is judged on whether transactions were genuinely incremental rather than ones that would have happened anyway.
The consequence is that reported returns are discounted. If a test shows a share of transactions would have occurred without the spend, that share is not credited, and the optimization algorithm is fed the lower number.
The advantage he keeps disclaiming
He repeatedly qualifies his own answers because of the resources behind them. The company employs around 20,000 developers with an AI team inside that, and hires thousands of engineers every year.
Marketing has direct access to that capacity, with around 300 engineers dedicated to marketing projects and the ability to borrow more from elsewhere in the organization.
He is explicit that this makes his experience hard to transfer. A traditional marketer lacks the tools and the scale, and is still relying on the platforms’ black boxes, which makes campaigns measurable but hard to truly optimize.
Where generative AI is already paying
The company’s generative AI work started with product listings rather than advertising. Improving listings using existing metadata is not a marketing initiative, but better listings produce better conversion rates, which changes the return on every performance campaign.
Reviews are the second case. On a popular item there may be thousands of reviews, which beyond a star rating are of little use, so generative AI synthesizes the top five reasons to buy and not to buy.
He is emphatic that this unglamorous work is the business. What information goes into a listing, what is left out and in what order it appears all inform how relevant the algorithm judges a product to be, and position in the results makes a large difference.
Outside marketing, generative AI agents now manage tens of thousands of legal claims without lawyers, and post-sales customer interactions are handled entirely by an agent. Internal coding productivity gains were real but smaller than expected.
The harder frontier is an assistant that accompanies a user through the journey, in shopping or in financial services. It sounds easy and is highly complex if bad recommendations are to be avoided.
Against silver bullets
His warning is about expectation. Marketers are promised large multiples of impact or productivity and go looking for a golden goose, while ten years of experience tells him there are no shortcuts and no silver bullets.
He expects everyone to produce one success case, and sets a harder test: proving generative AI has a positive impact on overall marketing day in and day out rather than in a sporadic showcase.
What limits progress is not resources but choice. He says the company has everything it needs and that it is a matter of priorities, chosen on expected business impact rather than on how interesting a use case is.
His complaint about published cases is the same. They read as interesting in a slide or on a conference panel rather than as true business impact.
On the fear of falling behind, he is unbothered. These things take years to materialize into something with impact.
Marketing’s seat at the table
He rejects the standard complaint about marketing lacking influence. In his account the seat was lost because marketing stopped speaking the language of business while sales and finance were dealing with day-to-day reality.
He does not expect recognition, and says his company will never call itself a marketing company. His measure is behavior rather than praise: every month the business gives marketing more money, which tells him it is delivering results.
His conclusion is that going back to basics and proving day-to-day business impact is what restores marketing’s standing.
He credits the habit of the company rather than himself: trying a lot, staying critical, learning new things constantly and unlearning old tricks.
Scope and limitations
The source is an automatically produced transcript with garbled passages, and the speaker repeatedly cautions that his company’s scale and engineering resources make his experience hard to replicate elsewhere.
Source
This page summarises Transcript: Eva Linderborg and Sean Summers (MercadoLibre), by Eva Linderborg, Sean Summers, March 3, 2025.