From keywords to conversations: helping advertisers make the GEA shift

If there is one product that fascinated me during my time at Google, it’s search. I had the privilege to observe the evolution of this technology as a customer success manager, GTM lead or product marketer for France and EMEA. It’s probably my background in literature and foreign languages and a difficult-to-explain fascination for words and semantics that have kept my interest and curiosity so high for so many years. How can natural language processing technologies infer meaning from keywords and match it to corresponding offers, commercial or not? There is a potential new big shift happening today with the launch of chatGPT ads in Europe, some would call it a paradigm shift or a revolution. A lot will be inherited from search as we know it and a lot will probably be brand new.

Same, same

Listening to multiple practitioner videos on Youtube and reading chatGPT ads related articles, it’s clear that this product innovation speaks to the SEA managers of the world. What chatGPT is often compared to is Google search ads. And the resemblance is shocking between the two interfaces. Same entities on both sides: campaigns, ad groups, ads, bidding, add-ons such as feeds or audiences. Early testers even confess sometimes copying-pasting their Google Search ads structures into chatGPT ads, as if each system were a mirror to one-another. After all, it’s about matching keywords, or semantic intents to content found in LLM conversations. When setting up a campaign in chatGPT ads, it is requested to add “Context hints” described as follows: “conversations, topics, or keywords where your products or services may be relevant; these hints guide matching but aren’t exact-match targeting rules”. Keywords are indeed included in the targeting entities but any Google Search practitioner would find themselves lost without match types. How can a keyword be used as a targeting mechanism if it is not controlled through its semantic expansion potential (low in exact, intermediate in phrase, broad in…broad match type). Something inconceivable in Google Search given how the spending and exposure patterns differ from one another. The “context hints” selection in chatGPT ads might feel uncomfortable for any practitioner used to mastering the coverage potential of such keywords. And with limited reporting available yet, customers are probably in the dark when it comes to how and where (in what conversations) their ads are served. In the age of AI-powered marketing, customers are invited to focus on performance first and foremost, irrespective of where ads are shown (I have been in the position to promote this thinking for many years at Google) and it makes total sense but reporting is what has the power to grow customers’ understanding of their audiences.

Same but different

All agree though to say that those context hints are the most challenging and thus most interesting part of the campaign set up. Should advertisers describe the conversations they want to target, the topic, with a prompt, with conversational personas or else? Essentially, advertisers are asked to define the conversations they want to be part of and it’s where the most innovative approaches, new skills and tooling will need to be built in the advertising ecosystem. Products and services are currently defined “at face-value”, with little distance between what they are and how they are named. Tomorrow, customers will need to think about how people talk and converse about their services and products. It’s a very positive evolution as it will force all marketers to put themselves even more in the shoes of prospective customers, in their minds and aspirations, imagining the specific words and sentences they use to describe what they are looking for, especially during the discovery phase of their journeys. Or even what they don’t know they are looking for. 

I run an AI vision strategy workshop called Narra using corporate fictions as brainstorming artifacts to align teams around a shared vision. I tried Google search ads to advertise this offer and there was barely any traffic on the exact terms describing the offer such as “corporate fiction”, “AI strategy” being too broad and too expensive for my very tiny budget. Everyone in the industry is looking for the sweet spot of not-too-broad-not-too-specific search terms to balance cost and volume. With chatGPT ads that I am experimenting with, the challenge is very different as I have to put myself in the shoes of the customers and think about the questions a business leader might ask chatGPT. “How can I foster a collective AI vision in my company?” is an example of a prompt I would love Narra to be displayed alongside. Being really close to this business idea, I can testify how hard it is to describe customer aspirations and how far it is from buying “AI strategy workshop” as a keyword. You could even say that this is a product management exercise of falling in love with the problem, not “your” solution. Dave Dugan, OpenAI Head of Global Ads Solutions, in this video, says that “people come to chatGPT with a job to be done”, a central concept in product management.

So the shift is major as it implies several things: moving away from describing objects as they appear to how people might talk about them, moving away from how things are named to how they could be referred to as, moving away from how, you as a company, want to describe your products or services to how people might describe them in their own words. We could even call this shift a decentering effort. In psychological terms, “decentering refers to the cognitive ability to shift your perspective away from your own immediate, egocentric point of view to consider multiple aspects of a situation or the viewpoint of others”. I’m probably going too far as we’re not talking about psychology but digital marketing yet finding the right conversations to target calls for a decentering effort on the advertisers sides. Will I say that empathy is the path to chatGPT success? No I won’t but I’m tempted to. So all in all, defining those context hints is much more ambiguous, much more uncertain than keyword targeting but potentially much richer as a well described intent allows for a much better matching and potentially much better performance, an early feedback advertisers share already (source). 

Necessarily different

Now advertisers will want to know what conversation they have appeared against. In a typical SEM fashion, they’ll need a search term report (a Google ads report showing what queries were matched to the bought keywords). There are high chances that this is already in the works given how critical it is. What will OpenAI display there? Grouped themed conversations? Keywords? Topics? Will it be useful for advertisers to include, exclude, bid higher for certain conversation groups? Or will they share nothing where digital marketing KPIs and conversion rates will be the only valid indicators of campaign performance? Speaking of indicators, new metrics could emerge, and I’m stealing great ideas from this article: conversation continuation rate: whether people keep chatting after an ad appears; semantic distance score: how closely the ad matches the conversation’s meaning; intent progression: whether subsequent questions move closer to a decision; and query refinement: whether users start asking more precise questions. Together, these could help assess not just whether an ad gets clicked, but whether it fits the conversation and helps someone move forward.

Totally new?

In traditional SEM fashion, keywords can be grouped by their popularity. Generic terms generate a lot of traffic when a very long tail of specific searches generate few searches that are less expensive in the auction. Given how specific each conversation might be in LLMs, will this pyramid flatten, making more conversations specific and long tail? Asad Awan, a leader in the OpenAI ads organization, gives the example in this video of his preference for instant vegan ramen, a dish he enjoys. This dish is an example of a "weird concept" and hard to imagine wanting without prior knowledge. A good AI product can help users discover these specific niche audiences and products they didn't even know existed, a clear opportunity for SMBs?

Matching intent and content, be it through Google Search Ads or chatGPT will get better and better over time as technology progresses. But the introduction of ads in large language models usher in a brand new era of innovations where the core pivot lies in the shift from intent expressed as searches to intent expressed as conversations.