A Man, a Machine, and the Vanishing Blue Link
Here is a fact that ought to trouble anyone who has ever paid for a Google ad. When Ethan Smith's clients ask ChatGPT to name the best website builder, the winner is not the company whose link appears first in the citations. The winner is the company that gets named the most times across all of them. Ranking, that old religion, has been quietly replaced by something closer to gossip: whoever is talked about most, wins.
Smith is the CEO of Graphite and, by his own accounting, has been optimizing for search engines since 2007 — back when the sport involved auto-generating a hundred million landing pages of scraped product snippets and waiting for Google to notice. It did notice. It launched an algorithm named Panda, and an entire industry of clever people watched their traffic evaporate overnight. Smith calls that the biggest change he has ever witnessed. This new one — the migration of curious humans from the search box to the chat box — he ranks second. Which is itself the point: everyone else is calling it the apocalypse.
"A lot of people are seeing this as everything is different," Smith says. "Nothing we've done before is going to work. We have to rethink everything." He does not believe this. What has arrived is not a new universe but a new layer — search, plus summarization, plus a few novel inputs. The engineers call it RAG, for retrieval-augmented generation: the model runs a search, gathers citations, and then does the genuinely new thing, which is to blend them into a single confident paragraph. Everything that worked in the old SEO still works. It's just that the old SEO no longer describes the whole board.
You need to get mentioned as many times as possible. Usually the first answer will be the one mentioned the most in the citations.— Ethan Smith
To understand why this matters, consider the two ends of the animal. The head — the popular questions, the "what's the best X" — behaves differently now. In Google, if your blue link showed up first, you won, full stop. In an answer engine, a first-place citation buys you nothing if a rival is name-checked five times and you're name-checked once. The machine is not choosing a winner; it is taking a poll and reading you the average.
The tail, meanwhile, has grown a second life. The average chat query runs to roughly twenty-five words, Smith notes, against Google's spare six. People ask, then follow up, then follow up again, trailing intent behind them like a lit fuse. Whole categories of hyper-specific questions — the ones about whether some obscure payment-processing API integrates with some obscure analytics tool — have never been typed into a search bar in the history of the internet, because search could never support that much specificity. Chat was built for exactly it. And a company that answers a question nobody else has bothered to answer can become, delightfully, the only citation in the room.
The startup that couldn't win Google can win ChatGPT tomorrow
This is the part that upends the usual advice. When early-stage founders came to Smith asking for SEO help, his standard reply was blunt: don't. Not yet. Google rewards domain authority, and domain authority is a slow, expensive fruit — a Series A or B luxury, not a launch-week one. But answer engines have no such patience tax. "You can get mentioned by a citation tomorrow and start showing up immediately," Smith says. A Reddit thread, a YouTube video, a blog post — a brand-new YC company that everyone is suddenly talking about can appear in an answer the same afternoon. The long tail, pronounced dead in the SEO era, has been resurrected in chat, and this time the little companies can reach it.
You can get mentioned by a citation tomorrow. Early stage companies can win. They can win quickly.— Ethan Smith
And the traffic, it turns out, is not merely early — it is better. At Webflow, Smith's team measured a sixfold difference in conversion rate between LLM referrals and ordinary Google search visitors. Sixfold. His theory is that a person arriving from a chat has already been primed by a long, meandering conversation, has narrowed their want to a fine point, and shows up not as a browser but as a buyer. Eight percent of Webflow's signups now come through this channel. It is not the largest channel. Paid still is. But it is, unmistakably, a top one — and, as everyone agrees, growing.
Then there is Reddit, which occupies a strange and central role in this story. It is among the most-cited sources inside the large language models, and it is the single thing Smith's clients ask about most. The instinct of every growth hustler is identical and identically doomed: spin up a hundred fake accounts, auto-post, upvote yourself, manufacture a trust score, and declare your product the best in every thread. It does not work. The accounts get banned; the comments get deleted; the community — which is very, very good at this — swats them like flies.
What works is almost insultingly simple. At Webflow, a couple of real employees find the right thread, say who they are, say where they work, and offer one genuinely useful piece of information. Five such comments can do more than ten thousand fabricated ones. The machines lean on Reddit precisely because humans police it, and the moment it stopped being trustworthy, the engineers tuning the citation algorithms would stop citing it. Authenticity, in other words, is not a virtue here. It is the exploit.
Smith is careful to locate all of this on the correct side of the machine. None of it touches the training data — the frozen, next-word-predicting core that knows the capital of California is Sacramento. It lives in the retrieval layer, the live search the model runs before it answers. That is the surface a mortal can move. Influencing the core model is, he suspects, nearly impossible and profoundly not worth your afternoon.