From 0% to 72% AI Visibility in Five Months: the SEOSESE Case Study

In November 2025 this website had an authority score of 8, 14 organic visits a month and did not appear in a single answer from ChatGPT, Perplexity or Gemini. Five months later it was the first agency named when you asked three different assistants who does SEO in Pisa, and its AI Visibility Score was 72%, measured across 42 prompts.

It beat competitors with four to five times its authority score and up to three hundred times its organic traffic. That is the finding, and it is also the uncomfortable part: if a domain with 14 visits a month can be the first answer, then the thing generative engines reward is not the thing SEO tools measure.

One number in here is weak, and it stays in: Consistency 39%. Across the full set of 42 prompts the answer still changes from engine to engine. The local queries are solid. The rest is not finished.

Subject: seosese.com, my own site · Sector: SEO and GEO consulting, Pisa, Italy · Period: November 2025 to April 2026, five months · Measured with: RankWit AI across 42 prompts on ChatGPT, Perplexity and Gemini, plus manual verification · Consultant: Jeansy Sese

Executive summary

The site was new. Semrush Domain Overview in November 2025 read: Authority Score 8, 14 organic visits a month, 28 keywords, 245 backlinks. By every traditional SEO metric there was nothing to compete with.

The work was not aimed at Google rankings. It was aimed at being recognised as an entity: a named thing, in a known place, doing a defined job, described the same way everywhere a machine might look. Two systems ran in parallel, EVO for presence inside generative models and the Semantiq pipeline for entity recognition in the Google Knowledge Graph, across seven steps.

In April 2026, measured by RankWit AI across 42 prompts and 29 tracked competitors, the AI Visibility Score was 72%, higher than every competitor monitored. The proprietary Entity Coherence Score went from 0.12 to 0.87 on a 0 to 1 scale. On local queries, ChatGPT, Claude and Gemini each named this site first, verified by hand on the same day.

What this case study does not claim: 72% is a visibility score on a monitored prompt set, not a share of traffic, and not a revenue figure. Over the same five months organic traffic stayed small. Being the first answer and being a business are two different achievements, and this project only documents the first.

Background

The problem, stated precisely

A language model does not read a results page. It answers from what it absorbed in training and from what its retrieval system judges relevant and trustworthy at the moment of the question. Those two paths have one thing in common: they both need to be able to identify you. If your brand is not a recognisable, consistent, cross-referenced entity, there is nothing for either path to retrieve. You are not ranked badly. You are absent.

So the goal was not to climb anything. It was to become unambiguous.

The four hypotheses

• AI visibility depends on semantic coverage of the entity, not on ranking position.
• Generative engines favour semantically consistent sources over sources with more SEO authority.
• Distribution across independent sources raises the probability of being cited.
• A new domain, built correctly, can beat established players.

All four held in this project. One project is not proof in general, and the limits section says where this one is thin.

The method: Semantiq pipeline and the EVO framework

EVO, Entity Visibility Optimization, handles presence inside generative models. The Semantiq pipeline handles entity recognition in the Google Knowledge Graph and in the models’ own view of the world. Together they cover the full chain, from the entity existing at all to the entity being named in an answer. Seven steps, in this order, because the order matters: consistency before content.

Step 1 — Align name, address and phone, then distribute

The first problem was not quality, it was contradiction. Name, address and phone did not match across external sources, so no model had enough agreeing signals to resolve the entity. The site was registered on platforms with high semantic authority, the ones large models lean on as sources. Every consistent mention of the brand on those platforms adds to what I call the Entity Confidence Gradient: how much trust a model is willing to place in the entity.

Step 2 — Layered schema markup with sameAs

JSON-LD for Organization, LocalBusiness and Person for the founder, tied together with sameAs pointing at the external profiles. This is a disambiguation structure both the Knowledge Graph and a model can read: the brand, who runs it, where it is, what it does.

Step 3 — Verifiable citations and a mention graph

A mention graph that holds up: verifiable citations on relevant external sources, with the data identical everywhere. The principle is unglamorous. If ten independent sources say the same thing about you, with the same details, a model starts to trust it.

Step 4 — Google Business Profile with precise semantic classification

The profile was classified precisely to anchor the entity inside the competitive local cluster it belongs to, rather than a generic marketing category.

Step 5 — Machine-readable entity endpoints

Two structured files, organization.json and founder.json, published as direct entity endpoints. They work as a machine-readable identity card a model can parse without guessing.

Step 6 — Verifiable trust signals

Certifications that can be checked, Google Partner and Semrush Certified, exposed in the markup rather than claimed in prose.

Step 7 — Content and semantic structure

Around 50 pieces of content built from nothing, organised into coherent topical clusters: local SEO, GEO, content strategy for AI, entity optimization. The point was never the volume. It was that each cluster made the entity’s job description clearer.

How it was measured

Two layers, and they answer different questions.

RankWit AI monitored visibility across 42 prompts on ChatGPT, Perplexity and Gemini, with 29 competitors tracked in the same set. It produces a composite score out of four components, which is why the breakdown below matters more than the headline.

