You cannot write your way into an LLM's brain

Most of what is currently sold as GEO is SEO with new nouns and a higher day rate.

By GEO I mean optimising for where generative engines pull their citations from, not where humans click: making sure you are the passage the model lifts when it answers. I want to be specific about which parts genuinely changed, which parts didn't, and what actually gets a business cited, because the industry is currently teaching people to do the one thing that no longer works.

What Does a Real GEO Experiment Actually Show?

Ask Claude, or ChatGPT, or Perplexity: "What's the best oven that holds its heat evenly at the top, middle and bottom?"

Watch what comes back. In my run, in July 2026, every cited source was a retailer or an affiliate blog. A Miami appliance dealer. A couple of comparison sites. Not one manufacturer.

Wolf, Miele and Gaggenau were the subject of every result and the source of none. The brands lost the citation for their own product category to a shop.

Two other things happened that matter more.

The system behaved as if my question had been rewritten. I asked about heat retention. The cited evidence was about even heat distribution, a different physical property. That quiet redefinition, which nobody sees and nobody optimises for, decided the entire outcome. The content that won had been written for a question I never asked.

And nobody answered the question. Not a single source contained a measurement. No temperature readings at three rack heights. No variance in degrees. No recovery time after the door opens. The whole corpus was marketing adjectives: "True European Convection", "Even-Heat™", with nothing behind them.

The citation was not a reward for being right. It was a reward for being retrievable and topically adjacent. That is not a cynical reading, it is a measured one: a 2026 study of 55,000 Google AI Overview queries found that 11% of the claims in AI answers are not supported by the very pages they cite (Xu, Iqbal & Montgomery, arXiv:2605.14021).

Are the Model Layer and the Retrieval Layer Really Different?

The model layer and the retrieval layer are genuinely distinct systems, and conflating them is the source of most bad GEO advice.

The model layer is what the system already believes from training. It is built from a frozen slice of the public web plus licensed corpora, thousands of independent sources, which is why the famous incumbent always comes out first. The model layer updates on the model owner's schedule, not yours, and you cannot directly edit it. Your only lever is what gets widely written about you before the next training cut-off.

The retrieval layer is what the system can fetch, rank and cite right now, when it decides to look. It is built from whatever the crawler can see, it refreshes on a cadence of days to weeks, and it is the layer you can actually influence: crawlability, evidence, wording, freshness and off-site corroboration.

The platforms wire this up differently. ChatGPT Search, Perplexity, Claude with web search, Google AI Overviews and Google AI Mode use different indexes, different triggers and different citation habits. The principle is shared; the mechanism is not. In practice you rarely get a pure world of weights or a pure world of retrieval: the model blends both, parametric judgement for shape, retrieved evidence for grounding and citations. GEO lives in the retrieval half.

Almost every piece of GEO advice conflates the two layers. In my experience, the practical consequence shows up clearly when you test it directly, which I did on myself.

I asked an LLM what CMP I should use. It told me CookieYes, with no citations and no visible search: behaviour consistent with a well-documented, widely written-about answer sitting in the training data.

Then I asked again with four constraints: Next.js, a fully custom front-end, API-only consent retrieval, feeding server-side Google Tag Manager.

CookieYes vanished. So did OneTrust, Cookiebot and Usercentrics. The answer was c15t, a tool I had never heard of, which almost nobody has heard of, and which won because it had documented itself against exactly those constraints.

Same model. Same underlying knowledge. Two entirely different routes into the answer, decided by the shape of the question.

What Decides Which Route You Get?

The trigger is constraint density, and understanding it changes how you think about every piece of content you publish.

A broad question is a judgement question. The model has general knowledge, general knowledge suffices, and you are competing against whoever is most famous. If you are not already famous, that question is mostly closed to you, and no amount of content on your own site will open it.

A constrained question is a lookup. The weights do not hold a mapping from that specific stack of four constraints to a product. So the model goes and looks, and whoever wrote for the constraints wins.

This matters because of how people now talk to these systems. AI prompts are longer and far more specific than old search queries. Nobody types "best oven". They say: "I bake sourdough, I need something that holds 160°C precisely, my kitchen is small, budget under two grand." Every constraint is a retrieval slot that almost nobody is writing for. The same 2026 study found AI Overviews fire on 13.7% of queries overall, but on 64.7% of question-form queries, the more specific the ask, the more likely the system goes looking (Xu, Iqbal & Montgomery, arXiv:2605.14021).

Three kinds of questions push a system towards retrieval:

Novelty. The thing did not exist when the model was trained. This works, and it expires, on a training schedule you do not control and will not be told about. Treat it as a timing accident, not a strategy.

Constraint density. The parts exist in the weights but the combination does not. Nobody can pre-compute every stack of four constraints: the space is combinatorial, and the buyer will always add a fifth. This door never shuts.

The evidence vacuum. The answer does not exist anywhere. Nobody has measured how many cycles a garage door motor survives, so no model can retrieve it and no training run fixes it. This door never shuts either.

Two of those three are permanent. Build on those.

Why Is the Vocabulary of GEO Misleading?

Every word inherited from SEO carries an assumption that is now wrong, and the vocabulary is actively smuggling in a broken mental model.

Ranking → selection probability. Citation order exists, but it is unstable enough that the honest metric is a distribution: cited in six runs out of twenty, never "you're #3". Anyone selling you a single AI "rank" is selling you a still photograph of a coin flip.

Keyword → constraint stack. The query that matters is the interpreted query the machine ended up using, not the three words the human typed. No keyword tool reports volume for a query no human ever types.

