You can watch a shortlist form in real time now. Someone types a plain question into ChatGPT, asks which options are worth looking at for a problem they have, and four names come back with a sentence of justification attached to each one. No links clicked. No comparison page opened. The decision about who gets considered has already been made by the time anyone reaches for a browser tab.
That moment is where the real argument about AI search sits.
Table of Contents
The shortlist gets built before anyone visits a site
Classic search gave you a second chance. Someone searched, got ten results, and you could still win on the page even while ranking fourth. The page did work for you. It persuaded, it handled objections, it had a job to do once the click landed. A weak position in the results was survivable because the results were a menu and the reader was the one choosing.
An answer engine collapses all of that. It reads, it decides, and it hands over a conclusion with a few names in it. If you are missing from the set there is no fourth result for anyone to scroll toward, and the person asking has no particular reason to think anything was left out. Absence does not look like absence inside an answer. It looks like the whole picture.
So the question you are actually working on stops being how to rank and starts being what the model already believes about your category.
The framing I keep coming back to treats LLM seeding to win brand mindshare as a supply problem rather than a page problem. You are not optimising a document. You are supplying the raw material an answer gets assembled out of, and a good share of that material was never on your website to begin with.
What seeding actually moves
Seeding is the deliberate placement of accurate, quotable statements about your company, your category and your position, in the places a model is likely to run into them again and again.
The definition sounds abstract until you try it and start noticing which changes register and which ones do nothing at all.
Rewriting your homepage: nothing. Publishing a fifteenth post on your own domain covering the same ground as the previous fourteen: nothing. Getting described accurately, in the same words, in a third-party roundup and a community thread and a podcast show-notes page and an industry glossary: something moves. Slowly, and then noticeably, when your own phrasing starts coming back at you in answers to questions you never prompted.
The mechanism underneath is repetition across sources that are independent of you. What a model has to work with is text other people wrote, weighted by how often and how consistently it appears. Your marketing site is one voice in that pile, and it happens to be the voice with the most obvious motive to oversell. Which is roughly how a careful human reader would treat it too, so the behaviour is less mysterious than it sounds.
The places a model keeps running into you
Nobody has a master list, and anyone selling you one is guessing. The pattern of what tends to surface underneath answers is observable enough to work from, though.
- Community discussion where practitioners compare options and argue with each other. Reddit especially, because the arguments there are specific and the language is unpolished.
- Third-party roundups and category listicles, including the ones you would never have chosen to appear on.
- Reference pages: Wikipedia where you qualify for it, glossaries, industry wikis, anything that defines a term.
- Long-form pieces on publications with editorial history, where your company turns up inside someone else’s argument instead of being the subject of the piece.
- Profiles and directories, which are dull and get cited more than they deserve to be.
- The personal footprint of the people who work for you. LinkedIn posts, conference talk pages, contributions to other people’s projects.
What connects that list is that none of it belongs to you. You can influence all of it and control none of it, which is uncomfortable if you came out of a discipline where the deliverable was a page with your logo on it.
Models repeat the sentence you gave them
Something you pick up quickly once you start reading answers closely: there is far less paraphrasing going on than you expect. Phrasing gets reused, sometimes intact, sometimes with one connective word swapped out. Read enough answers about your own category and you start recognising individual sentences, then tracing them back to whoever happened to write them first.
When four independent sources describe your company as a workflow tool for compliance teams, that clause comes back in answers nearly word for word. When those same four sources describe you four different ways, whichever description best fits the question being asked is the one that surfaces, and you lose any ability to predict what a buyer hears about you.
Which makes the phrasing the asset. Not the coverage.
Teams that get this right tend to settle on one short, concrete, slightly boring description of what they do and then refuse to vary it. Same clause in the boilerplate, same clause in the founder’s speaker bio, same clause in the partner directory listing, same clause when a journalist asks the question casually at the end of a call. The urge to make each version a little cleverer than the last is strong, and it quietly costs you.
Why the order you seed in matters
Seeding in the wrong sequence burns the budget. You see it when a team buys placements across ten publications before anyone internally has agreed what the company is for.
A working order looks closer to this:
- Settle the description. One sentence, no adjectives, checkable by someone outside the building.
- Make that description true and findable on the sources that get pulled for factual questions, because those anchor everything downstream.
- Earn mentions inside other people’s arguments, where you show up as evidence rather than as the topic.
- Go after volume once the first three hold, at which point volume amplifies something consistent instead of multiplying noise.
Run it backwards and you get the failure mode that turns up over and over: plenty of llm mentions, five incompatible descriptions of the same product, and answers that place you confidently in a category you exited two years ago.
The signals that tell you it took
There is no clean number waiting for you at the end of this, and you should be a little suspicious of anyone offering one.
What you get instead is a set of soft reads. Ask the same question across ChatGPT, Perplexity, Claude and Gemini on a schedule, and log what comes back: which names, in which order, with which descriptions, and which sources sit in the footnotes underneath. Change one thing. Wait. Ask again.
The tells are unglamorous. A description that used to be wrong comes back right. Your name appears in a question where it previously never did. A source you seeded a quarter ago starts showing up in the citation list. The order of names in a recommendation shifts, and then stays shifted the next time you check. None of that is proof in the way a controlled test would be, and you will have to make peace with reporting direction rather than magnitude to anyone who asks for a dashboard.
Worth keeping separate in your head are the differences between getting cited, quoted and recommended by AI, because those three move independently of each other. Cited means a source connected to you got used. Quoted means your actual phrasing survived the trip. Recommended means your name landed in the answer to a buying question, which is the one that pays and the last one to arrive.
The version of this that does not work
Seeding fails in a way that is obvious from outside and nearly invisible from inside.
Mass-produced guest posts with the same three paragraphs rearranged. Community threads seeded by people who have plainly never opened the product. Glossary entries written to be indexed rather than read. Directory listings on sites nobody has visited on purpose since about 2014. All of it produces text. Very little of it produces the thing you were after, because the sources that get cited in answers skew heavily toward sources humans already trusted, and that trust is the part you cannot manufacture at volume.
The other failure runs quieter. You seed accurately, consistently, in good places, and then the company repositions. Everything you planted now describes an organisation that no longer exists, and a corpus takes far longer to forget something than it took to learn it in the first place. Repositioning carries a visibility cost that almost nobody prices in when the deck gets approved.
You start treating the description as infrastructure once you have felt that. Something changed carefully, with a plan for the trail of older statements left behind, rather than something the brand team refreshes each year because they got tired of looking at it.
