Many still believe AI optimization is an oxymoron. You can’t optimize something as complicated and inherently unpredictable as a neural network. Even if it could be done, what AI chatbot would grant outside companies database access like that? That’s how they’re competing - on the proprietary quality of their data and algorithms. And why bother? AI chatbot training databases change infrequently. Even if you succeeded, it might take months before you saw results.
Don’t tell that to Ed Sussman, a New York City based specialist in reputation management. He thinks AI optimization is the next frontier of the business he’s been in for 15 years. And he’s spent the last two years with some of the top minds in AI building a software company, Citate.ai to prove it. Creepy as it sounds, he says that what AI chatbots spit out about you could someday become as big an arbiter of truth as the news media or social media. He thinks AI optimization will likely become as big as the $70-plus billion search engine optimization industry.
Sussman says his tech is based on the fact that AI chatbots don’t just rely on their training data for answers but also crawl the web furiously for as much current content as websites will allow them to gather. Change the diet of current information AI chatbots can crawl, he says, and you can quickly impact the kinds of answers they generate. That then can eventually change their more permanent training database.
He’s only been operating a few months, but he already has about half a dozen clients. Walter de Broweur at Stanford, who is also the chief scientist at Sharecare and a cofounder of Snowcrash, is his science advisor. Gil Alterovitz at Harvard Medical School is his AI advisor. And Michael Carlon, former senior applied AI engineer at A12, the incubator of the Allen Institute, is his chief AI engineer.
He says you don’t need special arrangements with the AI chatbots to pull this off. AI chatbots may give one user the same or similar answers. But he says their answers to the same question across millions of users are much more variable. So by replicating the process of many users asking that same question dozens of times, he says you can begin to understand how each AI chatbot thinks about you, your company, or your cause. Armed with that information, you can then suggest ways that clients can alter that information diet, he says, until they start generating the answers they want to see.
Understanding online reputations and finding ways to influence them is all Sussman has been doing since 2009. Through his two other companies, Buzzr and White Hat Wiki, he helps customers set up approvable Wikipedia pages, helps them develop content for their websites, and ensures that search engines see all that in the way his clients want. Partisan editing of Wikipedia pages along with clever search engine marketing can be ground zero of sophisticated attack campaigns. Sussman helps make sure his clients’ story is getting told and that their reputation isn’t getting attacked. That’s how I met him. Two years ago he helped me build a Wikipedia page for someone in my family.
It’s too soon to say whether companies like Citate.ai are long term businesses, or just interesting features that get added into much larger platforms. Profound, a similar and newer AI optimization company, just raised $3.5 million from a handful of top VCs. And Sussman, whose business is already profitable, says he’s been told he can raise six times that when he wants. But it’s also possible that this service just becomes a new feature that big marketing firms begin to offer or even a tool the AI chatbots supply, akin to Google Analytics. After all, it’s not even clear yet what AI chatbots will be allowed to crawl long term. That’s the subject of myriad lawsuits including a massive suit by the New York Times against OpenAI and Microsoft.
But just the fact that AI optimizers now exist is yet another data point suggesting we are at an important inflection point, where the foundations of online information for the past 25 years - search, Wikipedia, news, user-generated content, and social media - give way to new technologies and applications we are just beginning to understand.
Sussman, who has built a career in the old ways and as one of the early adopters of the new waves, seemed like a perfect person to talk to not just about his new company but about where the new information economy is headed. The following Q&A is drawn from two interviews I conducted - one this week and another at the beginning of June.
Q: Why did you start Citate.ai?
A: For the last 20 years, there's been a (search engine) optimization industry. How do I get to the top of Google? How do I get to number one (in search results)? And there's been this cat-and-mouse game between Google and the optimization firms. The same thing is true of AI. How are you going to get AI to tell your story if you're a nonprofit that cares about gun violence regulation or you‘re a company that cares about selling coffee makers?
