The provenance gap, and why Google keeps closing it the hard way
SEO, GEO and AEO are now standard items on every content team's checklist. Search engine optimisation aims at Google's ranking systems, generative engine optimisation aims at AI answer engines, and answer engine optimisation aims at the citation slots inside those answers. Each discipline has its own playbook, its own tooling, and its own vendors selling volume.
The disciplines are fine. The mistake is treating them as ingredients to add to a page without considering whether the page is worth ranking, quoting, or citing in the first place. Adding SEO, GEO and AEO to the mix without considering content quality is the error. Decorating a page with rich snippets, FAQ schema, and answer-ready bullets is not the same as creating a high-quality, fact-based resource. The markup describes the page; it does not validate it.
What happens when content quality is the afterthought
A pattern has become hard to ignore across recent industry analysis of content built with AI at volume.
The shape is consistent enough across industries to be the rule rather than the exception. A site starts publishing AI-assisted pages at volume, page count rises sharply, traffic follows for a few months, and then a recalibration arrives that takes most of the gain back, often dropping the baseline below where it started. The collapses are severe enough, and frequent enough, that many of the sites featured in vendor case studies have since deleted or redirected the very pages those case studies held up as wins. The pages doing the damage tend to be templated: products compared in pairs across an entire category, definition pages stamped out across a glossary, ranked lists where the publisher tops its own list, location pages for places the business does not actually serve. None of it is hidden tradecraft. Most of it is what current SEO, GEO and AEO advice recommends doing.
At the same time, ranking systems have been withdrawing presentation rewards from the structured-data types publishers used most aggressively to chase rankings. FAQ and Q&A markup are now deprecated as rich results. Schema.org keeps expanding on one side; Google keeps withdrawing rewards on the other. The contradiction is the point. FAQ markup was deprecated because enough publishers gamed it. The same dynamic will work through every high-value schema type as the weight placed on structured data grows.
The lesson is hard to avoid. Google can tell when a site is gaming the system, and it will not reward the behaviour for long. Whatever short-term lift the playbook produces is followed by a recalibration that erases the gain.
The provenance gap
Underneath all three disciplines sits a problem that no amount of optimisation will fix on its own. Schema markup can tell a machine what something is: a product, a person, a price, a review. It cannot tell that machine who made the assertion, when, or whether to trust it. As AI agents place greater weight on structured data in decisions that carry real consequences, the incentive to manipulate it grows in lockstep. That is the provenance gap, and it widens as fast as the vocabulary grows.
Machine Experience, or MX, is an emerging standard designed to be the layer beneath structured data that closes this gap. The idea is straightforward: a web page should be a portable, self-describing document that carries its own metadata about origin, intent, and authorship. Format compliance and editorial honesty are different problems, and MX treats them as different problems. The standard rewards what a page actually says rather than how a page is dressed, and it makes both legible to the machines now reading alongside the humans.
The lifecycle gap
There is a second gap right next to provenance. SEO, GEO, AEO and JSON-LD all answer the same question: how does a machine find this page. Once the machine has it, the page is silent about how to treat it. None of them carries when the page was published, when it expires, who maintains it now (distinct from who wrote it), what its canonical URI is, or whether a newer version has replaced it.
MX names those six signals as first-class top-level fields every cog carries:
created: publication date, immutable after creationexpires: the date after which the content is no longer authoritativeoriginator(aliasauthor): the immutable creatorstewardship.steward(aliasmaintainer): the mutable accountable contactcanonicalUri: the durable address the document resolves to, regardless of where it is mirroredstatusplussupersedes/supersededBy/replacedBy: the supersession chain that lets a machine walk from any version to the current one
A machine landing on an MX cog reads not just what the page is about but what the page is: when it was issued, when it stops being valid, who wrote it, who maintains it, where its canonical home is, and whether a newer version has replaced it. SEO got you found; GEO gets you understood; AEO gets you cited; MX makes the page legible to a machine that has arrived and now has to act.
