“There may be now ample proof, collected over the previous few years, that AI techniques are unpredictable and troublesome to manage.” That’s Dario Amodei in January, writing in regards to the expertise his firm sells.
Evaluate with what’s in your LinkedIn timeline this week. Right here’s the script: Schema markup ensures AI engines parse your content material. The primary sentence of each part have to be the reply. Optimize for chunk-level retrieval. There’s a 13% quotation elevate obtainable in case you do X, a 2.8x conversion enchancment in case you do Y.
It’s one of many cleanest patterns going proper now, and the business has elected to not discover. The individuals closest to those techniques are more and more cautious about claims of management. The individuals furthest from it are more and more sure they know the way it works … they’ve cracked it. That gradient runs the improper approach.
What The Folks Who Constructed It Truly Say
Anthropic printed its primary interpretability analysis submit in Might 2024. It opens:
“We principally deal with AI fashions as a black field: one thing goes in and a response comes out, and it’s not clear why the mannequin gave that specific response as a substitute of one other.”
Anthropic, writing about its personal mannequin, two years in the past.
Issues haven’t gotten extra assured since. Neel Nanda, who runs Google DeepMind’s mechanistic interpretability group, gave an interview to 80,000 Hours in September 2025 by which the headline discovering was that essentially the most bold model of mech interp might be lifeless. He doesn’t see a practical world the place the self-discipline delivers “the type of sturdy ensures that some individuals need from interpretability.” Price re-reading.
The individual whose job is to learn AI minds is publicly conceding that the challenge, as initially conceived, gained’t get there.
At NeurIPS 2024, Ilya Sutskever, co-founder of Protected Superintelligence and previously chief scientist at OpenAI, accepted his Take a look at of Time award and used the platform to say one thing the room wasn’t anticipating from him:
“The extra it causes, the extra unpredictable it turns into.”
Sutskever’s profession is basically the scaling speculation with a face on it. Listening to him say the following section produces much less predictable outputs is itself an admission.
Now scroll again to your timeline. The gradient is Dunning-Kruger redrawn at an business scale: Mt. Silly with a pricing web page, and the valley of calibration the place the precise work occurs.
What The Folks Promoting It Truly Say
A practitioner posts a four-pillar framework for “Technical GEO.” A guide ensures inclusion in AI Overviews. An company markets a 13% elevate in quotation chance, derived from information the company itself produced in regards to the company’s personal prescriptions. A extensively shared submit guarantees that sustaining a 300-character paragraph restrict dictates how a vector database chunks your content material. A vendor claims a 78% “share of mannequin.” A senior determine in your inbox describes a 2.8x enchancment in conversion from being cited in SGE.
The vocabulary is deterministic: “ensures,” “ensures,” “dictates,” percentages exact to the decimal, frameworks confidently named. None of it sounds something just like the language the individuals who constructed these techniques use when describing how the techniques behave.
That is the half I preserve getting caught on. The consultants are assured in regards to the techniques they’ve measured in opposition to themselves. Run the identical playbook on a number of shoppers, watch some metric transfer, name it proof. No management teams, no pre-registered hypotheses, no measurement of what the tactic is definitely claimed to vary. That’s the bar an actual check has to clear; every thing else has been affirmation in costume. The issue is the boldness stage, which is improper by an order of magnitude no matter whether or not the underlying tactic does something. The identical mannequin that Anthropic publicly says it can not totally account for is being optimized in opposition to by individuals who confidently declare to know precisely what they’re doing.
Both Anthropic has been suspiciously modest in public, or anyone else is suspiciously sure.
When Someone Exams
On Monday, final week, Ahrefs printed a research by Louise Linehan and Xibeijia Guan with a title that ought to ideally be not possible: We Tracked 1,885 Pages Including Schema. AI Citations Barely Moved.
The methodology is the type of work you’d anticipate to be customary, if the self-discipline cared about requirements. 1,885 pages that added JSON-LD schema between August 2025 and March 2026. 4,000 matched management pages. Quotation modifications measured 30 days earlier than and 30 days after the schema was added, throughout Google AI Overviews, Google AI Mode, and ChatGPT. Distinction-in-differences on the matched teams.
The discovering: No significant uplift in citations on any platform. AI Overviews truly confirmed a small however statistically important decline. The report notes the chances of a spot that giant being likelihood are roughly 1 in 2,500. The schema-makes-LLMs-understand-your-content thesis, examined at scale in opposition to a managed baseline, didn’t survive the check.
That is the empirical affirmation of the technical case I made every week in the past in The Entire Level Was the Mess: that LLMs learn unstructured language, and that schema-and-chunking prescriptions are reasoning about an structure that doesn’t exist. From first rules, two weeks in the past. From managed measurement, final Monday.
It’s price sitting with that. The dominant prescriptive class in the whole GEO playbook has been empirically falsified beneath managed circumstances, by a vendor with a considerable viewers, within the open. And the frameworks preserve promoting.
Then Google Itself Answered
On Might 15, 2026, Google printed official documentation on optimizing for generative AI options in search. The web page mythbusts the GEO prescriptions in writing: llms.txt information aren’t wanted; chunking content material isn’t required; rewriting content material for AI techniques isn’t mandatory; particular schema markup isn’t required; pursuing inauthentic mentions doesn’t assist. The framing is unusually direct for a Google developer web page:
“Many steered ‘hacks’ aren’t efficient or supported by how Google Search truly works.”
