Start with facts that survive a change of chat
A particular executor's conversation should not be the project's only storage location. AURA retains the business description, questions, tasks, publications, observations and actions. A new authorized agent first verifies the domain and reads available data. It should not resubmit every URL or rewrite already published text simply because it was not present in the previous session. Ask which step remains unfinished, which confirmation is awaited and whether the page changed after the last report. This matters especially after a break or a change of contractor.
A completion record needs substance
An applied-change task should include a file description, diff and checks performed. However, the executor's ‘done’ message remains an agent report. Independent confirmation—such as text and canonical at a public URL or an actual search-provider response—is stored and read separately. If a task only proposed content, preparing it does not mean publication. If a sitemap was submitted, that does not mean indexing. Every plan item needs a clear criterion: which observable change will allow this particular stage to count as complete?
Saved answers and a new run serve different tasks
Reading existing observations is sufficient to analyze past citations. Evaluating a current answer requires a new request to the actual service. Preserve the full text, links, time, actual platform, recorded model and collection method. An agent's own research must not be placed under another AI's logo. Run comparisons make sense with identical question wording, market and language, and comparable recorded models and modes. A partly completed run or missing access must not look like visibility falling to zero. An unknown result remains unknown.
Finish with a short, clear summary
A good summary separates completed work, collected evidence, open tasks and one necessary owner action. For example: canonical corrected, public check passed, notification accepted, a new Gemini answer not yet obtained because sign-in is required. That report can be verified and continued even if the next executor never saw the previous chat. Save stop reasons and return to the relevant task without duplicating results on a technical retry. History helps choose the next step but does not by itself prove that one particular change caused later citation growth.
Conditions and limitations
- An agent report is not independent confirmation
- Rereading is not a new AI request
- A difference in observations does not prove causality