Is GEO only in our heads, or something real?

Where the numbers come from. Every first-party figure here is from our survey of 150,074 domains, taken in Tranco rank order. 88,484 of them published something machine-readable, and every rate below is counted over those 88,484, never over the whole list, because a host that never answered is not evidence either way. The same checks run as npx agentic-readiness <site>, so anything here can be reproduced on one domain.

Hallucination is not one thing

I was deep in thought about how much of a model hallucination is RLHF-based assumption in the absence of training data or context. Pretraining gives the model the ability to confabulate. Post-training often influences whether it chooses to confabulate rather than say "I don't know."

A base language model is trained to predict plausible continuations. If the evidence needed to answer is absent from its weights or context, there's no fundamental mechanism in next-token prediction that says "stop, evidence unavailable." It still has a probability distribution over tokens, so it generates the most plausible continuation. Hallucinations therefore exist without RLHF. TruthfulQA demonstrated this with pre-RLHF base models such as GPT-3, GPT-2 and GPT-Neo/J: they confidently reproduced misconceptions learned from text, and the larger models were less truthful, not more.

But this is a slightly narrow point of view for web people like us. For us, when we think of hallucination we think of a model telling an untruth. Which the above is really a subset of.

And we don't interact with bare models anyway. We interact with fully fledged ecosystems with so many moving parts: model, system prompt, memory, retrieval, tools, planners, execution loop, verifiers, reward shaping, UI constraints, external state. What fancy people call harnesses. In that environment an untrue answer can arrive for a much wider variety of reasons, and they are not equally anybody's problem.

Why even think about any of this? Well, because if you're to believe GEO is a real discipline, you have to believe you can impact some of these factors in a controlled way as an outsider. Otherwise there'll be nothing to optimise, right? The important distinction is control over the model versus control over the evidence environment the model operates in.

The twelve, sorted by leverage

Twelve ways an agent ends up saying something untrue, ranked by how much a business outside the model can actually do about each one.

Six of them are yours. Wrong or stale retrieval, evidence the model misreads, the wrong tool, the planner's invented assumption, the partial result read as complete, and lost provenance. That's a much bigger surface than conventional GEO, and it is the only part of this list worth selling remediation for. For the other six, report the problem and say plainly whose it is.

Where a fix requires you to operate a machine endpoint, the leverage is conditional on your doing so. Three of the six carry that condition.

Wrong or stale retrieval is the biggest opportunity

RAG exists precisely because supplying current external information reduces dependence on faulty parametric knowledge. Models frequently update an incorrect parametric belief when given correct external evidence, although strong priors can still cause resistance. Retrieval quality matters enormously: wrong, irrelevant and conflicting documents measurably degrade answer correctness.

A business can work on this directly. Suppose models say "Lumar doesn't support JavaScript rendering." You want a canonical URL that states the fact plainly, rather than expecting the model to derive it from 1,800 words of marketing copy.

Does Lumar crawl JavaScript websites?
Yes. Lumar renders pages using Chromium and crawls content generated by client-side JavaScript.

Then make sure that page is crawlable, internally linked, canonical, indexed, current, server-rendered or otherwise trivially extractable, and not contradicted by three old documentation pages.

And here is the part I can measure rather than assert. The retrieval substrate isn't in good shape. Of the 88,484 publishing hosts, 61,587 publish an XML sitemap and 21,448 of them break the format. The single most common fault anywhere in the study is a sitemap sitting in a subfolder that lists URLs outside it: 14,176 sites. Google reads those URLs anyway if you submitted the file through Search Console, which is why so many teams have never seen it break. Hand the same file to anything that isn't Google and the out-of-scope entries are dropped.

The sitemap protocol was agreed in 2005. If the oldest, most boring piece of retrieval plumbing on the web is broken on a third of the sites that bother to publish it, the idea that the evidence environment is already optimised and only model behaviour is left to tune doesn't survive contact with the data.

Basic freshness engineering matters too: Google explicitly uses trustworthy lastmod values as a recrawling signal.

One correction worth making precisely, in an article about precision. For ChatGPT Search, OpenAI says that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links, and that they recommend allowing it. That's narrower than "must be allowed to be eligible": you lose the answers, not the link. GPTBot is the separate one, for content that may be used in training, and OAI-SearchBot is what builds the search index.

It's also a decision very few sites have taken deliberately. 33,626 sites name at least one AI crawler in robots.txt, and only 4,289 of them name OAI-SearchBot at all, about one in 8. A caveat I won't skip: naming a crawler isn't blocking it, and a crawl can't see which way a rule points. What the number shows is how few sites have made any explicit decision about the one bot that governs appearing in ChatGPT answers.

None of this is speculative "LLM SEO." It's improving the evidence retrieval systems have available.

Evidence the model misreads is also yours

This is where I'd invest substantially. Models do worse as irrelevant context increases, and where the evidence sits on the page affects whether a claim gets verified against it. So stop asking "is this information somewhere on our site?" and start asking a harder question.

Can a machine retrieve the smallest sufficient passage that answers the question without inference?

Those are very different standards. The first is satisfied by a sentence buried in the eleventh paragraph of a product page. The second is not.