Google names Reply Engine Optimization and Generative Engine Optimization by their full phrases and rejects the playbook outright.
That’s the search engine the consultants declare to be optimizing for, telling its personal developer viewers that the optimizations don’t work. From first rules, two weeks in the past. From managed measurement, final Monday. From Google itself, final Friday. Three unbiased sources of the identical reply, all inside a fortnight. All ignored by the individuals promoting the alternative.
The Value Of Asking
That is the place the analysis stops being well mannered.
Assured claims compound on these platforms in a approach that skeptical corrections don’t. The distinction is in who pays. Posting a assured declare prices you nothing. It will get engagement, builds an viewers, generates inbound, makes the slide deck look forward-looking. If it seems to be improper, nothing occurs. By the point anybody notices, everybody’s moved on to the following acronym.
Posting the correction prices you. It picks a struggle. It marks you as a contrarian, or worse, as anyone who doesn’t get it. On LinkedIn, the place most of this occurs, it really works in opposition to your skilled model. The algorithm won’t reward it. The unique poster owns the remark part and may ignore your methodology query whereas participating with the congratulatory replies. Your reply lives in a collapsed thread.
There’s a selected transfer price naming right here. Ask a GEO guide to elucidate, in plain phrases, what their methodology truly does, what mechanism it acts on, what would depend as proof, what would falsify it. The response escalates into jargon. “Vector-space alignment.” “T1 question optimisation.” “Chunk-level semantic retrieval.” Actual phrases from machine-learning analysis, glued into mixtures that sound rigorous and resist plain-language verification. The sample works as a result of it might probably. Asking “what does that really imply” appears naive, and observers with out the precise technical data can’t inform which mixtures are actual and that are improvised on the spot.
Learn the feedback on any high-engagement GEO submit. Fifteen replies in, 12 are agreements or “right here’s one other talent so as to add to your listing.” Two or three supply diplomatically-framed skepticisms: “I’d like to see extra information,” or “the listing is correct, however…” The writer engages substantively with the philosophical objection as a result of pushing again in opposition to “that is too technical” is simple. The methodological objection, that the prescribed abilities produce assured hypothesis with out a measurement layer beneath, will get the politest burial.
What this provides as much as is gaslighting at business scale. The individuals studying the expertise accurately get positioned as those who haven’t caught up; the prescriptions that managed checks simply falsified get bought as forward-looking. GEO has labored out methods to make calibration seem like the deficiency.
A current X experiment captured the dynamic exterior search engine optimisation. Somebody posted a Monet portray and claimed it was AI-generated, asking the replies to elucidate its inferiority to an actual Monet. Lots of responded, confidently cataloging the “AI tells.” Flat brushwork, soulless composition, no cohesion, no soul. They have been analyzing a Monet. The body decided what they noticed.
The unique submit, the place quite a lot of the preliminary replies have now been deleted.
It’s the identical trick. Vocabulary substitutes for substance; framing prompts affirmation bias earlier than any examination begins; the efficiency of research turns into what’s bought fairly than the evaluation itself; “that is X” arrives earlier than anybody checks whether or not it’s. As soon as the body is ready, the evaluation follows.
So the individuals most geared up to push again, the practitioners who’ve truly tried to check issues, the technical SEOs who know what schema does and doesn’t do, those who can spot a fabricated elevate quantity from throughout the room, keep quiet.
The consequence, on the timelines the C-suite reads, is a one-sided market.
The associated fee falls on the individuals who purchase the declare. Shoppers pay for schema audits the Ahrefs research simply falsified. Junior practitioners construct careers on methodologies that gained’t survive a managed check. And the self-discipline burns credibility it can want later, when conventional search displaces additional, and SEOs are anticipated to take a seat in rooms with engineering groups who’ve simply spent two years watching the sphere confidently mis-call the expertise.
Information advances by attempting to disprove your speculation, not verify it. GEO does the alternative, runs research designed to validate what it’s already promoting. If the professionals claiming this experience gained’t even attempt to falsify themselves, who can we anticipate to imagine us?
The Absence Is The Knowledge
Strip the discourse, and what stays is the absence.
A severe technical discipline watches a managed check contradict its dominant prescriptions, and the prescriptions preserve promoting. At that time, asking whether or not the prescriptions are improper stops being the attention-grabbing query. That has been answered. The more durable query is what’s improper with a discipline that watches and doesn’t right.
Identical with the gradient. When the individuals who constructed the techniques hedge and the individuals optimizing for these techniques assure, asking who’s proper stops being attention-grabbing. The researches and builders are proper. No person who has labored on inference attribution thinks in any other case. The more durable query is why the sphere lets the ensures journey unchallenged.
The trustworthy reply is that the incentives don’t pull towards correction. Confidence sells in methods warning can’t. The reportable framework wins the funds; the wise evaluation loses. And hedged language doesn’t match on a pricing web page the place a assure suits completely.
None of this wants villains. The marketplace for consideration rewards confidence over calibration, each time.
You may preserve watching the gradient run the improper approach. Or you’ll be able to learn what it truly is: an business standing on Mt. Silly, charging for the view.
Extra Sources:
This submit was initially printed on The Inference.
Featured Picture: Roman Samborskyi/Shutterstock
