When the AI changes without warning, whoever has no foundation disappears

Your brand may have dropped out of the AIs’ answers this week without a single line of your site changing. No SEO dashboard will warn you, because there is nothing to warn about: the page is still live, still indexed, still getting the usual traffic. What changed was on the other side.

Why does this happen?

A new model gets a launch note. The rest does not: the routing between versions, the system prompt, the verification layers, the index that feeds search-grounded answers. None of that is announced, and all of it moves what the model considers trustworthy enough to cite. When the bar rises, whoever was leaning on very little falls first.

And there is a part that shifts without anyone having touched anything.

What fluctuates and what gets built

Language models are stochastic. Two runs of the same prompt, on the same model, on the same day, return different answers — and that is not a defect, it is how the thing works. That is why our battery runs two executions per model in every audit: measuring once is measuring noise and calling it a result.

The commercial consequence is uncomfortable, and we put it in writing. G-SoV (Generative Share of Voice), which is how much your brand shows up in the answers, is a number we report and never promise. Anyone promising first place in ChatGPT is promising what they do not control, and the next run already contradicts them.

So there is nothing to be done? There is. It is just not where everyone is looking.

What does not fluctuate is the foundation. While the surface of the answers moves, the layer where entities are resolved obeys data-engineering rules — and those rules are stable, verifiable and workable. It is the only part of the problem that sits on your side of the table.

Your own voice is your weakest point

There is a widespread idea that, to be cited by the AIs, you just have to publish more on your own channels. We argue the opposite: the fact that an AI knows about a brand only what that brand publishes about itself is the defect, not the solution.

What does the model do when it has to decide whether to cite you? It cross-checks. If your site claims you created a method and no independent source confirms it, what it has is a claim without corroboration — the equivalent of a résumé without a single reference. It is not that it distrusts you. It is that it has no way to tell your claim apart from anyone else’s.

A brand that has only its own voice on the internet goes invisible at exactly the moment the model most needs confirmation: when the question is hard and the answer is going to cite someone.

There is the discomfort. The part of your presence you control on your own is the part that weighs least.

The result is not showing up, it is the entity resolved

Showing up in an answer is a symptom. The result is your entity being resolved: an identifiable node, with coherent structured data and sources that confirm one another.

That is what EPS (Entity Prominence Score) measures, across four pillars.

• 1. The Schema.org markup on your site: responsible for translating to the machine what the page says to the person.
• 2. Presence in Wikidata: an open database many systems use as an entity reference (the Wikidata QID works like your brand’s ID card inside the models).
• 3. The density of your knowledge graph: which maps how many and which authority connections tie you to other recognised entities.
• 4. Entity coherence: an analysis that checks whether what is claimed about you matches identically everywhere.

The difference between that number and G-SoV is the spine of our method. G-SoV depends on a third party’s algorithm and varies on its own. EPS is deterministic and auditable: the Wikidata record either exists or it does not, the markup is either on the site or it is not. One we report. The other we move, and we prove that we moved it.

In the audit, AI-Scan queries 15 AI systems, and the valid ones from each round are declared in the report — because a platform that fails, or that answers only halfway, cannot go into the count as though it had answered in full.

The small contradictions that blind the machine

Entity coherence is not a copy-editor’s fussiness.

An address that appears one way on the site and another way on the Google profile. The company name spelled three different ways. An old phone number in a directory nobody remembers exists. Each of those, on its own, looks like a trifle. Together, they are contradictions the machine reads — and it is not going to investigate which version is the right one. It lowers its confidence in the entire entity and moves on to the next.

How many of those does your brand have today?

Four things concentrate almost all the damage, and are worth checking:

The first is the spelling of the name, which must be identical on every public platform, with no improvised abbreviation. The second involves address, phone and e-mail being the same in the site footer, in the profiles and in your sector’s directories. The third refers to the URLs that serve as references, which must be live or redirected, since a broken link is a dead end for whoever is crawling. Finally, the fourth is the authorship of publications, with each author existing as an entity of their own, tied to the company.

I, Bruno Rosa, of Sumaúma AI Presence, ran this measurement on our own house on 16 August, before any client asked for it. The coherence pillar came out at zero. We had the company name declared three different ways in our own graph, the address abbreviated in one source and spelled out in another, and our Google profile not cited where it should have been. We are fixing it, and the baseline is on the record: if the number goes up, we can prove it went up. If it does not go up, the method is wrong, and we will say that too.

Body of work beats followers, and there is a condition attached

Generic content is not weak content. It is invisible content. What the model cites is what only one source says, and text any competitor could have written gives it no reason to choose you.

This is where the maths flips for whoever has a body of work. For generative AIs, a professional’s authorial work is worth more than a follower count. The model does not count likes: it counts semantic density and consistency tied to a name. That is why someone who wrote twenty books can beat someone who filmed two thousand reels.

Except there is a condition, and it is a hard one. The machine has to be able to read the work. A publisher’s PDF book, a talk on video, an article on a portal nobody crawls — none of that reaches the model as a source. Work the machine cannot read does not weigh.

That is what the Digital Twin exists for: modelling the author’s verbal identity and knowledge core, so that what goes out under their name is theirs, in form and in substance. And it is why we turn down anyone without a body of work. With no work of their own, the twin has no raw material and the delivery comes out generic. We would rather not sell.

If someone asked an artificial intelligence today for a recommendation in your market, is the name that comes out yours?

What gets built to last

We do not know when the next model will change its criteria. Nobody does. The easy path is to chase the tweak of the week with a writing trick and hope it sticks. The path that works is building a foundation that does not depend on guessing: a resolved entity, coherent structured data, sources that confirm one another, and your own body of work in a format the machine reads. That is what we build and maintain in GenPres.

What this text does not settle: how long it takes. The foundation is built in weeks and consolidates over months, and any closed deadline someone gives you about showing up in an answer is a guess dressed as a contract.

If your brand has a legacy, if its people have a body of work, and you suspect none of that is reaching the machines that now run the first screening of your market, that is exactly the work we do at Sumaúma AI Presence.