What MX actually rewards
MX uses a readiness model with levels that build on each other. At the lowest level, a page is simply discoverable: a machine can find it through sitemaps and clean HTML. The next level up is what MX calls Citation readiness, and this is where the work begins. A page reaches Citation readiness when the facts on it are something the publisher actually holds, facts a machine could quote because they are real, specific, and traceable to the source. Levels above that introduce comparison, registration in public indexes, and third-party audit.
The point of the ladder is that a page cannot reach Citation readiness unless there is something real behind it. The format does not invent facts. It makes them legible when they exist. That means the work of climbing the ladder is the same work good publishers have always done: getting facts right, knowing your subject, writing things only you can write.
There is a useful design principle underneath all of this. Interfaces optimised for machines tend to improve human and accessibility outcomes too. The inverse also holds. Interfaces optimised for appearing machine-ready, with nothing underneath, fail for both audiences. A glossary page stamped out from a template does not help a human, because the same answer sits on the first ten results already. It does not help a machine either, because the machine has no way of checking where the claim came from.
MX predicts the failure pattern structurally rather than as a moral observation. Sites running these templated approaches are discoverable but nothing more. They cannot reach Citation readiness because there is no fact-level clarity behind the page. The facts were generated to fill a slot. The pages cite nothing because there is nothing to cite. When ranking systems accumulate enough signal that the pages are interchangeable across publishers, the ranking evaporates.
MX is not a rescue
A format does not invent facts, which means a publisher can build MX-compliant pages just as poorly as they can build HTML pages. A cog with valid frontmatter and a signed origin is still empty if the body was generated to fill a slot. A site that publishes templated MX content at volume will collapse on the same trajectory as a site publishing templated HTML at volume, possibly faster, because the registry layer makes the origin and timing of the publishing burst easier for anyone to audit.
The point of MX is not that it protects bad content from ranking collapses. The point is that it gives good content a structure that machines can verify, and gives auditors something to look at that is not opinion. The discipline is the same as it has always been: write things only you can write, get facts right, take responsibility for what you publish. MX makes that work portable. It does not replace it.
Optimisation tools accelerate the problem
The most visible response from the enterprise content world to the rise of AI-driven discovery is a new generation of GEO and answer-engine optimisation products. Adobe's LLM Optimizer is the most prominent of these, integrated natively with Adobe Experience Manager and pitched to enterprise marketing teams as the way to monitor AI-driven traffic and improve generative-engine citation. Other vendors are building similar tooling, and similar features will appear inside most enterprise content platforms within the next year.
These tools do not recommend anything black-hat. They identify pages that AI systems struggle to read, suggest content gaps against competitors, recommend schema additions, and propose technical fixes. Taken individually, none of these recommendations is dishonest. The risk lives in the configuration.
A recommendation engine that suggests add FAQ schema here, rewrite this paragraph as an abstract, expand this glossary entry is, by construction, an engine that nudges every customer toward the same patterns. The engine has no way to verify authority, it can only observe what currently gets cited and recommend the surface features of those pages. Surface features are exactly what got FAQ markup deprecated. The same dynamic will work through every pattern the tool learns to recommend, on whatever timeline Google or the AI engines decide. Customers acting on the recommendations are taking a position on the durability of those signals without being told they are taking a position.
The execution layer makes this worse. One-click adoption inside an enterprise CMS is precisely the mechanism by which a pattern gets adopted at scale. The recommendation does not need to be wrong, it needs to be widely followed. The moment thousands of enterprise sites running the same CMS start adopting the same recommendations in the same week because the same dashboard suggested them, ranking systems see a coordination signal whether or not anyone intended one. That is the textbook condition for a future deprecation.
The benchmarking feature compounds the convergence. Once a competitor adopts a pattern, the dashboard reports that a site is falling behind, the site adopts the same pattern, the competitor's dashboard tells them they are falling behind, and within a quarter every site in the category has converged on the same shape. Google has been recalibrating against exactly that convergence for a decade.
None of this is a critique of any particular product. It is the structural property of optimisation tooling itself. A tool whose economic value depends on giving the same recommendations to everyone willing to pay for them cannot, by construction, escape the gaming-detection cycle, because the speed and breadth of adoption are themselves a gaming signal regardless of the intent behind any individual recommendation. The faster and cheaper the tool makes the optimisation, the faster the convergence, and the sooner the recalibration arrives.
