AI search visibility: how we measured which passages of transgenia.org an answer engine would cite, cut passages with no entity from 77% to 38%, and discovered new keywords

Your site can say true things that no answer engine could attribute to you. At Transgenia we found this out by measuring the AI search visibility of our own website: 53 of 69 passages never named the company, and the pricing FAQ answered "it depends." Here is what we measured, what we changed, what production shows today, and what we deliberately left out.

The real problem: 53 of 69 passages did not say who was speaking

The number that opened the case. On September 20, 2026 we evaluated 69 passages from transgenia.org: 56 FAQ answers and 13 entity passages (Organization.description, two meta descriptions, the opening paragraphs of About Us and Services, and the blockquote in llms.txt). 53 of 69 (77%) did not name Transgenia; an engine citing them would have no one to attribute the fact to.

The most visible was the least citable. The 13 entity passages averaged a composite of 0.52 versus 0.74 for the FAQs. Service heroes scored 0.11–0.36 with marketing_fluff of 0.78–0.91, 22 passages contained no verifiable fact at all, and 13 were evasive: they answered "it depends" or "contact us" without criteria.

The question a prospect actually types. For "how much does it cost to implement Odoo in a private clinic in Mexico," no passage reached P(cited) 0.70; the highest came in at 0.31. For "who implements Odoo in Mexico for SMEs," the best-ranked passage (0.92) was an English-language blockquote in llms.txt, which an engine would almost never show to a Spanish-speaking user.

AEO lesson: telling the truth is not enough; you have to tell it with a name, with a fact, and in the language of the person asking.

AI search visibility: what this discipline is called (AEO, GEO) and what it pursues

SEO chases the click; AEO chases the citation. The discipline goes by three names for the same thing: AEO (Answer Engine Optimization), GEO (Generative Engine Optimization) and, in the plain terms people actually type, "AI search visibility" or "SEO for AI." The goal is for ChatGPT, Claude, Perplexity or Gemini to take a passage from your site as a source and attribute it to your brand.

What this case adds. In The definitive guide to optimizing your website for AEO we explain the full definition, llms.txt and FAQPage structured data; we do not repeat it here. This case turns that guide into a per-passage measurement and a loop that discovers keywords the site does not yet answer.

How we measure visibility in AI engines: a probability per passage, not a hunch

Seven questions in one call. The rubric asks each passage 7 TypeSafe questions: 2 of type Score (answers directly, citation-worthy) and 5 of type Noul (self-contained, names the entity, verifiable fact, promotional language, evasive), resolved in a single system_one call to the jev-latest model, which on TypeSafe's production API resolved to jev-1.13.0 on 2026-09-20.

The composite lives in code, not in the model. Weights: 0.30 directness, 0.15 self-contained, 0.10 entity, 0.20 verifiable, 0.10 no-fluff and 0.15 non-evasive; citation_worthy stays out of the composite as an independent control. A second re-ranking Noul asks, for each (query, passage) pair, whether an engine would cite that passage: 6 queries × 69 passages = 414 calls, mean latency 198 ms.

Cheap and stable, which is why it works as a threshold. The full study (585 cached calls) cost ≈ $0.02 USD at the published rate of $0.042 USD per million input tokens. Across 5 repetitions on 8 passages, the mean per-question standard deviation was 0.0083 (maximum 0.0359) and the composite's was 0.0049; no passage crossed the 0.70 band or the 0.30 band. Do not read this as accuracy: the passages are in Spanish and Jev is trained mainly in English, so this is a measurement of this content, not a benchmark.

Isolating the cause: a single word moved P(cited) +0.51

A one-word control. Adding only "Transgenia" to an otherwise identical sentence raised names_entity from 0.01 to 0.98, the composite from 0.84 to 0.94 and P(cited) for "what is Transgenia" from 0.17 to 0.68. A second control on another FAQ moved P(cited) from 0.30 to 0.41. The entity name is not vanity, it is attribution: brand visibility in AI starts there.

