# Agentzia historical before-and-after: source notes

Study date: August 28, 2026.
Prepared from repository research notes for the article:
https://www.agentzia.dev/resources/geo-before-and-after-agentzia

This is an extract of Agentzia's own recorded research summaries, not complete raw transcripts, independent validation, or a new experiment. Model names and scoring labels are historical. No new runs were commissioned for this article.

## Provenance

Source files at repository commit 552bbe5:
- feedback/dogfood/01-baseline-report.md
- feedback/dogfood/02-after-pricing-proof-docs.md

Baseline run: 373856ff-45f5-460f-b835-9fcafedd2c43
Follow-up run: d76ebda3-b42b-4e2c-a0f3-fda89c998153
Target: https://www.agentzia.dev
Suite: blind-core-v3
Participant model identifiers in both notes: openai/gpt-5.4 and anthropic/claude-sonnet-5
Observer model identifier in both notes: openai/gpt-5.4
Coverage: six of six scenarios in each run.
Scenarios: direct evaluation, category discovery, competitive comparison, trust research, purchase decision, implementation research.

## Recorded intervention

Between runs, Agentzia added explicit free and paid usage terms on a pricing page, a self-study with methodology and cost scope, canonical MCP setup documentation, and narrower claims about observable assistant behavior. Several changes happened together.

## Recorded counts

| Signal | Baseline | Follow-up |
| --- | --- | --- |
| Discovered | 5/6 | 5/6 |
| Understood | 4/6 | 5/6 |
| Trusted | 0/6 | 2/6 |
| Considered | 3/6 | 4/6 |
| Selected | 1/6 | 1/6 |
| Historical summary score | 42/100 | 53/100 |

The aggregate discovery count mixes directed and unbranded tasks. The category-discovery scenario still missed Agentzia. It is not a five-of-six unprompted search appearance rate.

## Follow-up observations in the notes

Four participants found pricing on the pricing page. Direct evaluation still reported missing pricing on the homepage. Trust and purchase research cited missing operator/legal information. These observations concern the historical website, not its current state.

## Limits

- One baseline run and one follow-up run; no repeated matched series or randomized control.
- Six different scenarios, not six interchangeable buyers.
- Multiple simultaneous edits; no isolated effect for an individual page.
- Citation-level evidence was recorded, but exact browser replay was unavailable.
- Search results, retrieved sources, and model responses can vary.
- No measured effect on human conversion, traffic, rankings, or revenue.
- Do not compare these early-suite counts numerically with the later four-profile sample reports.

Conclusion: selection was unchanged in these two recorded runs. Higher intermediate counts and a higher summary score do not establish a causal improvement or a commercial outcome.