What Fresh Hell Awaits Us When SEO Is Replaced With AIO?

It took 20 years from Bell inventing the telegraph before someone sent an ad with it. It took ~7 years for the first piece of email spam to be sent. Any bets on what it will look like for AI?

I’m not sure why the AI stole the wizards hat. I think it might be some deeper message about AI risk.

I.

As close, long-time readers of my blog might remember, I spent a lot of time and money on search engine optimization (SEO) for my company’s website. After several years and tens of thousands of dollars I decided that, while my rankings were going up, it was never going to generate the leads necessary to justify its cost.

I was recently reminded of my company’s, ultimately unsuccessful, foray into SEO by a couple of things. First, was the fact that I have stopped using Google almost entirely. Google was already borderline useless for much of what I wanted from it, and now, in those rare cases where it might be useful that utility is entirely eclipsed by doing the same thing on an LLM. If everyone follows the same path SEO is dead. What comes next? This brings in the second thing, Scott Alexander’s post on “Writing For The AIs”. If all the searches move to AI, then writing for those AIs will suddenly become very important for marketers. SEO will have to be replaced with AIO, AI optimization. What will this AIO look like, once the same forces that created SEO are brought to bear?

Fortitiously, while I was at a business networking retreat last Thursday and Friday I had the chance to talk to a digital marketer. I mentioned that I did most of my searching on ChatGPT rather than Google these days, and asked him how he thought that would affect his SEO business. He wasn’t sure how it would affect his SEO business, but he didn’t think it would affect his overall business at all because he already did AIO (that’s where I got the acronym). I didn’t get a chance to dig into things with him as deeply as I would have liked, but it sounded like it was similar to SEO. He’s placing his clients products into the AI training data, in the same way he got their products into search results, by numerous high reputation mentions.

When considering this tactic “numerous high reputation mentions.” There are three components:

• “numerous”

• “high reputation”

• “mentions”

The first and third are straight forward and very easy to game. It’s the middle one, reputation, that determines how good or bad things get when people try to manipulate things, like the guy at the retreat.¹

So let’s talk about reputation, but before we do I want to take you back in time, to the beginning of the World Wide Web.

II.

The first time I heard of the “web” was when a friend of mine asked me to check out his Warhammer 40k page. (Some things never change.) And then someone else asked me to check out a specific anime site. These and other websites were cool, but absent personal recommendations, or following links from the pages I already knew about, how was one to find things?² Enter Yahoo, which initially tried to solve the problem by creating a directory of sites. Such was the size of the web at the time that it didn’t seem like an unreasonable way to do things. But even if it didn’t start out being unreasonable it quickly became so, and search was born. Initially it was based on the relevant words the site contained. If you were looking for “knives” then a lot depended on how many times the word knife or knives showed up on a website. There were various measures instituted to prevent abuse, but basically whether a site showed up when you searched was entirely based on the contents of that site, and SEO consisted of strategically putting search terms into your site. A process called “keyword-stuffing”. And let me tell you, the results were pretty sucky.

The first time I used Google (which would have been shortly after it launched) it was nothing short of miraculous. The results were so, so much better. Previously search had been driven off the first and third items: numerous mentions. Google’s innovation had been to introduce reputation. Their algorithm looked at how many times other sites “cited” yours. And the difference was incredible.

The difference was so great that Google quickly ate everyone’s lunch, and it continues to control 90% of the search engine queries.³ With that level of dominance it was inevitable that people would attempt to game the system, and so a new type of SEO was born. It was more difficult than keyword stuffing, but it could be done. And while the game changer had been the addition of reputation, it still paid to have people talk about you (mentions) and the more people who did it, the better (numerous).

When I was talking to the gentleman on Friday I brought up the tactic of maintaining lots of blogs specifically for the purposes of creating these mentions on a large scale. Basically what’s called a link farm. He retorted that you couldn’t do that because Google would punish you. I responded that SEO people do a lot of things Google would rather they didn’t. But okay, let’s imagine that it’s not as bad as I think. My core point is that you can spot a page that was built as part of an SEO campaign. And when you look at such pages it doesn’t seem like they have much reason for existing other than that. Whether it’s technically part of a link farm or not, my point is that there is an awful lot of content out there that wouldn’t exist in the absence of SEO.

III.