One sentence rewritten in the schema. The home page's Organization.description scored 0.81 and did not mention that Transgenia is an OpenAI Select Partner within the OpenAI Partner Network, a credential the schema did carry elsewhere. Rewritten with the legal name, founding year, partners and languages, it rose to 0.95, and P(cited) for "which Mexican company is an OpenAI Select Partner" went from 0.05 to 0.96. The home page meta description (159 characters) went from 0.59 to 0.92. Structured data for AI only helps when the text inside it is citable on its own.

Rewriting to get cited: the metric forced us to publish prices

The ceiling of "it depends." The first rewrite of the clinic pricing FAQ without publishing a figure (three scope variables, a 15-minute QuickLook and a free demo) raised the composite from 0.53 to 0.73 and P(cited) from 0.14 to 0.55, but evasive stayed at 0.45. The cookbook said it plainly: only a published range would move that query further.

We published aggregate billing, not quotes. Transgenia pulled the ranges from its own invoicing: two private clinics in Mexico City billed in 2025-2026, between 24,000 and 92,000 MXN in the first year (average ≈ 58,000 MXN; monthly consulting plan from 3790 MXN), and B2B distributors billed between 2023 and 2026, typically 110,000–320,000 MXN per year (average ≈ 190,000 MXN). All before 16% IVA (Mexico's value-added tax) and subject to scope. With the range published, the clinic FAQ went from 0.731 to 0.930 composite, P(cited) from 0.53 to 0.89, and evasive fell from 0.50 (the gate's baseline) to 0.03; the distributors FAQ raised P(cited) from 0.51 to 0.91.

Sanity check. The rewritten distributors FAQ, evaluated against the clinic query, did not improve (0.12 → 0.10), as it should be: a distributor answer should not win a clinic question. Hard rule: measurable transparency, not promises.

The gate before publishing: canary, holdout and zero fabrication

Three layers in every pull request. The first is deterministic and needs no API: every JSON-LD parses, the visible text of each FAQ matches the JSON-LD, and faq.json and i18n parse. The second runs the rubric against ci_baseline.json: an unchanged passage cannot drop more than 0.05, an edited one must improve or reach a composite ≥ 0.60, and a new one must reach 0.60. The third is the canary: the maximum P(cited) per prospect query cannot fall more than 0.05 from baseline.

A loop that never writes copy. The autoresearch loop (a JevOps pattern: propose, evaluate in isolation, apply only what scores better, remember) receives old/new proposals from a person or an agent. Jev evaluates the rubric old vs new, facts against site_facts.json (21 verified facts, each with a source and date) and the canary; it accepts only if the composite gain is ≥ 0.05, unsupported claims ≤ 0.50, names_entity ≥ 0.85 and no canary query drops from ≥ 0.70 to < 0.65. Every decision is logged in autoresearch_memory.jsonl.

The batches, failures included. P4–P6 (heroes, cards and methodology FAQs): 23 rewrites, mean composite 0.517 → 0.756; one got worse (0.553 → 0.543) and three stayed below 0.60. P8: 10 of 10 accepted (9 scored plus a Chinese mirror that is not scored), mean composite of the scored ones 0.559 → 0.904; one proposal was rejected on the first pass by a text-presence check (presence_ok=false) and accepted after fixing it. P9: 4 of 4, 0.493 → 0.913. The draft-first and human sign-off framework behind the gate is in How Transgenia operates with governed AI agents: the method that builds governance first.

Discovering new keywords with AI: the questions your site does not answer

From Google Suggest to prospect questions. We start from seeds in Google Suggest (es-MX), such as "posicionamiento en IA" (AI search visibility), "cómo aparecer en ChatGPT" (how to appear in ChatGPT) or "auditoría SEO con IA" (AI SEO audit), and collect their autocompletions: real demand, written the way people write it. Some seeds returned no autocompletion at all, among them "cómo saber si ChatGPT cita mi sitio" (how to know if ChatGPT cites my site): with no Suggest autocompletion there is no measured demand for that phrasing, but it is exactly the question this case answers, which is why we used it in the FAQ. From seeds and completions we draft questions just as a prospect would type them.

The threshold defines the new keyword. Each question is scored against every live passage on the site. If no passage reaches P(cited) 0.70, there is demand and no citable answer: that question is a new keyword, content gets written for it and it is measured again. AI-assisted keyword research comes from evidence that the answers are missing, not from intuition.

