The page Google indexed and the model misread
Meta description: When a model misreads a page, diagnose the actual answer and the public information environment first — then intervene and inspect what changed.
Miklos Roth
The flagship page in these stories is a point of pride. It ranks where it should, it converts, the brand team signed off on every word. Then someone checks what a model says about the company’s offer, and the summary is wrong — a secondary service described as the main one, an old positioning line repeated as current fact, a capability attributed that the company does not have. The technical team checks the markup and finds nothing broken. They are right: nothing is broken. The page was indexed correctly and understood wrongly, and those are different events.
When an AI assistant presents an incomplete or misleading account of a company, the immediate reaction is often to write another page telling the model what it should say. I prefer to diagnose first. What, exactly, did the assistant say? Which alternatives did it mention? Does the error concern a product, price, service area, professional qualification, or broader reputation? A vague complaint that the AI does not like a brand is not yet a workable brief.
What do teams assume is shared between the index and the model?
The assumption I encounter is that reading is reading: if Google indexed the page correctly, then the information is “in the system,” and any model drawing on that system will get it right. This was a reasonable assumption when retrieval meant returning the page. Models do not return the page; they extract from it, weight it against everything else they hold, and compress. What survives compression is not what the brand team would choose. It is what the page makes most extractable — which is often a slogan, a stray phrase, a legacy description repeated for years.
Where do the two readings split on a real page?
I look backward from the visible result. The assistant’s answer is the observation; the public information environment is the field I can inspect. I examine the website’s technical accessibility, the content available about the company, how its entities are described, the competitor pages visible in traditional search, and the off-site publications or references associated with those competitors.
I have described the method as regression-like reasoning. That phrase is important: it is an analogy to looking for the factors that might explain an observed outcome, not a claim that I have fitted a statistically validated model to a proprietary AI ranking system. My aim is practical inference under uncertainty.
In the cases I have examined, the split is consistent in kind. The index reads structure: titles, headings, the hierarchy the SEO team labored over. The model reads text: sentences it can lift and reuse, claims it can verify against other sources, phrases that appear in multiple places. A page can be perfectly structured for the first reading and treacherous for the second — the careful nuance in paragraph four loses to the punchy, slightly wrong sentence in the hero. The team optimized the skeleton. The model read the skin.
Why do technical fixes miss an interpretation error?
The false explanation is technical: a schema fix, a crawl issue, a rendering problem. I have watched technical audits run clean while the misreading persisted, because the misreading was never in the markup. It was in the language — what the page says plainly versus carefully, what it repeats versus what it says once. Technical fixes are attempted first because they are fixable. An interpretation error requires rewriting what the company says about itself, and that is a harder meeting to schedule.
What pattern marks distortion rather than absence?
Absence is simple: the company is not cited. Distortion is stranger and, in my experience, worse: the company is cited, confidently, for the wrong thing. I have seen a firm described as a vendor of a product it discontinued, a consultancy filed under software, a premium offer summarized in budget language — all stated in the model’s calm, neutral voice. Distortion does not trigger the alarms absence triggers, because the company appears to be present. It is present as a version of itself it would not recognize.
Sometimes the intervention is on the site: an accurate clarification, a better structured service page, or a useful answer to a customer question. Sometimes it is outside the site: improving the business’s public representation, seeking appropriate editorial coverage, or correcting an outdated listing. Not every external mention has equal credibility. A large quantity of low-quality references may create noise rather than improve understanding.
After the change, I would revisit the target question and related questions rather than evaluate one isolated answer, retaining the before-and-after wording, the date, and the test conditions. Even if the brand begins appearing, attribution should be modest.
[AWAITING MIKLOS — scene: second scene]
What question exposes the split?
The teams in these stories monitor rankings and audit markup, and neither instrument touches the split, because the split lives between the page and the model’s account of it. There is a question that makes the split visible — it involves comparing what the page says with what is said about the page — and I have noticed it is almost never in the audit template.
The discipline is not mysterious: observe carefully, compare credible evidence, propose a mechanism, intervene, and inspect what changed. I cannot guarantee that an assistant will repeat a corrected answer. I can help a company improve the accuracy and availability of the information from which people and systems may form an answer. Which source the model is trusting — that is the question to locate first.
PR excerpt (80–120 words, Miklos’s own words): When an AI assistant presents an incomplete or misleading account of a company, the immediate reaction is often to write another page telling the model what it should say. I prefer to diagnose first: what exactly did the assistant say, which alternatives did it mention, and does the error concern a product, price, service area, or broader reputation? A vague complaint that the AI does not like a brand is not yet a workable brief. The answer is the observation; the public information environment is the field to inspect.
By Miklos Roth — AI marketing consultant, Budapest. More: Miklos Roth — AI marketing consultant.
