When results change overnight, whoever pays wants an explanation. Algorithms rarely give a complete one, but there are useful questions to ask the platform, the manager and the account itself. Practical transparency comes from records, comparison and clear hypotheses.
The essentials in a definition
Algorithm transparency, in advertising, is the ability to understand what data and rules influence results and to get useful explanations from the platform and manager when something changes. It does not mean seeing the code, but logs, reports and clear answers. The AI Act and the Digital Services Act reinforce transparency obligations in certain contexts.
What can and cannot be known
Platforms explain partially: change-history reports, auction statistics, recommendations and diagnostics. They do not reveal the algorithm’s details, and no explanation covers every cause of a variation.
So practical transparency depends less on knowing how the algorithm works and more on documenting what changed around it: budget, ads, data, competition, seasonality.
Questions to ask the platform
Were there policy or system changes in the week results shifted? Did the account go through a review or restriction? Which signals are being used in optimisation? Are there diagnostics available on the quality of conversion data?
Keep the answers. Even when generic, they help eliminate hypotheses and can be useful in support escalations.
Investigating a change in results
- Confirm that data and tracking are correct
- List internal changes and external events
- Formulate a testable hypothesis
- Test and record the conclusion
Questions to ask the manager or agency
What changed in the account over the last two weeks? What is your hypothesis for the variation? What test do you propose? What data would confirm or refute it? When will we have an answer?
A good answer is specific and testable. “The algorithm is learning” may be true, but with no deadline and no criterion it is not an explanation.
When you cannot see inside the algorithm, look at what changed around it.
A record that speaks for itself
Keep a change diary with date, author, reason and expected result. Add external events: promotions, holidays, website outages, competitor changes. When there is a variation, the record lets you cross-check likely causes.
Over time this history becomes an asset: it lets you learn patterns, justify decisions and train new colleagues.
Decision table
| Question | To whom | What to expect |
|---|---|---|
| What changed in the account? | Manager | Dated change history |
| What signals and data does the campaign use? | Platform and manager | List of data sources and conversions |
| Why did delivery fall? | Platform | Diagnostics, statuses and recommendations |
| Which decisions are automatic? | Manager | Map of what is manual and automatic |
Common mistakes to avoid
- Accepting “it is the algorithm”. Vague explanations are not enough; demand verifiable facts.
- Not keeping history. Without a log of changes, no explanation is possible.
- Expecting full access. Define what is reasonable to ask and what the platform does not disclose.
How to measure the result
Measure average incident diagnosis time and the percentage of drops with an identified cause. If the first rises or the second falls, improve records and questions to the platform.
Limits and exceptions
There are limits to what can be known about algorithms: platforms do not disclose their detailed workings, and part of decisions depends on signals the advertiser does not see. Practical transparency comes through your own records, concrete questions and comparisons, not from waiting for complete explanations. When a decision has relevant impact, document the context, request formal support and, if necessary, seek a second opinion from an independent specialist.
Illustrative example
Illustrative example: after a sudden drop in conversions, the advertiser sends the manager and the platform a timeline of changes made, disapproval statuses and site changes. The analysis finds a set of products was disapproved in the feed. Transparency drawn from an organised log shortened diagnosis from days to hours. The NIST approach values exactly this documentation.
Frequently asked questions
Do platforms always explain variations?
Rarely in full. They give partial diagnostics that are worth cross-checking with your own data.
What is a good change log?
A diary with date, author, reason and expected result, including external events.
Should I escalate to platform support?
When you suspect an error or an account limitation and have already ruled out internal causes.
Can I ask the platform to explain a drop?
You can ask for diagnostics and support; bring data, dates and changes to get useful answers.
What should I demand from my manager or agency?
Reports with clear definitions, change history and an explanation of automatic and manual decisions.
Practical transparency means change logs, maps of what is automatic and concrete questions to the platform and manager, not access to the algorithm’s code.
Sources and recommended reading
Last editorial review: October 2026. Informational content; always confirm current conditions in the official sources before deciding.