Canary and holdout so we do not fool ourselves. We keep 11 canary queries, used for optimization and checked on every PR, and 8 sealed holdout queries that only the production audit opens. The gap between the two (mean maximum P(cited) 0.93 versus 0.70) exposes copy overfitted to the questions we already knew. The four holdout gaps with 0 passages ≥ 0.70 were inventory regulated by Cofepris, Mexico's federal health regulator (0.52), Odoo migration (0.36), an agent that verifies CFDI, Mexico's electronic tax invoices, against the SAT, Mexico's tax authority (0.69), and per-project margin (0.68). P9 wrote that content (migration composite 0.22 → 0.90, CFDI-SAT 0.24 → 0.95); on 2026-09-21 all four moved to canary and four new holdout queries were sealed.

What production shows today: the AI SEO audit on the live HTML

We audit production, not the repo. On 2026-09-21 we compared 27 files of the live transgenia.org HTML, with cache-busting, against origin/main: 27 identical, 0 with drift. The gate passed on that snapshot with 168 ES+EN passages in ok status, and the re-ranking cost ≈ $0.022 USD (534,976 input tokens).

By page, and what is left. Mean composite in production: About Us 0.95, index 0.94, clinics 0.85, distributors 0.85, soluciones-ia 0.83, implementacion-ia 0.79, servicios 0.76 and join-us 0.61. 11 passages remain below 0.60 and join-us is still last. Audit verdict: GREEN, with the four holdout queries flagged at 0 as explicit exceptions.

What we deliberately did not measure. Every figure in this case is P(cited) as estimated by jev-1.13.0, not observed citations in ChatGPT, Perplexity, Gemini or Claude. Nor did we measure traffic, Google rankings, leads or revenue attributable to the change: no source exists, and Transgenia does not invent one.

What it costs and how to start: in-house or with an AI search visibility agency

Compute costs cents; judgment does not. The rubric over 69 passages cost ≈ $0.0032 USD (75,911 tokens, 233 ms mean latency), the re-ranking ≈ $0.0079 USD and the production audit ≈ $0.022 USD. What you should demand from any AI search visibility provider, Transgenia included, is not the compute but the gate: facts verified with a source, a sealed holdout and human sign-off before publishing. At Transgenia we apply that gate to our own site before proposing it to anyone.

1. Diagnosis. Extract the live passages from your site (FAQs, meta descriptions, Organization, heroes), run the 7-question rubric and the re-ranking against the questions your prospects type. Flag the passages with no entity and the queries with 0 passages ≥ 0.70.
2. Read-only pilot. Rewrite with facts you have already published, evaluate old vs new with the same model, and accept only gains ≥ 0.05 that break no canary query. Nothing reaches production without a diff and sign-off.
3. Governance. Put the gate in CI, seal a holdout and audit production with cache-busting after every deploy. The holdout queries that fail are your next batch of keywords.

If you want to see the stack in real operation before deciding, our sector demo runs on 100% simulated data and the QuickLook takes 15 minutes.

→ See the demo →

Frequently asked questions

About the author and Transgenia

Efraín Carreón Ortiz is CEO of Centrum Transgenia, a Mexican technology boutique. He holds the official Claude Code badge (Claude Partner Badge) issued by Anthropic, verifiable on Credly.

Transgenia is an OpenAI Select Partner within the OpenAI Partner Network and a registered partner in Anthropic's Claude Partner Network. We run twelve governed AI agents in our own production and support verifiable implementations in the healthcare and B2B sectors.

Let's talk: LinkedIn, book 15 minutes or the contact page.

Keep reading

• The definitive guide to optimizing your website for AEO: the full definition of AEO, llms.txt and FAQPage that this case takes for granted.
• SEO, GEO and AEO: how we orchestrate four AI models to move our own company's /goal: how we split SEO, GEO and AEO work across four models; this case is the measurement layer.
• How Transgenia operates with governed AI agents: the method that builds governance first: the draft-first and human sign-off framework behind the gate and the autoresearch loop.
• Transgenia signs the Partner Select partnership with OpenAI: the credential the home page's Organization.description did not mention and that raised P(cited) from 0.05 to 0.96.
• Our governed AI solutions.