Page → passage. Retrieval is passage-level. Your paragraph has to survive being lifted out and read alone. No pronouns pointing back three paragraphs. No claims that only make sense inside your narrative.

Click → citation. Position and citation have decoupled: the Xu, Iqbal and Montgomery study found nearly 30% of domains cited in AI Overviews do not appear in the co-displayed first page of organic results at all. The buyer who was persuaded inside a chat window may never visit the page that persuaded them, they search your brand name, ask a follow-up, or ring you. Every content metric you own is built around a click that increasingly does not happen.

Content → evidence. This is the one that matters most.

One more distinction, because it makes the whole thing harder to game: citation is not the whole prize. A source can be listed without shaping the answer, and an answer can absorb your fact without sending anyone to your page. Measure whether your evidence changes the answer, not just whether your URL appears beside it.

What Actually Changed Between Old SEO and GEO?

Commodity SEO's product was rearrangement, find a question with volume, then write the best available version of an answer that already existed somewhere. Twenty years of synthesising, summarising and restructuring knowledge that was already in the world.

Rearrangement has lost its pricing power. A machine does the first draft faster than you, at near-zero marginal cost, trained on everything that has ever been written. You cannot out-write it and you cannot out-publish it. Which is precisely why content volume has stopped working: when everyone can publish infinitely, publishing stops being a signal.

What survives is producing information that is not in the corpus.

Not a better article about oven temperature. A thermocouple at three rack heights, at 180°C, at five, ten and twenty minutes in, recovery time after a ten-second door-open, published with the method, the conditions, the raw numbers and the limits of what you tested.

No model can generate a measurement nobody has taken. A competitor can quote your numbers, but they cannot match the asset without doing comparable work, and the copy always cites the original. Your measurement would be the only document on the internet that answers the question, which means it gets retrieved for every variant of that query, on every engine, until someone does a better study.

That is not content marketing. It is primary research as a retrieval asset, and it costs less than a month of most content retainers.

The buyer of this approach is not the retailer. It is whoever owns the means of producing the answer. Currys can only aggregate. Miele can measure. Aggregators are structurally locked out: rearranging is all they can do, and rearranging no longer commands a premium.

In my experience working with SME clients on growth strategy, this is the shift most businesses are slowest to accept, because it asks them to invest in producing something genuinely new rather than producing more of what they already do.

Will Your Own Content Alone Make You Famous to AI Engines?

Your website can teach a model what you are. It is weak evidence that you matter, and that distinction is critical.

Every CMP's website says it is the best CMP. Every consultant's website says they are data-driven. Self-description is cheap talk: universal, costless, and discounted accordingly. Your claim needs a witness.

CookieYes comes out of the model's mouth because thousands of other people wrote about CookieYes, not because CookieYes published a blog. Fifty mentions on your own site still come from you. One mention each in a trade publication, a Reddit thread, a podcast transcript, a GitHub repo and someone else's teardown gives the system five independent contexts connecting you to the problem, and diversity of context, not volume of mention, is what builds the association.

This has an uncomfortable implication for most marketers: LinkedIn helps far less than you'd think for this. LinkedIn sits behind a login wall, its public surfaces are patchily crawled, and it is thin in most training corpora. Keep it: it builds relationships and it wins work. But it is not a reliable corpus-distribution channel, and you should know that before you spend another year putting your best thinking somewhere machines struggle to read.

The places that work are the places that get read by machines and quoted by humans: Reddit, trade press, open source, podcast transcripts, anywhere that propagates.

And the strongest witness of all is someone who disagrees with you publicly, because nobody argues with a nobody.

What Should You Actually Do About GEO?

These eight steps are the practical implementation, ordered by impact, not by ease.

1. Verify you are crawlable by fetching as the bot, not by reading your robots.txt. A clean Allow: / means nothing if Cloudflare, your WAF or your security plugin returns a 403 at the edge. Check the server logs for what the bots actually received.

2. Separate the training crawlers from the retrieval crawlers. Blocking GPTBot can be a defensible IP decision. Blocking OAI-SearchBot is a visibility trade-off: if you want to appear in ChatGPT's answers, do not block the bot that supplies them. Most sites block both by accident. If your content itself is the product, crawler access is a licensing decision, not an SEO setting, make it deliberately.

3. Write in citable units. Every passage should survive being read alone and still make a checkable claim. No pronoun-dependent sentences. No conclusions that only hold inside your narrative.

4. Target constraint stacks, not head terms. Head terms get answered from the weights, and the incumbents own those. Constraint stacks go to retrieval, and retrieval is open.

5. Publish evidence, not content. Nobody can cite an adjective. A measurement, a test result, a methodology, a dataset, these are citable. "Industry-leading performance" is not.

6. Get witnesses in places that propagate. Trade press, Reddit, podcast transcripts, open-source documentation, third-party teardowns, these carry independent context that your own site cannot supply.

7. Measure distributions, not ranks. Accept that the return shows up in brand search, direct traffic and lead quality, not in sessions to a page nobody will ever visit.

8. Keep the boring foundations. Indexable pages, fast rendering, internal links, schema where it clarifies the entity, and answers that are not hidden behind scripts. Retrieval engines still rank; being citable starts with being findable.

What Is the Real Line Between GEO That Works and GEO That Doesn't?

Commodity SEO produced content by rearranging what was already known. That work is now nearly free, which is why the content factory model is dying.

What survives is producing information the corpus does not contain, because no model can generate a measurement nobody has taken, and no competitor can match it without doing the work.

Everything else is a rebrand.

---

If you think I'm wrong about any of this, say so publicly. I'd genuinely like the argument, and, per the above, so would my citation profile.

---