In this article

Separate retrieval, recommendations and campaign settings
Meta describes Andromeda as a retrieval system, narrowing a large set of candidate ads before later recommendation stages. Its engineering explanation was published in December 2024. [1] Meta’s October 2026 GEM article describes its central ads recommendation foundation model and further scaling work. [2]
These are related parts of the advertising system, not interchangeable campaign switches. A publication date does not establish the date a particular account changed. Keep the platform source separate from an account-specific observation and any hypothesis connecting them.
Do not assume a decline proves an algorithm effect
Start with complete dates, a consistent conversion definition and real outcome counts. Review spend, impressions, traffic and business results. Determine the first stage where performance changes materially, and identify the campaigns or markets contributing most to the change.
Keep website deployments, offers, stock, tracking releases and campaign edits on the same timeline. A conversion drop limited to one destination deserves a different investigation from a delivery interruption across the account. An algorithm explanation should earn its place through evidence rather than become the default answer.
Review creative concepts and input quality
Inspect the message, proof, offer and destination for the assets actually running. Different buying objections can justify different concepts, but a fixed creative-count rule cannot guarantee a profitable campaign. Measure whether the assets reach relevant users and produce the intended business outcome.
Also verify the optimization event and its collection. A form start, submitted enquiry and booked appointment are different signals. The account review needs to know what the platform is counting and whether downstream qualification evidence exists.
Use a decomposition instead of an arbitrary recovery recipe
A higher cost per outcome can arise from higher traffic costs, lower conversion rate, different allocation or measurement changes. Compare those components within stable segments before choosing an action. Show counts alongside rates so sparse results are obvious.
For a hypothetical campaign, stable visitor volume with fewer completed registrations points toward a different test from fewer visitors at the old registration rate. Both can create the same final cost increase. A fixed recipe to consolidate, refresh and wait a set number of days cannot distinguish them.
Design one reviewable intervention
State the hypothesis, exact proposed change, what remains stable, the relevant outcome and a review window that reflects buying delay. If the action is a protective response rather than an experiment, say so. Keep budget constraints and account scope explicit.
Retain the evidence even when the test is inconclusive. A destination fix that restores form completion supports a different conclusion from a new asset that receives a different audience mix. The next report should describe what was observed rather than promise a universal recovery percentage.
Investigate the account with 1ClickReport
Connect the client’s Meta account and relevant analytics source. Ask the Analyst to compare complete periods, identify the largest contributors and state unavailable evidence. Native billing notices, delivery diagnostics and certain new metrics may need a separate check.
For current developments, read the focused guides on the three October brand updates and GEM. For a sudden decline, use the October diagnostic. This broader page explains the system distinctions so those investigations begin with a clear question instead of an unsupported claim that all Meta ads stopped working.
A first task to try
Investigate this client’s Meta performance over comparable complete periods. Separate documented platform changes from observed account changes. Decompose the result into delivery, traffic cost, conversion and allocation, list competing explanations and recommend the next test with no assumed algorithm cause.Questions readers ask
Does an algorithm update explain every performance drop?
No. Verify account delivery, tracking, traffic, destination and business outcomes.
Is there a guaranteed recovery recipe?
No. The useful action depends on the measured failure and the client’s constraints.
Sources and editorial method
Research checked October 26, 2026. Platform facts link to primary sources below. The worksheets and operating examples are our editorial recommendations; hypothetical examples are not customer results. This article is published by 1ClickReport.