Marketing teams have spent two years treating ChatGPT like another ranking algorithm to reverse-engineer. Every week, someone is promising a new seven-step process to win AI citations or lock down ChatGPT search recommendations.
There’s one major problem with all of that. You are optimizing against a system that is still under construction and operated by companies whose actual business is not search.
That’s not a diss towards the AI companies or their tech. Rather, it’s a reminder that we’re dealing with marketing channels in their toddler-era.
The tell is in how they fix things
Consider what happened recently inside ChatGPT’s search retrieval process.
When you ask a question that requires current information, the model can break your prompt into background searches and assemble an answer from what comes back. SEO veteran Lily Ray aggregated what several independent researchers have measuredabout how those searches have changed.
In roughly one model generation, use of thesite:operator went from a fraction of a percent of background searches to somewhere between a quarter and two-thirds of them, depending on whose collection method you trust most. Retrievals per answer roughly doubled. The number of unique domains cited in responses went down.
Product pages and official sources gained share while listicles, roundups and comparison pages lost it. Over the last two years, that latter group has been the darling of AEO and GEO specialists.
The model also started feeding more searches to government domains, established review platforms, recognizable brands, Reddit and other sources it already has reason to trust. Then in a week or two, Reddit itself saw a plummeting reference count from ChatGPT. This haphazard and unfinished approach seems aimed at letting the LLM bypass trust evaluations for websites.
There are some obvious problems with that approach. Researchers have documented “Chat” constructing searches against domains the brand does not own, including parked domains available for purchase. A system that reliably understooda parked domain. Google is smarter than that 99% of the time
LLMs? Still getting search potty training done. Spending heavily to optimize for each patched strategy assumes it will survive the next one.
Why this was predictable
I wrote on LinkedIn earlier this year that several thingsI predicted three years agowere holding up: LLM platforms would prove easy to manipulate in search, they would run search as cheaply as possible and eventually they would have to learn to do many of the same things Google does with its own index, algorithms and anti-spam systems.
I also expected the worst spam problems on free tiers, with paid tiers performing better, particularly where the platforms were borrowing from Google and Bing. And I was right! Better retrieval appeared first in premium reasoning models and only recently reached the free default.
Google has spent more than two decades and enormous amounts of cash figuring out how to rank pages decently in an adversarial open web. OpenAI is now encountering many of the same problems with ChatGPT search and if it wants to catch up quickly, will borrow some of the same solutions.
Meanwhile, much of the SEO industry continues to treat those fixes as new marketing playbooks.
Search is a feature, not the business
Google had strong incentives to make search understandable to publishers and advertisers because search was the business. Ads were the moneymaker but the core search product was the lure that got folks to the website.
In return, we got Search Console, published quality guidelines, spam policies and an advertising marketplace. While it was frequently frustrating, it’s been stable enough to plan against.
OpenAI, on the flipside, makes money from subscriptions, API usage and enterprise deployment. Retrieval and citations seem to make ChatGPT more useful for searchers, but there is no equivalent webmaster ecosystem and no guarantee that the retrieval behavior you optimize for today will still exist by the end of the year.
One of the big drivers will be the introduction of ads and their success or failure. Internal projections reported last year put free-user monetization above $1 billion by 2026 and approaching $25 billion by 2029. Wewrote about that trajectory in detailwhen the first AI ad formats surfaced.
If you are allocating a meaningful marketing budget to AI visibility, you should price that instability into the decision.
What survives every version
Build authority.
Be the brand the model knows enough about to search for by name. One study found brands appearing in a model’s initial background query were cited roughly 69% of the time, while pages merely fetched were cited around 2%.
You don’t get there with prompt engineering. Being conversation-worthy does. Years of coverage, reviews, reputation, customer activity and visibility across the web give the model a reason to know who you are before the search starts and also provide cards to play in other marketing channels while LLMs get their act together.
The last two years sold executives a story in which everything they had built was obsolete and a new set of AI tricks would replace it. Now, some of those tricks are already losing value as the platforms change how search works.
Some AI visibility tactics will work. Some may work very well for a period of time. I would still put the larger share of the budget into becoming the company these systems are already looking for at the start of their whole process, whatever it may be.
