Suggested segments from activity and measures

What it does

Suggested segments have the system propose customer groups you didn't think to ask for, either from activity data using recency, frequency and monetary analysis, or from a measure using a decision tree model that finds which attributes move that measure. You review the suggestions and turn the useful ones into real segments.

Key facts

Preview. Both flavours — activity-based and measure-based — are documented as preview, so not for production-critical work, and the docs do not name a GA wave.
Two entry points, both on the Suggested segments tab: suggestions based on activity, and suggestions based on measures.
Activity-based suggestions need activity data already ingested and mapped. They run an RFM-style analysis and produce groups like most active customers, highest revenue contributors, and customers who've gone quiet.
Measure-based suggestions need at least one measure to exist. You pick a primary attribute — typically a measure such as total spend per customer — and then the influencing attributes to test against it, for example membership tier, tenure or occupation.
The measure route works in reverse too: pick a customer attribute as the primary (say, rewards member) and find out what influences it.
The model is a decision tree. It surfaces rules where the population behaves meaningfully differently from the average, and it only shows suggestions that are statistically significant — so an empty result is a real answer, not a failure.
A primary categorical attribute is capped at 10 categories. If yours has more, group them down before using it.
An influencing categorical attribute is capped at 100 categories.
Microsoft flags this explicitly as profiling under privacy law, and warns against using sensitive attributes such as race, sexual orientation or gender as inputs. That is a real conversation to have with the client's DPO, not a footnote.

When to use / skip

This is a discovery tool, not a delivery tool. It earns its keep in the analysis phase, when the client has data unified but no clear view of which customer groups matter — you point it at a KPI measure, it comes back with "customers on tier 2 with tenure over three years spend materially more", and that becomes a proper segment plus a proper conversation.

Skip it when the client already knows their audiences. Retail and financial services clients usually arrive with segment definitions that came out of a strategy deck two years ago, and no AI suggestion is going to displace them. Skip it too when the measure you'd point it at is thin or newly built — the model will find patterns in whatever you give it, and the significance filter only protects you from noise, not from a badly defined measure.

Because it's preview, don't design a delivery around it. Use it, harvest what it finds, and make the resulting segments stand on their own definitions.

Configuration decisions

Which flavour matches the question: activity-based when the question is "who's engaged and who's drifting", measure-based when the question is "what drives this number".
Which measure or attribute is the primary. This choice frames every suggestion that comes back, and it's worth doing two or three runs with different primaries.
Which influencing attributes to include. Restrict to attributes the client can actually act on — including a system-generated identifier as an influencer is a waste of a run.
How to group high-cardinality categoricals down to the 10-category limit for a primary attribute. That grouping is a business decision, not a technical one.
Which suggestions get promoted to real segments, and whether they're rebuilt in the segment builder with an explicit definition rather than kept as generated output.
Whether any of the candidate attributes carry protected-characteristic risk, before running anything.

Gotchas

Preview status is the headline gotcha. Terms differ, behaviour can change between waves, and support expectations are not the same as GA features.
The 10-category limit on primary categorical attributes catches people out because the natural primaries — product category, region, tier — usually exceed it.
Suggestions are only as good as the measure behind them. A measure with a quiet data quality problem produces confident-looking suggestions built on rubbish.
A run with no results reads like a broken feature. It isn't: it means nothing crossed the significance bar. Explain that before the client sees it.
The profiling warning has genuine compliance weight in the UK and EU. Using attributes that proxy for protected characteristics — postcode, first name — is the trap, not the obvious ones the docs list.
Segments created from suggestions land in the same active-segment pool as everything else, and exploratory runs generate more of them than anyone expects.

Consultant notes

Demo this in analysis workshops, not in the go-live demo. It shows well when the client's own data produces a surprise, and it shows badly when it produces three obvious groups they already knew about.
Frame the output as a hypothesis to test, not an answer. The credibility of the whole CI Data programme suffers if a suggested segment is presented as fact and turns out to be a data artefact.
Get the privacy position agreed in writing before the first run, listing which attributes are in scope as influencers. UK clients with a DPO will ask, and having it already done is worth a lot.
Before go-live, check whether anything in production actually depends on a preview-generated segment. If it does, rebuild the definition explicitly in the segment builder.

Worth another look once suggested segments reach GA, or if the category limits change.