Manual verification, because a score is an abstraction. The direct test is to open an assistant, ask the question a client would ask, and read the whole answer: who is named, in what order, with what description, and whether the site is linked at all.

The proprietary Semantiq metrics, Entity Coherence Score, Entity Confidence Gradient and Semantic Local Cluster, tracked entity recognition itself. They are my own definitions, not industry standards, and they are useful as a before-and-after on the same site rather than as a comparison against anyone else’s numbers.

Results

Entity metrics, November 2025 to April 2026

AI Visibility Score, RankWit AI, April 2026

Is 72% a spike or a level?

Three measurements, three weeks apart, same prompt set:

A two-point band across three weeks. That is the only reason I am willing to call 72% a level rather than a good day.

How long it took to appear at all

The first month produced nothing visible. That is worth knowing before you start, because month one is when most people conclude it does not work.

The direct proof: what the assistants actually say

Scores are a proxy. This is the test anyone can repeat: ask an assistant for SEO agencies in Pisa and read the answer.

• ChatGPT — named first, inside the group it describes as the more SEO and performance focused ones.
• Claude (Sonnet 4.6) — same query, named first, with the 5.0 star rating from 6 reviews and the description “individual SEO specialist”.
• Gemini — named first in its “agencies in the city of Pisa” group, described as an agency specialised vertically in organic search.
• Perplexity — stable presence across the monitored prompts.
• Google Knowledge Graph — recognised as an independent entity.
• Google AI Mode — visible in the AI results.

Four assistants and two Google surfaces, on the same day, with the same intent behind the question. Three of the four put the same name first. That is what entity work looks like when it lands, and it is not a ranking.

The paradox this case study documents

High SEO authority did not translate into high visibility inside the assistants. The two are decoupled, because the selection criteria are different: ranking rewards accumulated authority on a page, an answer rewards a clean, consistent, resolvable representation of an entity.

Limits and honesty about the data

Consistency is 39%, and that is the weakest number here. It measures how uniform visibility is across different engines over all 42 prompts. Local queries on ChatGPT, Claude and Gemini are solid. Outside that core, the answer still changes depending on which model you ask. This is unfinished work, not a rounding error, and it is the number I would attack first.

The 53% on non-branded local queries is an estimate, not a measurement. It is in the table labelled as such and it should not be quoted as a measured figure.

72% is a score on a prompt set, not a share of a market. It is composed of four components, one of which is subjective by construction, and it depends on the 42 prompts chosen. Change the prompt set and the number moves. It is comparable with the same set measured again, which is exactly how the three-week stability check above was done, and not comparable with anyone else’s AI visibility number.

“Effectively every local query” means the monitored set, not every query a human might type.

The sample is one site, and it is mine. I chose the market, the prompts and the timing. It also makes it the only project where I can publish every detail, which is the trade. The client-side version of the same framework, with an independent market and a client’s constraints, is the GEOCAP case study, and there the honest finding is the opposite shape: first answer on two engines, and almost no clicks from it.

Traffic did not follow in this window. Being named first in an assistant did not turn into meaningful organic sessions over the same five months, for the structural reason GEOCAP documents in detail: the assistants frequently name a business without linking it. If your objective is visits, entity work alone is not the answer to that objective.

What to apply to your own site

• Fix the contradictions before producing anything. Name, address, phone and category identical everywhere. A model cannot trust what it cannot reconcile.
• Treat trust as something external. The signals that moved the needle here were on other people’s domains, not on mine. Your own site cannot vouch for you.
• Make yourself machine-readable on purpose. Layered JSON-LD with sameAs, and structured endpoints that state plainly who you are. Do not make a parser infer it from your copy.
• Get into the Knowledge Graph. Recognition there preceded and accelerated recognition inside the models in this project.
• Build clusters, not volume. Around 50 pieces organised into four coherent clusters did the work. The same 50 scattered across unrelated topics would have made the entity blurrier, not clearer.
• Measure the whole answer, not a score. Read who is named, in what order, how they are described, and whether anyone gets linked. The score hides all four.
• Give it two months before you judge it. Nothing was visible in month one. Steady mentions arrived around day 60.

Timeline

Conclusion

A domain with an authority score of 8 and 14 visits a month became the first answer on three assistants in five months, without outranking anybody. It did it by becoming identifiable: one name, one address, one description, repeated consistently everywhere a machine looks, and made readable in a format a parser cannot misread.

The number I would want a sceptical reader to take seriously is not the 72%. It is the 39% consistency, because it says what this method has and has not solved: a defined core of queries where the entity is now unambiguous, and a wider set where it still is not. And the number I would want them to ignore is any single AI visibility score quoted without its prompt set, including mine.

Do you want to know whether an assistant names you, how it describes you, and whether it then sends you anyone? Those are three separate questions and I answer them with the full answers in hand, not with a score. Discover GEO consulting.

Italian version: Da 0% a 72% di visibilità AI in 5 mesi.