If you take the long view, AI is going to write history and/or rewrite history. It’s going to be, maybe, a bigger arbiter of what the truth is than the news media, than social media. It's going to become so pervasive that people are going to (constantly) turn to it. It's going to tell a story, and that story might be about your brand, it might be about your country, it might be about you as a person.
There are going to be people who say, “Oh, don't try to influence them (the AI chatbots) at all.” And that's just nonsense. I'm just supposed to let history write itself without me? I'm not supposed to write my webpage being aware of what AI is looking for. I mean you have to if you want to be part of this new world. What we do is basically about measuring and looking at stories that the AI is telling the world in a way that's very scientific, and evidence-based. We're going to publish studies that this actually works - that we're not just making up gobbly gook.
Q- How does the tech work?
When you ask an AI chatbot (for example) is Starbucks a good place to work? And then, when you zero in on (start asking it) whether they have good health insurance or what the hours are like, you can start to get a very accurate analysis as to whether the story that those answers are telling about your company is positive or negative. We can detect bias in AI responses. We do accuracy detection using semantic matching against our own corpus, our own training data.
But the key here is you can't just ask once. You have to ask (the same question) maybe 50 times before you can figure out what's happening (what kind of answers the chatbots are generating) on any particular day. If you sample enough the responses will fall into a distribution (a graphable range of answers) And so we spent twelve months building an analytic suite, which also has optimization tools. This obviously could be misused severely. So we have a strict code of ethics. We'll only help people who are organizations that are not going to abuse it to put forward bias, inaccuracies, that sort of thing.
Q: I could see how the AI chatbots might notice a firm asking the same question dozens of times and try to keep you from doing that. Has that been a problem?
A: We've managed to find third-party vendors, who I won’t name, who are providing us the volume of data that we need for all of the major chatbot platforms. The third-party provider is under strict instructions that we are only interested in data that is free to the public and we are not going behind any paywalls.
Q: How can your customers be sure of the reliability of your third-party data?
A: One of the things the third-party data providers give us is the full text and all the links, And we share that with the client. So if, for example, there might be 50 different replies in a day to a particular prompt, in addition to our analytics, the client gets to see all 50 replies. If you want to go into the details on a reply-by-reply level, you can. In some cases, we're even providing screenshots in addition to the actual replies and data.
So they get our analysis, our meta-analysis of what's happening over a week, a month, three months, and they can see all the different trends. And at the same time, if they want to dig in further, they can look at the actual AI output … they can look at each of the answers.
Q: What does AI optimization mean for the giant search engine optimization business as well as for widely used information sources like Wikipedia?
A: Search engine optimization people who basically are going to be out of a job unless they figure out how to do this new thing. The impact on Wikipedia will take longer. Right now Wikipedia is viewed as a super authority by these chatbots. If you want one of the fastest ways to influence AI, get the (client’s) Wikipedia page changed.
But yes, in the long term, (Wikipedia will be less important) Wikipedia can write about six or 10 million topics. AI can produce essentially an article about unlimited billions of topics for whatever you want. So Wikipedia is going to become like Encyclopedia Britannica. It's just going to become a small, less important project. It's not where you’re going to turn, because even if it's better and more in-depth, it's not as comprehensive. So people's behavior will change when they look for answers. They'll get those answers from AI, not from Wikipedia.
Q: Do you worry about what happens if this technology gets into the wrong hands?
A: Yes. There is no doubt that there's going to be an information war to try to influence and control AI, on the political front, the commercial front, and on the social landscape. There are people who are going to be doing it (manipulating AI) in a malicious way. So I'm going to try to do it in a positive way.
Like SEO, this will be a cat & mouse dynamic, with both the LLM and the feeders trying to maintain the upper hand. I've never felt good about even white hat SEO. With LLM training data in the zettabytes, it seems unfathomable to me how the feeders are going to bias the LLM to give the answers their clients want. Of course, as soon as (if?) LLMs start to ingest their own synthetic data, this game is shut down forever afterwards.