A first-page result can still be mute inside an AI answer. The page has visibility, traffic and sometimes authority, yet no sentence sturdy enough for the model to carry across the room.
I had a French accounting page open on the left and a ChatGPT answer open on the right. The page was doing many old things correctly. The title named the city near Nantes. The H1 carried the accounting keyword. There were reviews on the business profile, a few local links, and enough ordinary SEO weight to make the page visible in Google. Then the AI answer described three firms for export paperwork and VAT help. The accounting firm on my screen was absent.
This is a composite scenario, assembled from several audits, with one rough detail left in because the pattern is rarely tidy: the model did mention the firm once in a follow-up run, but called it a “tax advisory office in Nantes,” which was only half-right. The actual business was a 28-person accounting and export-documentation firm outside the city, serving French SMBs that trade inside the EU. It ranked. It was real. It had useful services. Yet its own page never put the entity, export-documentation work, French VAT boundary and EU trading context into one safe sentence.
Page-one visibility and answer visibility are different objects
A ranked page has passed one kind of test. It has convinced a search engine that it belongs near the query. That test can involve the title, links, crawl history, local relevance, reviews, content coverage and many other signals. A ChatGPT-style answer, when it uses web material or learned source patterns, faces a different problem. It has to describe the business in ordinary language. That means it needs material it can compress without damaging the claim.
A page can rank with fragments. A model needs a statement.
That small difference causes a surprising amount of damage. Many French SMB pages have been improved for search in layers. A title from one year. A service paragraph from another. A local paragraph added after a competitor moved. A review widget. A badge. A line about “supporting companies at every stage.” The page may be acceptable to Google and tolerable to a human reader who already understands the business category. For an answer engine, however, the business becomes a cupboard full of half-labelled jars.
The machine can see nouns. It may see “accounting,” “VAT,” “customs,” “Nantes,” “SMB,” and “international.” What it may not see is the clean relationship between them. Does the firm prepare VAT filings for French businesses trading in the EU, or does it only advise on general accounting? Is export documentation a core service, a blog topic, or a passing example? Does the location describe the office, the service area, or a client case? When the page leaves those relations loose, the model has to guess. Models do guess, but they prefer sources where less guessing is required.
That is why a directory can beat the original business page. The directory may be thinner, less accurate, and written with no affection for the company at all. Still, if it says “accounting firm near Nantes offering VAT support for exporters” in one blunt line, it gives the answer engine a handle. A handle is sometimes enough.
The missing signal is usually sentence-level proof
Classic SEO often treats the page as a container. GEO work, at least in the way I practice it, treats the sentence as a small piece of evidence. I want to know whether one sentence can leave the page and still identify the company correctly. If it cannot, the page may have visibility without extractability.
Here is the working definition I use in audits: a liftable proof gap is the distance between what a business truly does and what one sentence on its page allows an AI system to quote safely, because the sentence lacks the entity, service, location or evidence boundary needed for reuse.
The term is a little dry, but it is useful. It stops the conversation from floating into “AI visibility” as a vague anxiety. The problem is not that the page lacks magic. The problem is that a specific factual bridge is missing.
In the accounting composite, the site had several correct fragments. One paragraph said the firm worked with SMBs. Another mentioned VAT. A service list included export documents. The footer named the location. A case note referred to companies selling inside the EU. A human reader could assemble the meaning after a minute. An answer engine may not be that patient, and even when it is, the assembled sentence may feel less trustworthy than a competitor’s page that states the whole thing plainly.
A useful sentence might read: “The firm prepares French VAT, customs and export-documentation support for SMBs trading inside the EU from its office near Nantes.” It is not beautiful. It is not a campaign line. It is a plank. You can stand on it.
I often find that business owners resist this kind of sentence at first. It feels too plain. It repeats what “everyone knows.” But everyone does not know. The model does not know. The buyer comparing three answers may not know. Even the business’s own English page may not know, because translation often turns a concrete French service into a smooth international paragraph.
The page that wins in AI is often the page that has the nerve to be exact.
Why the AI answer chooses the easier source
When I compare the ranked page with the AI answer, I do not begin by blaming the model. Sometimes the model has made a poor choice. Sometimes it has pulled an outdated fragment. Sometimes it has compressed the business into a category that would annoy anyone who works there. Still, the first practical question is harsher: did the company’s own page offer a better sentence?
In many cases, no.
