A Real Before-and-After: From Zero Citations to Being Cited by ChatGPT

This is our own brand, not a client. We publish it because it is the one case where the entire before-and-after can be shown in public — including the unflattering starting numbers and the conclusions we decided we could not defend.

Disclosure: TriloAI builds GEO by TriloAI. The figures come from our own system, which you cannot independently audit. The website changes described below, however, you can verify yourself.

Starting point: August 17, 2026

The more revealing figure sits underneath: that scan collected 39 cited source URLs, and triloai.org appeared in none of them. The models knew the name but never retrieved the pages — mentioned, never cited.

The four actions the report gave us

1. Add llms.txt defining the brand, products, and key pages in one place
2. Add sitemap.xml covering service pages, FAQ, and case pages
3. Upgrade the existing FAQPage schema for consistency with service pages
4. Add a one-line positioning statement and use case to every service page header

All four were implemented. You do not need to take our word for it: triloai.org/llms.txt, triloai.org/sitemap.xml, and triloai.org/robots.txt are public files, and the FAQPage structured data is visible in the homepage source.

Result: ChatGPT retrieved the page

In the August 23, 2026 scan, ChatGPT cited this URL while answering a question about consultants specializing in fitness studios and sports venues:

https://triloai.org/services/n8n?utm_source=openai

The utm_source=openai parameter is appended by ChatGPT itself, which is evidence that it actually fetched the page rather than recalling the brand name. Against 0 own-domain citations on August 17, that is a change of state.

What we deliberately do not claim

You might expect a percentage improvement here. We checked, and that comparison does not hold up.

• The platform set differed. The baseline covered four platforms; the re-scan did not include Google AI Overviews. Different denominators cannot be compared.
• The engine model tier differed between runs. Measuring the same brand with different model tiers moves the score on its own.
• AI answers vary run to run. A single difference is not a trend.

So we claim only the part unaffected by all three: whether ChatGPT actually retrieved triloai.org is a binary fact. On August 17 it was zero. On August 23 there was one verifiable citation.

How to run a valid before-and-after

1. Fix the prompts. Reuse the exact same questions; change them and the scores stop being comparable.
2. Fix the platform set. Dropping one platform changes the weighted denominator.
3. Fix the model tier used for measurement.
4. Disable caching, or confirm the cache has expired — otherwise you may receive the previous run's answers and read it as “no change”.
5. Record cited sources, not just scores. Scores drift; whether your domain was cited is far harder to misread.

We violated three of these on our own first re-scan, found out afterwards, and withdrew the conclusion. That is why this page contains no percentage.

What this case can and cannot show

It shows that the problems the report flagged were real, and that after fixing them we started appearing where we had not appeared before. It cannot show that the same steps work for every brand, or how long they take: this is one brand and one observation. The four actions were also done together, so we can't tell which one made the difference, and no major AI platform has said it uses llms.txt to decide what to cite. We will keep recording and update this page when there is more data.