Version 2026.11.21
On this page
Atlas4 new7 changed2 fixed
App
Invited teammates get a welcome email
When someone creates their account from an invitation to your organization, Adverity Atlas now emails the invited address to confirm the account is ready, with a link to sign in. If the person who owns that address didn't create the account, for example because the invitation was forwarded, the email tells them to contact your organization's admin.
Chat can read what an automation run wrote
Ask the assistant to review an automation's latest run and it now reads the run's full written report, not only its findings. Report-style automations that record no findings no longer come back as "nothing found". You can also ask for an automation by part of its name, or list only the automations you created. In an automation's run history, a run with a report but no findings now shows the start of the report instead of "No findings".
Data apps reload from cache
Reopening or reloading a data app within 15 minutes now shows its results straight away instead of running every query against your warehouse again. Data shown in a data app can therefore be up to 15 minutes old. Editing the app still shows fresh results at once.
Changes to a data source's access settings in a workspace, such as a narrower row filter, hidden tables or columns, or a removed permission to run SQL, also take up to 15 minutes to reach people who opened the app shortly before. Removing the data source from the workspace takes effect immediately.
Profile photos from Google
If you sign in with Google, your Google profile photo appears, after your next Google sign-in, next to your name across Atlas: the top bar, your profile, member lists and sharing dialogs. Atlas keeps its own copy of the photo, so viewing it never sends anyone's browser to Google. Later photo changes are picked up the same way. People who sign in another way keep their coloured initials.
SSO setup shows Adverity's values before you configure the IdP
Setting up SSO used to be circular: your identity provider needed Adverity's entity ID and ACS URL before it would give you the values a connection requires, but Settings > Single Sign-On only showed Adverity's values after a connection was saved. They are now shown from the start, together with the setup guides and a working Download XML for the service provider metadata. The download previously failed with "Error generating metadata." for every organization, and now returns the file.
Old interrupted replies no longer block a conversation
Some conversations refused every new message and every retry with "A reply is still being written. Wait for it to finish.", even though nothing was running. This happened when an earlier reply in the conversation had been interrupted and never completed. Those conversations accept messages again, and an interrupted reply only holds up the conversation while it could still be running.
Data sources
Choose the BigQuery projects a service account uses
When you connect Google BigQuery with a Service Account JSON key or Workload Identity Federation, Adverity Atlas now lists the Google Cloud projects those credentials can see and asks for two:
- Storage project: the project whose datasets Atlas reads. It starts on the project the key was created in, when the key can see it.
- Compute project: the project the queries run in and are billed to. It starts on Same as storage project.
When the credentials see only one project, Atlas picks it for both and skips the question. Workload Identity Federation no longer asks you to type a project ID.
Atlas saves both choices with the data source. Reading from the storage project and querying in the compute project follow in an upcoming update; until then, a service account keeps using the project named in its key.
Choose the BigQuery projects after signing in with Google
When you connect Google BigQuery with Sign in with Google, Adverity Atlas now asks for two projects instead of one, from the Google Cloud projects your account can use BigQuery in:
- Storage project: the project whose datasets Atlas reads.
- Compute project: the project the queries run in and are billed to. It starts on Same as storage project.
You choose them right after signing in, before the connection test, so the test checks the project you picked. When your account sees only one project, Atlas picks it for both and skips the question. Signing in again, or with another account, asks again.
Re-authorizing an existing BigQuery data source shows the same two choices, starting on the projects it uses now, so that is also where you change them.
Atlas saves both choices with the data source. Querying in a separate compute project follows in an upcoming update; until then, the data source reads and queries the storage project.
Change Google Ads With Admin Approval
Ask in chat to change a Google Ads account and the agent can prepare it: pause campaigns, ad groups, ads or keywords, add negative keywords, apply labels, and set budgets or CPC bids. Every change waits for a workspace admin to approve it, and the approval shows exactly what will change, with the current and new values. Right before the change is applied it is checked again, so if someone edited the same item in Google Ads in the meantime, nothing is written and you are told why. The result, applied, refused, rejected or expired, is posted back in your chat.
BigQuery queries run in the compute project
Google BigQuery data sources now use both projects you chose when connecting them:
- Storage project: Atlas reads datasets, tables and their data from here.
- Compute project: syncs, profiling and chat questions run their queries here, and the queries are billed here.
A central service account can now query data in other projects. It needs to read data only in the storage project and to run queries only in the compute project.
The connection test now also checks that Atlas can read the storage project. When it cannot, the test names that project.
Each data source's details show both projects. Existing BigQuery data sources were given both projects automatically, set to the project they used before, so nothing changes for them.
Extract Knowledge checks every item against your data and shows where it found it
Extract Knowledge now explores each data source the way an analyst would and checks every finding against the full data before saving it. It finds brands, products, markets, channels and campaign objectives, how your campaign names encode them, which tables are copies of each other, and data caveats such as days with no rows or clicks recorded without impressions. A period such as 2026Q1 is no longer saved as a market, and a brand that shows up on a handful of another client's rows is asked about instead of saved as a fact.
Every extracted item now says where Atlas found it: the data source, example campaign names with the matching part highlighted, and how many campaigns and how much spend it covers. When Atlas is unsure, the item shows Needs your answer with the question it could not decide, and Yes, keep it or No, archive it settles it. When a second data source holds the same brand or market, it is added to the existing item rather than saved twice.
Extraction runs once for each group of workspaces that see the same rows of a data source, and the run shows on the Tasks page as Extract knowledge.
BigQuery connection test names the missing role
The project picker for Google BigQuery lists every project your account has any role in, so it was possible to pick a project where Atlas cannot actually work. The connection test now checks both projects you chose and tells you what to fix:
- Compute project: when the account cannot run queries there, the test asks you to grant it the BigQuery Job User role in that project.
- Storage project: when the account cannot see any datasets there, the test asks you to grant it the BigQuery Data Viewer role in that project.
Each message names the project and comes with the gcloud command that grants the role.
A connection test that fails because the credentials were rejected no longer mentions Snowflake for data sources of other types.
Data apps on BigQuery load again
Data apps that filter a Google BigQuery data source, for example by a date range, showed "Query failed: … Syntax error: Unexpected string literal" instead of their results after the last release. They load again.
Data apps on Databricks no longer fail the same way when a query quotes a column name with backticks.