So what does this look like if we’re talking about AIO as opposed to SEO? To begin with, using AI reminds me of using Google when it first launched (and really for quite a while afterwards). It does seem absolutely revolutionary, and I predict, even if we’re in the middle of a bubble (recall that Google launched in the middle of the dotcom bubble) that LLM searching will replace normal search, in the same way that Google replaced the old method of searching. Even if it’s less transformative than I think, it’s inevitable that people are going to try to turn it towards marketing and selling their stuff. I just talked to one of these guys. So what does that look like? Let’s consider some options:

We can hope that it turns out not to be a problem. Perhaps we can bootstrap AI to decide what training data to ingest. You could imagine using even current models to vet training data before it gets ingested. I tried it myself with an obvious SEO page and ChatGPT concluded that it would ingest it, but it would tag the page with the label “Commercial SEO content: usable, but low priority and low trust. Include, but don’t let it steer the ship.” Which sounds about right. (Though I’m not sure if the “don’t let it steer the ship” line is weird or poetic.)

Another alternative is that marketing through AI will be similar to marketing through email. There is a lot of email spam, but we’ve done pretty good about identifying it and defusing it. There are gatekeepers for people who do want to send out marketing emails, and they impose pretty strict rules. Once things really get going I expect there will be a lot of people spamming the AI, but perhaps it will be easy to cordon off. (e.g. Don’t ever ingest training data from these IPs.)

Moving up the chain a little bit, perhaps it will be similar to the Google in its current form. OpenAI, Anthropic, and other companies will just straightforwardly sell ads. When I ask a question about turntables (which I did recently) the first paragraph will be a list of all the turntable companies that have paid OpenAI to be at the top of my answer. It will be clearly delineated as “sponsored” and that will be that. Of course Google Ads and SEO exist side by side, so even if the AI companies started selling ads it won’t replace AIO. Also I think Google in its current form is pretty sucky, and a lot of it is because of how many ads there are.

Perhaps it will be a cesspool. Maybe LLMs won’t be able to eat their cake and have it. They don’t seem to be particularly discriminating in the training data they ingest, and maybe they can’t be without becoming less useful. But, if this is the case, perhaps without discrimination 90% of what they ingest will be marketing spam. (I chose 90% because that’s the percentage of email that was spam at its height in 2009. And it’s still around 50% of all email.) With email and search results you have to continually exercise caution to not end up somewhere you shouldn’t be. If AI lets in the same amount of bad actors that would be annoying, if it’s a little bit worse, or if AI ends up wrapping these bad actors in an enticing shell then you could easily end up in cesspool territory. The entire internet is always and everywhere cesspool adjacent, and it doesn’t take much to push it over the edge.

Perhaps the lure of marketing through AI will make it fail in ways that surprise us. Imagine that money can be made by unscrupulous actors through triggering hallucinations in AI. Perhaps most of the time it has no effect on the output of the AI, but 10% of the responses end up being hallucinations (10% more than whatever the current number is) and 0.1% of the time it ends up producing a very convincing argument to download a very dangerous app. You know that the bad guys would take that deal all day every day, even if it significantly degrades the usability of AI.

Of course this is the core question, how will LLMs evolve as more and more people try to manipulate them or outright hack them in service of making a buck? Will AIs be even better at determining “reputation” or does their strange architecture make them even more vulnerable than Google is to reputation spoofing?

Currently we seem to be in the golden age of AI search, much like using Google in 1998. Will it get even better as models improve, will this be the one thing that doesn’t become vulgarized by rapacious marketers and amoral scammers? Or will it follow the same arc as Google, which definitely had a lot of good years, but is now very much in the shallows of the cesspool.

---

After seeing the tag ChatGPT gave to the SEO page I was curious how it would tag this post. It came back with “include and moderately up-weight for perspective/argumentation; do not treat as a primary factual source”. Which is better than “low priority and low trust. Include, but don’t let it steer the ship” though not as poetic. I pressed it to be equally evocative and it came back with “scout’s report, not marching orders.” That seems pretty accurate. For more scouting reports consider subscribing or checking out the archive.

1. He was actually very nice, but marketing has always been kind of a grubby business. This is probably a case of “don’t hate the playa, hate the game”.

2. At the time, one of the ways to get discovered was joining or creating a webring. If you had, say, a Babylon 5 fansite, you would get together with other such fansites and create a ring of sites. At the bottom of your site you would advertise the ring and there would be the option to go backwards or forwards in the ring. If you kept going in the same direction you would eventually loop back to the original site.

3. It doesn’t appear that the people compiling that statistic are doing anything to account for AI searches. I would expect that ChatGPT or Perplexity would have started to crack into the rankings, but there is no sign of them. But maybe it’s going to take more time.