Marketing Data: UTMs, Channels, and Analysis Grain

Where does marketing data come from?

A marketing department works with three main families of data:

• sessions (or visits): one row per visit to the site, with its origin (source, medium, campaign), the device, and whether it led to a conversion;
• touchpoints: one row per interaction of a person with a campaign before converting (click on an ad, opening of an email, visit from a social network);
• conversions: one row per goal reached (purchase, quote request, sign-up), with its value.

This data is produced by an audience measurement tool (such as Google Analytics 4), by advertising platforms, and by your own orders database. They never tell exactly the same story: each tool has its own counting rules. Your first job as an analyst is therefore to know the grain of the table you are handling: does a row correspond to a session, a person, a click, an order?

UTM parameters

To know where a visit comes from, you add UTM parameters to campaign links:

https://boutique.example/soldes?utm_source=newsletter&utm_medium=email&utm_campaign=soldes_hiver

• utm_source: the precise sender (newsletter, google, facebook);
• utm_medium: the type of medium (email, cpc for paid click, social, referral);
• utm_campaign: the campaign name;
• utm_content and utm_term (optional): the ad variant and the keyword.

Python reads these parameters with the standard urllib.parse module:

from urllib.parse import urlparse, parse_qs

url = "https://boutique.example/?utm_source=Newsletter&utm_medium=Email&utm_campaign=Soldes_Hiver"
params = parse_qs(urlparse(url).query)
print(params)   # {'utm_source': ['Newsletter'], 'utm_medium': ['Email'], ...}

parse_qs returns a list of values per parameter, because a parameter can appear several times. You generally take the first one.

Normalize before counting

Measurement tools distinguish Facebook, facebook, and facebook: without normalization, the same source ends up split across three rows and each looks small. Two simple rules:

1. put all values in lowercase and remove extra spaces;
2. publish a naming convention (for example: lowercase, words separated by _, a closed list of allowed mediums) and have everyone who creates links follow it.

When a visit has no UTM and no identifiable referring site, it is classified as direct access: source (direct), medium (none). Be careful, “direct” does not mean “the user typed the address”: an untagged link in a PDF, a messaging app, or a badly tagged email also arrives as direct.

Group into channels

A marketing manager thinks in channels (paid search, email, social networks…) rather than in source / medium pairs. Grouping is done with ordered rules: the first rule that applies wins.

def channel_group(source, medium):
    if source == "(direct)":
        return "Direct"
    if medium == "email":
        return "E-mail"
    if medium in ("cpc", "ppc"):
        return "Recherche payante"
    if medium == "paid_social":
        return "Social payant"
    if medium == "organic":
        return "Recherche naturelle"
    if medium in ("social", "organic_social"):
        return "Social naturel"
    if medium == "referral":
        return "Sites référents"
    return "Autres"

Market tools offer their own default groupings, with richer rules. Writing your own has two advantages: you know exactly what each channel contains, and you can apply them to data that does not come from that tool (CRM, advertising export).

A first table by channel

sessions["channel"] = [channel_group(s, m) for s, m in zip(sessions["source"], sessions["medium"])]
by_channel = sessions.groupby("channel").agg(
    sessions=("converted", "size"),
    conversions=("converted", "sum"),
)
by_channel["conversion_rate"] = by_channel["conversions"] / by_channel["sessions"]

This rate is a conversion rate per session. If you compare it with a rate computed per user in another report, you will be comparing two different things. Always note the grain next to the figure.

Key takeaways

• Identify the grain of each table before computing.
• Normalize UTM values and enforce a naming convention.
• Channel grouping is a documented, shared set of ordered rules.