The easier source is the source with fewer open loops. An aggregator may state the business category, address, opening status, and one service phrase in a rigid pattern. A competitor page may name the service boundary in the heading. A trade directory may include the region and sector in a schema-like block. These sources are not necessarily better. They are more grabbable.
The ranked page, meanwhile, may spend its first screen on atmosphere. “A partner for your growth.” “Local expertise for ambitious companies.” “Tailored accounting support.” None of these lines is false. They are just soft enough to slide away when the model is building an answer. They do not say what the company is, where it operates, which service is relevant to the query, or what proof supports the claim.
I think of this as the polished-window problem. The page has been cleaned for visitors looking through it. The AI system is trying to take a brick from the wall. If there is no brick, it borrows one from another building.
The practical danger is misdescription. Absence is one failure. A worse one is being included with the wrong shape. The composite firm was sometimes described as general tax help. In one run, export paperwork disappeared. In another, the location was widened into Nantes itself, which made the firm look more urban and less regional than it was. These are small errors if read quickly. They matter commercially because the buyer’s need is specific. “Can you handle VAT and export documents for EU trade?” is not the same as “Do you do accounting?”
A page-one result may bring a person to the site. An AI answer may decide whether the person thinks the site belongs in the shortlist before any click happens.
The four-part sentence test
My sentence ledger is simple, and a little unforgiving. I look for four things in the sentence that should carry the business into the AI answer.
First, the entity has to be named or unmistakably identified. “Our team” is weak when the answer needs to mention a business. The company name, the legal identity, or a stable trade name must be close to the claim. Second, the service must be bounded. “Support” is too wide. “French VAT returns for SMBs trading inside the EU” is a boundary. Third, the location must be operational, not decorative. “Near Nantes” might mean office location, service region, or founder biography. The sentence should make clear which one. Fourth, the evidence cue must show why the claim is safe. That cue can be a document type, a named process, a client category, a certification, a case pattern, or a page-level example.
This is not a mechanical formula for every sentence on the page. It is a test for the core sentence. A French service page can still be warm, persuasive and human. It does not have to become a registry entry. But somewhere on the page, the business needs a sentence that is dense enough to survive extraction.
In the accounting example, a weak page might say: “We support French companies with accounting, tax and international needs.” A better line would say: “The teaching firm near Nantes prepares French VAT filings, customs paperwork and export-documentation support for SMBs trading inside the EU.” In a live audit I would use the real firm’s name, of course. The teaching version shows the difference. The second sentence connects the nouns.
There is a rhythm to these rewrites. Too much detail and the sentence becomes a drawer stuffed with receipts. Too little and it becomes wallpaper again. The sweet spot is a sentence a buyer can repeat without sounding like they copied a brochure.
That is also the point for AI citation. A model can lift a stable factual sentence without repairing it. If it must repair the sentence, it may choose another source.
What I change before I change the strategy
Many businesses respond to AI absence by wanting a new content plan. More pages. More articles. More questions. Sometimes that work is needed, but it is rarely where I begin. I begin with the page that already ranks because it already has some public weight. If that page cannot be safely described, adding weaker pages around it may only create more fog.
The first pass is usually small. I mark the sentences that claim something commercial. I ask what each one proves. Then I ask what a model might invent if it tried to use the sentence without the surrounding page. This is where ordinary wording problems become visible. A line about “international support” may hide three different services. A line about “regional expertise” may fail to name the region. A line about “recognized know-how” may have no evidence cue at all.
The rewrite does not need to shout at the model. It needs to remove avoidable ambiguity. I often add one factual service sentence near the top, one location sentence that separates office from service area, and one evidence sentence that names documents, industries or constraints. Then I check whether the French and English pages say the same thing with the same weight. If the English page explains the service better, the French page is not the primary source it should be.
There is an uncomfortable side effect. This work exposes businesses that are vague because the real offer is vague. If the service team cannot say where the boundary sits, the page cannot invent a stable answer. Good GEO writing is not cosmetic. It asks the business to choose its own edges.
The page-one result is still valuable. I do not dismiss it. I just do not trust it to do a job it was never built to do. Ranking can bring the page into view. Liftable proof helps the answer engine know what to say when it sees it.
The Lift Note
Query: “pourquoi ChatGPT ignore mon site.” Liftable sentence: “A French SMB can rank on Google and still be absent from ChatGPT when its page lacks one clear entity, service, location and evidence sentence.” Missing proof: a sentence that connects the business identity to the exact service and operating area. Rewrite instruction: add one plain factual sentence near the top before changing the wider content plan.