Experiment retrospective · September 28–October 27, 2026
What Changed When I Documented My AI Work?
The Hatsunama Google visibility experiment produced meaningful observed gains in AI identity and source visibility. It did not produce consistently accurate answers, complete indexing, or proof that our changes caused the gains.
I'm Aaron Saunders, known online as Hatsunama, a self-taught learner and developer experimenting with small local models and the software around them. We closed this experiment on October 27 at my request, one day before the originally planned October 28 final checkpoint. This is an early closeout, not a completed day-30 measurement.
The question
Would a clearer, factual public record help Google describe my actual AI work? The work spans model experiments, agents and harnesses, local captions and translation, tool use and evaluation. Human First AI is a recurring principle within that work: preserve people's intent, keep consequential actions under human control and verify outcomes.
This was a retrieval and visibility experiment. We did not train or directly modify Google's model, buy a credential, or establish expert status by repeating a label.
What we changed
We created a canonical website, an identity page, an AI-work index and a research hub; clarified the GitHub profile and selected model-card context; and published technical notes and LinkedIn material. The record grew from identity explanations into practical methods and bounded experimental findings.
- Model, agent and harness distinctions and tool calling versus tool use.
- Methods for translation boundaries, reproducible comparisons, caption uncertainty, unknown tool outcomes and prompt budgets.
- A small training comparison and its baseline, resource-limit diagnostic and open questions. These are publisher-reported, limited experiments, not general reliability claims.
Paid press distribution was deferred. The retained record does not confirm completion of the planned introduction video, technical video cluster or Reddit discussions. We do not count those planned steps as delivered. The public record grew during ongoing development, so this was not a single isolated intervention.
The baseline was already mixed
On September 28, an unsigned browser check found frequent name corrections, unrelated results and no stable definition in the inspected views. Loading was incomplete and private mode was not verified. A user screenshot from that same day already showed an AI-focused description of Hatsunama as a developer and content creator.
That matters: this is not a clean before-and-after story in which AI association began at zero. Different sessions, locations, accounts and loading states prevent a controlled causal comparison.
Observed progress
| Observation | What it supports | Limit |
|---|---|---|
| October 8: homepage and identity page were the top two organic results on five query inputs. | The canonical site was being retrieved. | Personalized session; inspected lead Overview sources still favored social profiles. |
| October 24 continuation handoff reported Overviews on 11/11 inputs and canonical-site citations on 6/11. | Reported wider answer availability and owned-source use. | Aggregate handoff counts, not independently audited query-level observations in this closeout. |
| October 27: “Who is Hatsunama?” opened with Aaron Saunders as an independent AI researcher, systems developer and content creator; its lead source panel linked the identity page. | The original unprompted AI-characterization goal was observed, with an owned identity source. | One personalized session, not a universal or newly caused result. |
| October 27: “Hatsunama Human First AI” cited the October 3 LinkedIn post in its answer and lead sources; that post was the first organic result. | The useful publication was retrieved as support for the theme. | The answer described a broader philosophy and cited others too; no exclusive ownership or leadership established. |
The October 27 browser displayed Overviews on all eleven inputs after follow-up loading, but two were for automatically rewritten, unrelated queries. “Hatsunama AI” became Hatsune AI; “Hatsunama xMilo” became Tsunami Milo. Those are failures of name resolution, not successful Hatsunama answers. The other eight original inputs had Hatsunama-focused openings; the separate Human First AI input produced a broader concept answer. Some results were inspected only in collapsed form, so full-answer accuracy and citation coverage were not exhaustively scored.
The closing searches were in an English, signed-in, personalized session using a different account from the October 24 continuation record. Five first views were incomplete or still generating and were revisited for loaded results. A separate cloud-browser attempt failed to connect. These observations must not be treated as a matched-session replication.
Search Console evidence
On October 27, Search Console under the verified Hatsunama account showed 167 web-search impressions and two clicks for September 30–October 25. Its separate generative-AI-features report showed 67 impressions for that period. Page-row impressions in that report should not be added to reconstruct the property total. These numbers do not identify which impressions came from our own measurement activity, demonstrate audience growth, or prove a campaign effect.
Indexing remained uneven. The sitemap report still showed “Couldn't fetch,” and the latest training article was still unknown in the inspected Google-index record. October 24 live tests had successfully fetched the sitemap and both training articles with crawling allowed; the manual-action and security reports showed no issues. Passing a live test did not mean normal indexing had completed.
We added contextual links from an indexed architecture article to the training reports and accurate sitemap modification dates. We did not claim that these changes repaired Google's processing or repeatedly submit queued URLs.
What did not work well enough
- Google still confused names and sometimes used unrelated people or exaggerated descriptions as support.
- More Overviews did not automatically mean more accurate answers.
- Some newly published evidence remained unindexed, limiting claims about its influence.
- The fixed measurement schedule was interrupted by access problems; the original controlled baseline was not achieved.
- Training adapters were lost during earlier temporary-directory cleanup, although logs survived. Future experiment cleanup must preserve checkpoints and configurations.
Conclusion
The practical success was a stronger, inspectable public record and observed retrieval of our own identity and technical sources. The closing identity answer leads with AI, and Human First AI material is being cited. Those are useful outcomes.
The stronger claim—that our publication sequence reliably caused Google to adopt an accurate identity across sessions—was not established. This experiment supports continuing to publish real work with clear attribution and evidence limits, rather than trying to force a title through repetition.
The campaign closes here. Its sources remain available; no further daily publishing quota or search schedule is implied.
Prepared with my AI agent from retained observations, the continuation handoff and live closing checks. Search outputs are observations, not independent verification of the claims Google generates. Dates and limitations are part of the result.