Every few years the search industry invents a new name for itself, and the current crop (LLM SEO, GEO, AEO) arrived with unusual speed and unusual pricing. Brands that watched an AI assistant describe their category without mentioning them are understandably nervous, and nervous budgets attract confident vendors.
Here's the framing that survives contact with the data: the game hasn't changed nearly as much as the referees have. Traditional search ranked pages with one well-studied algorithm. AI assistants pick their sources through retrieval systems with different habits, different biases, and different favorite websites. What those new referees reward, though, looks a lot like what the old one did. LLM SEO, done honestly, is authority building aimed at a new set of judges.
A language model answering from its training data cites nobody. It's paraphrasing compressed memory. Citations only exist when the system decides a question needs fresh or specific information and goes out to retrieve it, pulling a handful of pages and assembling the answer from them.
That's a smaller window than most people assume. Semrush's analysis of 17 months of ChatGPT clickstream data found that live search was enabled on just 34.5 percent of queries as of February 2026, down from 46 percent in late 2024. Roughly two thirds of ChatGPT answers never touch the live web at all. And when assistants do send people somewhere, the destinations concentrate hard: 30 percent of all ChatGPT referral traffic flows to just ten domains, with Wikipedia, Reddit, GitHub, and YouTube near the top.
Each assistant also referees differently. ChatGPT leans on community and reference platforms, Perplexity searches the live web almost every time, and Gemini draws from Google's own index with its own selection quirks. There's no single leaderboard to optimize against anymore, which is exactly why tactics tuned to one platform's current behavior age so badly.
Read those numbers together and the shape of the problem changes. Chasing per-answer visibility tricks targets a shrinking, fragmenting window. What compounds is becoming the kind of source that retrieval systems, and the training corpora behind them, already treat as load-bearing.
The academic side of this field reached the same conclusion early. The Princeton-led study that coined the term generative engine optimization tested nine different content interventions and found that adding quotations, statistics, and cited sources boosted visibility in generative responses by up to 40 percent, while superficial tweaks like keyword stuffing did little or backfired. The winning moves were the ones that made content more concretely evidenced, which is what quality editorial guidance has demanded all along.
The practitioner side agrees, sometimes grudgingly. In Digiday's survey of the GEO agency wave, uSERP's Jeremy Moser put a number on it: a GEO service that doesn't admit AI visibility is 80 percent fundamental SEO is selling snake oil. The genuinely new parts are real but narrow, mostly retrieval plumbing and platform integrations. The rest is the old discipline under a new invoice line.
Strip out the theater and three signals decide most of your citation odds.
The first is retrievability, and it's binary. A brand blocking GPTBot or PerplexityBot in robots.txt (often a leftover from a 2023 legal panic nobody revisited) is invisible no matter what else it does. The same audit should cover Bing indexation, since several assistants retrieve through it, and whether key pages render without heavy JavaScript.
The second is quotability. Retrieval systems lift passages, and the passages they lift state a claim plainly and back it with a number or a named source. A page that buries its point under a warm-up intro and hedged conclusions gives the model nothing to attribute. Restructure the highest-value pages so each section opens with the sentence you'd want quoted, because that is literally the selection mechanism.
The third, and the one that compounds, is being referenced by sources the machines already trust. Assistants assemble answers from a small pool of pages per query, and that pool skews heavily toward established publications rather than brand blogs. Earning links and mentions there is slower than tweaking your own site, which is why it separates durable winners from acronym chasers. It also raised the bar on placement quality. The screening bar at hetneo.link is a useful illustration of where that discipline has landed: a mention only counts if it sits on a page with live rankings and real traffic, because a page nothing retrieves passes nothing along, to either kind of search engine. A hundred links on dead pages now lose to five on pages that assistants actually pull.
The honest expectation setting: AI assistants still send most publishers a sliver of their traffic. Digiday cites Similarweb data showing Reuters and The Guardian each earn less than 1 percent of overall traffic from AI platforms despite being heavily cited. But the visitors who do arrive behave differently. The Washington Post reports subscription conversion rates four to five times higher from AI referrals than from traditional search. Small pipe, unusually good water.
So the sane program for the next year isn't exotic. Unblock the crawlers and verify Bing. Rewrite cornerstone pages so every section leads with its citable claim. Publish original numbers on a schedule, since data is the most reliable citation magnet the research has found. Keep those pages visibly maintained rather than letting them fossilize. And put sustained effort into earning references on the publications retrieval systems already lean on, because that authority transfers across every assistant at once while per-platform tricks expire with each model update.
Measurement deserves the same sobriety. Rank trackers don't see inside ChatGPT, so the practical check is unglamorous: keep a standing list of the fifteen or twenty questions your business should be the answer to, run them through the major assistants monthly, and log who gets cited and from which pages. Over a quarter, that log tells you more about your actual AI visibility than any dashboard currently sold, and it costs an hour a month.
The referees changed. The judges of the judges changed. The way you win in front of them, evidently, did not.