Your next customer will ask an LLM, not Google

A buyer needs an accounting tool that handles multi-entity consolidation. A year ago they typed that into a search box, scanned ten blue links, opened four, and three of those four were yours, your homepage, your comparison page, your blog post engineered for exactly that query. Today they type the same need into a chat box. They get back three named recommendations and a paragraph on why. They pick one. They never opened a single page. They never saw your site at all.

You weren’t beaten on that deal. You were never in the room.

That is the shift hiding inside every “AI is changing search” headline, and almost nobody has drawn the operational conclusion. The conclusion is blunt: your next customer will ask a model for the shortlist, and the model decides whether you’re on it. This post is about what that does to everything you’ve ever written, and what you write instead.

The funnel lost its middle

Here’s the old shape, and it’s worth saying plainly because it’s the thing that just broke.

Someone has a need. They search. They land on pages. They read, they compare, they form a shortlist, and somewhere on one of your pages a button says book a demo and they click it. Every step happened on a surface you could see, measure, and influence. The search engine sent traffic. The page caught it. The funnel had a middle, and the middle was made of clicks on your property.

The new shape deletes that middle.

Someone has a need. They ask a model. The model reads the open web, synthesizes an answer, and hands back a recommendation with reasons. The reading happened, somewhere, sometime, a crawler ingested the open web, but it did not happen on your page, in your session, with your analytics watching. The buyer reads a paragraph the model wrote, not a page you wrote. If your name is in that paragraph, you’re a finalist. If it isn’t, there was no page for you to lose them on, because they were never going to see a page.

The bottleneck never disappears. It just moves. For twenty years the bottleneck was ranking, being one of the ten links worth a click. The new bottleneck is citation, being one of the three names worth repeating. And the two are not the same skill.

The naive move: keep doing SEO, harder

So the obvious response is to do what worked, but more of it. Find the queries buyers ask. Stuff the pages with the keywords. Build the backlinks. Climb the ranking. If the model reads the open web, then ranking high means the model reads you first, right?

It feels right. It’s wrong in a specific, expensive way.

A search engine was a librarian. You optimized to be the book it pulled off the shelf and handed to the reader. The reader still opened the book. Your job was to be found, and a keyword was a findability signal, a flag a crawler could match against a query. The whole craft of SEO assumed a human would arrive on your page and the page would do the rest of the work.

A model is not a librarian. It is a summarizer who has already read every book and answers from memory, in its own words, out loud. It doesn’t hand the reader your page. It tells the reader what your page said, or what it thinks your page said, or what three of your competitors’ pages said more clearly. Keyword density doesn’t move it, because it isn’t matching strings to a query. It’s deciding which claims are clear enough, specific enough, and well-supported enough to repeat with its name on the line.

So you can win the old game perfectly and lose the new one completely. You can rank first for the query and still not get named in the answer, because ranking made you findable, and the model doesn’t need to find you. It needs to quote you. A page built to be found by a crawler and a page built to be quoted by a model are different documents, and the difference is the whole post.

Our move: write the sentence the model will steal

Here’s the reframe that changes what you write. Stop asking how do you rank for this query. Start asking what is the exact sentence you want a model to say about you, and have you written that sentence somewhere it can lift it whole?

Because that’s what the model is doing. When it composes a recommendation, it isn’t inventing prose from nothing. It’s assembling claims it found stated clearly, attributing the ones it trusts, and hedging or dropping the ones it can’t pin down. A vague page gives it nothing to lift, so it lifts from someone clearer. A page that states one precise, checkable, attributable claim hands the model a finished sentence, and the model, asked to be helpful and specific, takes the finished sentence.

Think about the difference between these two ways of saying the same thing.

The first: We’re a powerful, robust, industry-leading platform that streamlines your financial operations with cutting-edge automation. A model reads that and has nothing. There’s no claim in it, nothing specific, nothing checkable, nothing a buyer could act on. It’s adjectives wearing a sentence. When the model is asked which tool handles multi-entity consolidation, this page contributes exactly zero, because it didn’t say.

