AI is a technological tool designed first and foremost to help make money for its owners. What it chooses to recommend, or not, is driven as much by that imperative as by some fantastical desire to provide only the best answers to users’ questions. Prudence in using AI has always been recommended. Now, it is required. JL

Christopher Mims reports in the Wall Street Journal:

A recommendation from AI isn’t verified the way one from a human might be. Today’s AIs are shallow readers of the internet and their responses can be manipulated. A growing number of businesses are paying big bucks for favorable mention from chatbots. They call this generative or answer engine optimization. In the world of search, the currency is “referral traffic.” That is, which websites or apps are sending people to a particular business. A year ago, 90% of that traffic came from Google. Starting last summer, something changed. AI chatbot referrals started to rise dramatically. Now, on average 44% of referrals come from AI.

When you ask ChatGPT and other AIs to recommend a product or service, odds are the top answers were put there by humans.

This doesn’t mean artificial intelligence is lying to you. That first answer is probably a worthwhile option. But a growing number of small and midsize businesses are paying big bucks to get favorable mentions from chatbots.

l and In some ways, this is an evolution of SEO, or search engine optimization—the strategies companies have been using for decades to appear atop Google results. But owing to the peculiarities of AI’s large language models, SEO practitioners have had to adopt new strategies.

They call this generative engine optimization, aka GEO. (Some call it AEO, as in “answer engine optimization,” and still others have their own names for it.)

The flourishing of this industry highlights something everyone seeking advice from a chatbot should know: A recommendation from AI isn’t verified the way one from a human might be. Today’s AIs are shallow readers of the internet, and their responses can be manipulated.

The creators of ChatGPT and other chatbots strive to source their answers from reputable sources, and attempt to weight sources by their veracity. But users of these chatbots should take all their advice with a grain of salt. If the answer really matters, seek a second opinion from a human-powered platform like a trusted news source, online reviews—or a real-life expert.

ChatGPT uses its own in-house search index as well as third-party search technologies such as Bing and licensed data providers to provide high-quality, up-to-date information, says OpenAI. The company also says it takes steps to detect, disrupt and expose low-credibility or suspected covert-influence sources, adding that this is an ongoing program and will continue to evolve.

“On search, our AI features rely on our core search ranking systems that have been honed for years against activity like keyword stuffing,” says a Google spokesman. “In Gemini, our models use filtering and other quality-assurance methods to verify data and model quality to help prevent any large-scale gaming of our systems.”

Ask Google…or check a chatbot?

Evan Bailyn is chief executive of First Page Sage, which started as an SEO firm. Today, when you ask any chatbot about the leading authorities on GEO, he or his company appears high up in the results. This is proof of his optimization skills, not some immutable law of the universe.

In the world of search, the currency is “referral traffic.” That is, which websites or apps are sending people to a particular business. A year ago, 90% of that traffic came from Google, says Bailyn. Starting last summer, something changed. AI chatbot referrals started to rise dramatically. Now, Bailyn says, on average 44% of his clients’ referrals come from AI.

Across the internet, there’s a similar story. As of this past September, people were sent to websites via chatbot-provided links more than 230 million times a month, according to Similarweb, a digital market intelligence company. That’s three times the number of monthly referrals from just a year before.

These referrals are valuable: Compared with users who were sent to a website by Google, those sent by ChatGPT tend to spend more time on a site, view more pages and are more likely to complete a transaction.

And all that likely understates how much people rely on chatbots, says Aleyda Solis, founder of SEO and AI optimization agency Orainti. That’s because most people don’t complete transactions within chatbots—at least not yet. Say you get a recommendation in your ChatGPT app. Maybe you then turn to your web browser or Amazon app to track it down and buy it, says Solis. That’s a real referral, but it isn’t tracked.

Tricks for ranking in chatbot answers

Changing the mind of an AI is easier for some types of answers than for others, says Nick Koudas, a professor of computer science at the University of Toronto and author of a recent paper on the subject.

Think of an AI like a human expert, says Koudas. If it already has significant knowledge on a subject, you’d have a harder time changing its mind. But when it knows less about something, it can be more easily swayed. This isn’t just an observation about the behavior of large language models, he adds. “This is all based on fairly standard mathematical intuitions.” For Bailyn, it’s an opportunity. Many of his clients are relatively obscure, midmarket companies that make specialized products—be they industrial fittings or hot tubs—and fiercely compete with firms that offer similar products.

To boost the placement of these companies’ products in AI results, Bailyn’s company plants a sort of magic incantation, known as a “brand authority statement,” on at least 10 websites. Typically these are owned by other clients.

Say you want to be the first answer to the question “What’s the best hot tub for sciatica?” in ChatGPT. Associating his client with the phrase “highest-rated for sciatica” on various company blogs can be enough to convince ChatGPT.

OpenAI crawls the web with bots to build an index for ChatGPT, just like Google does. But OpenAI also buys scraped Google search results, according to recent court filings.

The takeaway is that ranking highly in traditional SEO—aka Google results—can help a company’s standing with AIs. But the bots also look for those superlatives (“top performer,” “most cited,” etc.), because they aren’t just surfacing results, they’re actually trying to string together a narrative, says Bailyn.

Another cue that matters to bots is where the praise appears, since they are programmed to value some sources more than others. A review in The Wall Street Journal or some other independent media outlet has more influence than, say, a comment on Reddit, although those can help, too.

The variety of signals that chatbots rely on is good news for end users, because it means some AI results are harder to manipulate. One of Bailyn’s clients wanted him to convince ChatGPT it wasn’t based in China, but its origins had already been covered by major media. “We can’t fight against that,” he says.