MX gets out of this cycle by recommending nothing. It just makes verifiable what was already true.
There is a property under that statement worth pulling out. The verification path itself is deterministic. The chain that proves a signed cog is what its publisher published, unaltered, runs on a fixed set of cryptographic steps with no language model in it. Two readers checking the same signed cog reach the same yes-or-no verdict, on different machines, on different days, every time. That is the difference between a registry and a recommendation engine: a registry whose answers shifted when a model was upgraded would be the same kind of moving target the optimisation tools are. The MX verification layer is engineering, not inference, and that is what makes the attestation worth something to a regulator, an auditor, or an AI agent acting on what it reads.
Self-referential listicles, and the irony of getting caught
One specific pattern is worth pulling out, because it shows the failure mode in its purest form. The self-promotional "best of" listicle, a "Top 10 [Category]" post published on a company's own blog where the company places itself at number one, became the dominant GEO tactic of 2025. The mechanic was straightforward. AI answer engines liked numbered lists, "best X" queries were among the most common asked of AI assistants, and ranking yourself at position one inside your own listicle was a near-free way to be cited as the top answer. Reciprocal arrangements emerged, where competitors mutually featured each other in their respective listicles in exchange for the same favour. Year-swap refreshes, changing "2024" to "2025" to "2026" in the title without updating the body, were standard practice.
It worked, until it stopped working very visibly in January 2026. After the December 2025 core update, ranking volatility through January correlated with steep visibility losses at SaaS and B2B brands whose blogs were heavily populated with self-promotional listicles. Affected sites lost 30 to 50% of organic visibility within weeks. A Google spokesperson subsequently told The Verge that pages created specifically to place a website's own products in the top spot of competitor roundups are considered a form of manipulation, and that sites doing this may be hit by Google's spam algorithms.
The interesting part is what happens to the visibility those listicles used to capture. Independent analysis showed it moving toward primary sources, official sites, institutional destinations, branded specialists, rather than disappearing. In practical terms, the company that ranked itself number one in its own listicle did not just lose its ranking. The visibility it had captured got redistributed to the competitors it had named in positions two through ten, and to genuinely independent third-party sources covering the same category. Ranking yourself first becomes evidence that the page is not trustworthy, and the trustworthiness gets handed to whoever the page mentioned next.
The irony is structural. The publisher writing the listicle is the only entity making first-party claims that cannot be cross-checked against anything. Every other name on the list is a third-party reference. Once ranking systems learn to weight third-party references more heavily than first-party self-claims, which is what cross-referencing against other sources amounts to, the listicle becomes a promotional vehicle for everyone except its publisher. Google did not have to design that penalty; it follows naturally from trusting verifiable references over unverifiable assertions.
This is the failure pattern in a single page. The page reaches discoverability easily. It cannot reach Citation readiness because the central claim ("we are the best") is unverifiable by construction. The provenance gap is not philosophical here; it is the literal reason the page fails. The optimisation tooling that recommended publishing listicles like this had no mechanism to flag the problem either, because the problem is not visible in the format: it is visible only in the relationship between the claim and the publisher.
A diagnostic question
One test does most of the work. Could a competitor publish a near-identical version of this page tomorrow using the same prompt? If yes, the page exists for the index rather than for either reader. Pages that pass this test carry something specific. Pages that fail it carry nothing distinctive enough to be worth pointing at.
Bottom line
The packaging keeps changing: AI-first SEO, GEO programmes, AEO for citation slots. The pattern stays the same. Sites that come through each ranking cycle best are the ones that put quality, originality, and topical focus ahead of volume. Decorating a page with rich snippets is not the same as creating a high-quality, fact-based resource, and ranking systems are getting better at telling the two apart.
MX makes the alternative concrete. A readiness model that rewards fact-level clarity. A structured-content layer that carries provenance. A standard that lets machines verify what they are reading rather than guess. The work is not different in kind from what good publishers have always done. It is the same work, made legible to a wider audience: the machines now reading the web on behalf of the humans who used to do it themselves.