Neighbors: 13, 18, 44.Article 16 — Misread AI search ecosystem
id: 16
status: integrated
slug: /insights/16-misread-ai-search-ecosystem
primary_query: how AI changed search for brands
miklos_slots_open: false
Search developed a second physics
Meta description: If AI search disappeared tomorrow, technical quality, useful content, visual communication and credible external presence would remain.
The growth reviews I sit in on this topic have a familiar shape. The organic channel is down, or flat, or “undergoing transition.” Two camps form. One declares SEO dead and proposes reallocating to paid. The other declares that nothing fundamental has changed and proposes patience. Both camps cite data. I have watched this debate recur across companies, and I have come to believe both camps are wrong in the same way: they are arguing about the fate of a surface, while the buyer has moved to a different one.
I sometimes ask a deliberately extreme question: what would matter if every AI search interface vanished tomorrow? My answer takes me back to the work I began doing roughly twenty years ago. A business still needs a technically functioning website, content that explains its value, and credible ways for people to discover and trust it. The packaging of online discovery may change radically, but these foundations do not become irrelevant simply because a new interface attracts attention.
What did the old search system reward that this one routes differently?
The old system rewarded position: relevance, authority, and technical hygiene, expressed as a ranking on a page of links. The buyer did the assembly — scanning, clicking, comparing, synthesizing. In the new system, the engine does the assembly. The buyer receives a synthesized answer with citations, and the clicking is optional. Everything the old system rewarded still exists — the links are still there, below the answer — but the routing changed: the buyer’s attention now stops at the synthesis. A company can win the old system completely and barely participate in the buyer’s actual experience.
Why are both “SEO is dead” and “nothing changed” useful to the wrong people?
The foundations, as I see them: what the website communicates when a visitor arrives — design is not separate from communication; content that makes products, services, expertise, limitations, and customer questions clear; technical quality, so search systems can discover important pages and the site behaves reliably; and a public footprint beyond the company’s own domain — legitimate signals that the company exists, does what it claims, and participates meaningfully in its market.
These pieces support one another. An attractive site without findable information does not solve discovery. A technically excellent site without an intelligible offer may bring little commercial value. The value comes from treating these elements as a system and judging the system by the customer’s eventual experience and the business result.
What I notice is the politics of the two positions. “SEO is dead” is useful to whoever wants the budget moved — the paid team, the new-agency pitch, the vendor selling the replacement. “Nothing changed” is useful to whoever owns the current arrangement — the incumbent agency, the internal team whose playbook is the old physics. Both positions produce confident decks. Neither deck describes what the buyer now actually meets, because neither camp’s budget depends on describing it accurately.
What metric still gets managed after it stops describing the surface?
Read more: Miklos Roth.
In these companies, the metrics outlive the surface they measured. Rankings are still reported monthly. Share of search is still tracked. Click-through curves are still modeled. Each metric was built for the list-of-links world, and each is technically accurate — and together they describe a system the buyer increasingly does not visit. I have watched dashboards stay green through a collapse in the company’s presence inside answers, because no metric on the dashboard could see the answer layer. The reporting was not lying. It was looking elsewhere.
What pattern shows a team operating on the old physics?
The tells are procedural. The content briefs still target a keyword and a position. Success is still a ranking screenshot. Cannibalization is still debated as a problem of pages competing with pages, not of passages competing with passages. I have learned to listen for the moment in a review when someone asks how the company appears inside the answers themselves — and to notice how often the question is treated as exotic, premature, or assigned to a tool pilot that reports next quarter.
The mistake that frustrates me most is the decision to avoid the difficult foundations because a quicker marketing activity looks easier to buy and explain. I describe the experience through a football analogy. The person responsible for the foundations can feel like the goalkeeper. When the defense works and the goalkeeper makes the expected save, few people celebrate. When the team concedes a goal, everyone notices. Technical and strategic SEO work has a similar asymmetry: problems become visible when something fails, but the effort that prevented failures may remain outside the headline numbers.
[AWAITING MIKLOS — scene: second scene]
What question resets the review?
The review cannot reset itself, because its agenda, its metrics, and its budget line all assume the old surface. The reset has to come from a question that makes the room look at the buyer rather than the channel — and in my experience, the question is simple to state and uncomfortable to answer, because answering it honestly invalidates much of the dashboard on the table.
My advice to a business owner is to ask whether the online presence would still be worth visiting, understanding, and recommending without the AI layer. If the answer is no, that is where the improvement should begin.
PR excerpt (80–120 words, Miklos’s own words): What would matter if every AI search interface vanished tomorrow? A business still needs a technically functioning website, content that explains its value, and credible ways for people to discover and trust it. The packaging of online discovery may change radically, but these foundations do not become irrelevant because a new interface attracts attention. I do not see AI visibility as permission to abandon old disciplines — it is a reason to apply them with more care and to observe new behavior.