What you're reducing is the amount of reasoning between evidence and claim. I'd call the metric inference distance: how many unsupported transformations are required to get from the source material to the factual answer. You want that close to zero.

Tool selection is something businesses can influence

This is where the agentic ecosystem gets interesting, and where I have the sharpest numbers. Tool-selection research shows measurable differences in how susceptible models are to misleading tool descriptions and overlapping capabilities. A tool called query, described as "Query the system", makes correct selection impossible. One called salesforce_get_account_by_id that says it is the authoritative source for account records, returns owner, tier and renewal date for one account ID, and is not for opportunities or historical snapshots, makes it easy. You can't force an agent to pick your tool. You can remove the ambiguity from the decision.

And this is where the numbers get embarrassing.

112 of 497 machine-readable API descriptions give an agent nothing stable to name a call by. That's not a theoretical concern about tool ontology. It is the exact failure this section argues about, already present in a fifth of the sites that went to the trouble of publishing a contract at all. And 40 server cards advertise an MCP server without saying where it is.

Don't merely expose an MCP server. Engineer its ontology so an agent can reliably determine when it is the authoritative source.

Unsupported planner assumptions are surprisingly addressable

Consider: "Why is Lumar more expensive than Screaming Frog?" If neither company has published meaningful evidence, an agent will generate something plausible. Possibly complete fiction. A company can anticipate those inference gaps and publish the answer: that they are priced for different jobs, that one licenses a desktop crawler per seat and the other bills on crawl volume and retained history across an estate.

Or, importantly, publish the negative. Negative knowledge is particularly useful. Companies publish enormous amounts of X does Y and very little X does not do Y. False assumptions live precisely in that missing space.

Provenance may be the biggest long-term opportunity

Generative search research has repeatedly found provenance problems. Liu, Zhang and Liang's Evaluating Verifiability in Generative Search Engines audited four engines by human evaluation and found only 51.5% of generated sentences fully supported by their citations, and only 74.5% of citations actually supporting the sentence they were attached to.

That study is worth citing carefully rather than waving at. It is from 2023, and it audited Bing Chat, NeevaAI, perplexity.ai and YouChat. NeevaAI no longer exists. It establishes that the problem was real and severe; it does not describe 2026 systems, and I haven't found an equally rigorous replication since.

For a machine endpoint you control, return facts with their source URL, the date they were true, what kind of basis they rest on, and what they supersede. Not because the model needs every property, but because every downstream agent then has the opportunity to preserve provenance.

And the gap here is close to total. 73 sites publish an A2A agent card and 15 of them are valid; 54 leave out a field the card is required to carry. Of the 191 sites publishing MCP authorization metadata, 121 are valid. The machinery for machine-readable, attributable answers exists, it is barely deployed, and where it is deployed it is usually wrong.

Where I'd be much more sceptical

This is where a lot of GEO marketing is bullshit. There's no good evidence that adding an llms.txt, rewriting a page a certain way, adding some schema, or mentioning a fact five times reliably causes frontier model parametric knowledge to change.

A company doesn't control when the next model trains, its corpus, deduplication, weighting, filtering, synthetic-data pipeline or post-training. Allowing GPTBot means content may be used in training. It absolutely does not mean a factual correction will be absorbed.

Worth noting what the most-adopted intervention actually is: 13,730 sites publish llms.txt, making it the most widely adopted agent-facing file on the web after robots.txt and the sitemap, and the one with the least demonstrated demand behind it.

The GEO literature supports that scepticism. Controlled studies do show that altering page structure and content can increase citation visibility. What is missing is any demonstrated stable, longitudinal, cross-platform causal method for improving organic discoverability. The strongest evidence concerns what happens once content is already available to the generative system. That is an enormous distinction, and it is how I'd communicate it to a customer.

There is a better business category hiding here

Not generative engine optimisation, and not making AI mention your brand more. Something closer to factual integrity for organisations: ask models hundreds of questions about the organisation, and for every wrong claim, classify which of the twelve failure modes produced it. Not a brand score. A statement of whether this particular error is addressable, and why.

Three of four models say Lumar cannot crawl JavaScript. The claim is wrong. There is no canonical page stating it plainly; the fact appears once, in the eleventh paragraph of a product page, beside two older documentation pages that contradict it. That is failure mode 2 and 3. Publish a canonical answer, retire the contradictions, re-measure.

That is empirically testable. After the change, rerun the same experiment. Not once, because these systems are stochastic, but enough times to separate a real shift from sampling noise. Diagnose all twelve, sell remediation for the six where the organisation genuinely has control, and for the rest report the problem and say whose it is. That distinction may be the strongest thing about the product, because it avoids pretending every wrong answer can be fixed with another fucking llms.txt file.

It also has a precedent in how we already report. In the same scan, the one security-impact check the scanner carries, whether a site has published private key material in its public signing-key directory, fired 0 times across 59 publishers. The honest way to report that isn't "the web is secure." It is: nobody in this sample did the dangerous thing, the sample is 59, and by the rule of three that is consistent with a true rate anywhere up to about 5%. A product that reports a null result that carefully is a product whose positive findings you can believe.