Retail analytics has an uncomfortable secret: most "customer behaviour" data is merchandise behaviour observed from a distance. RFID makes the merchandise side genuinely precise, and that precision is worth more than a vague story about footfall.

Item-level RFID measures garments, not people

Be clear about the unit of observation before designing anything.

A tag identifies an article of clothing, its size, its colour and its location.
A tag does not identify a person, an intention, or a household.
What you infer about people is a model layered on top of garment events.
The value is in the garment's journey: floor, fitting room, till, or back to stock.
Any claim about "customer journeys" should state which garment events support it.

The fitting room signal and what it proves

A garment entering a fitting room is the strongest non-transactional signal in a store.

Entry is a real event: the item left the floor and entered a private space.
Non-purchase after fitting is a specific, actionable failure, not a generic "lost sale".
The difference between "tried and rejected" and "never noticed" is the whole point.
Size availability explains a large share of fitting-room rejections and is fixable.
A pile of untagged returns from the fitting room means the process, not the customer, failed.

Store traffic: counting bodies is not identifying them

Traffic counting and item tracking answer different questions.

A door counter gives a denominator; without it, conversion has no meaning.
Traffic systems that count bluetooth probes or camera tracks carry legal obligations that vary by country.
Staff exclusion matters; a store where employees are counted inflates the denominator by a third.
Group entry breaks the "one person, one visit" assumption and should be acknowledged.
Never attempt to match a specific person to a specific purchase; that is where analytics becomes surveillance.

The join key problem

Item events and sales events are recorded in different systems with different clocks.

The POS knows the basket, the timestamp and the tender; the RFID system knows the garment.
The join is by EPC at the point of sale, and only if the tag was actually read at the till.
A till without a reader produces a sale with no item event, and the model silently loses it.
Time-window joins are a proxy, and proxies must be labelled as such.
Reconcile the counts daily: items sold, items read, and the gap explained.

What Odoo can hold and what it cannot

Odoo is good at transactions and inventory; it is not a behavioural data platform.

Sales orders, POS orders, stock moves and product variants are native and reliable.
Fitting-room events have no native model and need a custom table with a clear schema.
Traffic counts belong in a separate time-series store, not in the product master.
Keep analytics tables out of the transactional path, so a heavy query cannot slow the till.
Expose the results as a read-only reporting layer, refreshed on a schedule.

Attribution windows and the honesty of conversion

Conversion rate is the most quoted and least defined retail metric.

Define the window: same visit, same day, or within seven days, and state it everywhere.
A garment fitted at 14:00 and bought at 14:20 is a clean conversion; one bought next week is a guess.
Returns should reduce conversion, or the metric rewards the wrong behaviour.
Exclude staff purchases and known wholesale transactions from the denominator.
Publish the definition next to the number, because the argument is always about the definition.

Sell-through, dwell and the size curve

The operational value is in assortment and replenishment decisions.

Sell-through by variant tells you which size and colour combination is actually moving.
Dwell time on the floor separates "not seen" from "seen and rejected".
Fitting-room rejection rate by size is the clearest signal for a replenishment gap.
Use the data to fix the size curve, not to blame the store team.
A single week is noise; require at least four weeks before changing a buy.

Privacy: the boundary you must not cross

The legal line is closer than the marketing ambition.

Never link a tag to a payment card, a loyalty account or a phone identifier in the analytics layer.
Fit-room data should be aggregate and, ideally, not retained at the individual garment level beyond the season.
Staff must be able to explain, in one sentence, what is collected and why.
Signage and notice requirements differ by jurisdiction; assume the strictest one you operate in.
A privacy impact assessment before the pilot is cheaper than a deletion order afterwards.

Data quality: the reads that never happened

Missing reads do not announce themselves, and they bias every conclusion.

A fitting room without a reader records nothing, which looks like "no fittings today".
Metal fixtures, mirrors and crowded rails reduce read rates in ways that vary by hour.
Duplicate reads inflate dwell time; deduplicate by tag and by minute.
Items leaving the store without a till read appear as permanent stock, distorting availability.
Publish a data quality score alongside every dashboard, because trust is the actual product.

From dashboard to decision

A report nobody acts on is a cost, not an insight.

Every metric should have a named owner and a decision it changes.
Size-level replenishment is the most reliable first use case.
Fitting-room rejection review belongs in the weekly merchandising meeting, not in a monthly deck.
Kill metrics that nobody has acted on for two quarters.
Measure whether the decisions changed, not whether the dashboard was opened.