The second: This tool consolidates financials across multiple legal entities in one close, including automatic intercompany elimination, for teams running more than one company on a single books. A model reads that and has everything, a capability, a boundary, a who-it’s-for. When a buyer asks the consolidation question, this is a sentence the model can repeat almost verbatim, because it’s already an answer. It said what it does, for whom, with what limit.

The lesson generalizes into one line, and it’s the line worth keeping: the citation is the new backlink, so write claims a model can lift verbatim, not keywords a crawler can rank. A backlink was a vote that moved your ranking. A citation is a vote that puts you in the answer. You used to earn votes by being linked. Now you earn them by being quotable, by writing prose so specific that a model deciding what to say finds your sentence is already the thing worth saying.

What “quotable” actually means

It’s tempting to stop at write clear claims and treat the rest as polish. It isn’t polish. There are three properties that make a sentence liftable, and a page can have prose people love and still fail all three.

The first is specificity over adjective. A model cannot repeat powerful, repeating it would make the model sound like a brochure, and a model tuned to be helpful avoids sounding like a brochure. It can repeat handles consolidation across legal entities in one close, because that’s a fact, and facts are what a helpful answer is made of. Every adjective you delete and replace with the concrete thing it was hiding is a sentence you just made liftable.

The second is a claim with edges. A model trusts a claim more when the claim admits what it doesn’t cover. Best for teams under fifty people; not built for enterprise procurement is more quotable than built for everyone, because the bounded claim is the one a careful answer can stand behind. The page that says who it’s not for is the page a model is willing to recommend, because it’s the page that won’t make the model wrong.

The third is attributability, a claim stated in your own voice, on your own surface, in a form a model can tie back to you. The model is deciding whose name goes on the recommendation. It’s more willing to attach yours to a claim you stated plainly than to a vague impression assembled from scraps. You’re not optimizing to be found. You’re making it easy to be credited.

None of this is a trick. That’s the part that should be reassuring. The old game rewarded a kind of writing that was secretly for crawlers, keyword-dense, thin, built to be matched and not read. The new game rewards writing that’s secretly for humans: specific, honest about its limits, clear enough to repeat. The thing a model wants to quote is, almost exactly, the thing a careful buyer wanted to read all along.

The turn: a company that can’t be summarized won’t be recommended

Here’s the part that isn’t about content marketing.

A model recommending your company is doing, at scale and out loud, what every customer has always done quietly: forming a one-sentence version of who you are and what you’re for, and repeating it to someone who asked. The friend who says oh, you should use them, they’re the ones who handle multi-entity stuff cleanly is running the same loop the model runs. The model just does it for thousands of buyers at once, instantly, from whatever you’ve made public. Word of mouth got a crawler.

Which means the question underneath all of this isn’t how do we do GEO. It’s a harder, older one: can your company be said in one true, specific sentence? If the answer is no, if the clearest thing about you is a wall of adjectives, if even your own team would summarize you three different ways, then the model has nothing to lift, and it will recommend the competitor who can be said cleanly. The vague company doesn’t lose because the model is unfair. It loses because vagueness was never recommendable, and now the thing doing the recommending reads everything and forgets nothing.

That’s the work. Not gaming a new algorithm, the algorithms will shift, and chasing them is the trap that made the old web thin. The work is becoming a company with claims sharp enough to survive being summarized by a machine that has read your competitors too. Write the true sentence. Make it specific. Give it edges. Put it where it can be quoted. Then it travels, through a model, through a buyer, through a recommendation you never saw being made.

---

That’s a lot of what we think about at Apollo Space, not just building agents that do the work, but building a company that reads clearly enough to be recommended by the agents your customers now ask first. The buyer who never visits your page is not lost. They’re just one good, true, liftable sentence away. The only question is whether you’ve written it yet.