Overview
The agent's limits are built into its access and publishing paths. Each run uses your workspace credentials, and a store write requires the explicit instruction and permissions described below.
Publishing
The agent never publishes store content on its own. Not after a successful run, not because it is confident, and never without a person starting it.
Accepting a metadata proposal updates a Spek draft. Publishing that draft, submitting an event or product page, toggling visibility, or starting a real store test are each separate actions that a person must explicitly request, governed by what the store currently allows.
Agent chat can carry out a direct interactive publish command when the user's role and the app's store prerequisites allow it. It writes the requested or pending locale drafts and submits the eligible version to App Review. That authority exists only for the verified message and cannot be borrowed by a scheduled run, an attachment, quoted text, an earlier message or the model itself.
Store access
The agent reaches App Store Connect only through its official token API, using the credential your workspace supplied. There is no Apple ID login, no stored session, no cookies and no headless browser driving a console on your behalf.
This is a safety boundary rather than an implementation detail. It is explained in how store connections work.
Data
Every keyword the agent proposes came from a real store response. Nothing in the discovery path may emit a term that was not seen in a store's own search or autocomplete results.
The same applies to numbers. If a market's research has not resolved yet, the agent waits rather than drafting against partial data, and it will tell you a value is unavailable rather than producing a plausible one.
Customer-facing answers describe Spek research and Spek estimates. They do not name outside ASO tools or contrast a Spek number against another vendor.
Scope
Some things simply are not connected, so the agent cannot reason about them however you ask it to.
- Store impressions, product page views and downloads.
- Conversion rate, and anything derived from it.
- Apple Search Ads, including its popularity data.
- The results of store-run product page tests. The agent can draft one; it cannot read back how it performed.
- Revenue, retention and anything from inside your app.
What it does measure is keyword rank and the search visibility derived from it. See what Spek measures.
Beyond that, the agent works only inside the scopes you granted, on the markets you chose, for the app you set it up on. A capability that is not in its scopes does not exist as far as a run is concerned.
Paid research and screenshot generation start only from your direct, verified command. Stop cancels work that is still queued. If the paid provider call has already started, Spek does not send it again; it waits for the worker's real result or shows that reconciliation is needed.
Common questions
Could anything publish while I am away?
No. The agent never publishes store content on its own. Its output waits in the inbox until you look at it. Only a verified interactive command, the Agent inbox header, or an Editor publish action can start a listing publish.
A prepared proposal and an approved store action are different stages. Review the outcome of any action you explicitly requested, including one that continues in the background. The absence of an open browser tab does not turn a pending draft into an authorization to publish.
Does it see my downloads or revenue?
No. Store impressions, downloads, conversion, Apple Search Ads and in-app analytics are all unconnected. It measures keyword rank and the search visibility derived from it.
Can it invent a keyword?
No. Every term it proposes came from a real store response. If a market's research has not resolved, it waits rather than drafting against partial data.
What is the difference between asking for advice and authorizing a store action?
A question asks the Agent to explain or inspect something. A direct action request asks it to perform a supported task for the selected app, subject to your permissions and the current store state. Keep the intended app, locale and action clear, especially when publishing. A proposal, plan or research result does not independently authorize a later store write. Read the reported outcome after an action instead of assuming that a conversational acknowledgement means the provider has accepted the change.
Can I use the Agent's estimates as a business forecast?
Use keyword estimates to inform an ASO review within the relevant market. They are not a forecast of downloads, purchases or revenue, and the Agent cannot fill those gaps with data it does not receive. If a business decision depends on conversion or retention, consult your own store and app analytics alongside the search observations. Keep the sources and time periods clear in a report so a directional search metric is not accidentally presented to colleagues as a measured financial result.
Why can the same request work for one person but not another?
Workspace membership and role affect what a person can change, and provider permissions and app state place further limits on store operations. A colleague with publishing access may therefore have controls that a viewer does not. Check the account, workspace and selected app when comparing behavior. Asking the Agent to ignore an access restriction does not grant permission. If access needs to change, the responsible workspace or store administrator should review that change through the normal account controls.
What should I do when an action's result is uncertain?
Preserve the visible status and check the available recovery action. For a supported uncertain publish, reconciliation checks what the store accepted before another write is considered. If the product asks you to inspect the store directly, verify the exact app, version, locale and field there. Avoid repeating a paid or publishing command merely because a response took longer than expected. The useful next step is establishing the first operation's outcome, not creating a second operation whose effect may overlap it.