Analyse a website's channel mix with AI

The question is not "how big is this site" but "how does it get its visitors, and how fragile is that". A mix of 60% direct and 5% paid reads very differently from 60% search, even when total traffic is identical. Keep volume and mix separate in every answer.

1. Fix the scope before pulling numbers

Confirm with the user, and state in the output:

• Domain (exact host, e.g. example.com vs www.example.com). Subdomains are usually separate properties; ask before merging them.
• Country or global, and the period. Channel shares swing hard between months and countries, so a single month is a snapshot, not a trend.
• Desktop, mobile or both, if the tools expose it. Mobile sites often have a very different social and direct mix.

If the user has not chosen a period, use the most recent full period the tools return and say so.

2. Pull the traffic sources first

Start with get-websites-traffic-sources. It gives the top-level split between channels for the site. Read the shares, not just the labels, and check they sum to roughly 100%. If they do not, the breakdown is probably incomplete. Say so before interpreting.

Then get the size context from get-websites-traffic-and-engagement so that a 30% share is never described without knowing whether 30% of 2,000 visits or 30% of 20 million.

3. Drill into the channels that matter

Only drill down where the shares say something. Typical order:

1. Referrals (get-websites-referrals-agg): which other domains send traffic. Is one source carrying most of the referral share? That is a partnership or dependency, not a diversified channel.
2. Paid (get-websites-ppc-spend): estimated paid spend. Read it against the paid share. A high paid share with low estimated spend is a flag to check, not a fact.
3. Ad networks (get-website-analysis-ad-networks-agg): which networks the site runs or buys through. Useful for explaining display and affiliate exposure.
4. Marketing channels (get-segments-traffic-sources and get-websites-traffic-sources splits): when the question is about organic vs paid vs social rather than the raw source labels.
5. Similar sites (get-websites-similar-sites-agg) only if the user wants to know who shares this acquisition profile. Do not use it to explain the site's own mix.

Keep the drill-down to what the question needs. Five to six calls is usually enough.

4. Read what the mix implies

Translate shares into judgement, using these rubrics:

• Single-source dependence. One channel above roughly 50% of visits means exposure to that platform's algorithm, pricing or policy. Name the channel and the risk.
• Direct-heavy. High direct often means brand strength or returning users, but it can also hide untagged campaigns, email or dark social. Say that it is "unattributed", not "brand".
• Organic vs paid balance. A site with a large paid share and small organic share is buying its traffic; the question is whether that spend is sustainable against its engagement.
• Social share. High social with low engagement is usually a viral spike or link-in-bio traffic; check engagement before calling it a strategy.
• Referral concentration. Many referrers with a long tail is healthy; one or two referrers carrying the referral share is a partnership risk.

Always pair a mix claim with its engagement evidence where possible. A channel that brings a lot of visits with very short sessions is a weaker channel than its share suggests.

5. Compare fairly if a competitor is involved

If the user asks for a comparison, use the same domain scope, country, period and device for both sites. Compare shares, then sizes separately. Do not state "X gets more search traffic" unless the absolute numbers support it, since a smaller site can have a larger search share and far fewer search visits.

Common patterns and traps

• Shares that do not sum to 100%. Usually a missing or "other" bucket. Report the unexplained remainder rather than rescaling silently.
• Treating estimates as measured facts. Every figure here is modelled. Say "estimated" and avoid false precision (write "about a third", not "33.4%").
• Confusing source with channel. "Google" as a referral source is not the same as organic search. Check which label the tool uses before interpreting.
• Single-month conclusions. A campaign month or seasonal spike can make social or paid look dominant. Check at least two periods before calling it a pattern.
• Ignoring the subdomain split. A blog or help centre may carry most of the search traffic while the main domain looks weak. Confirm the property before concluding.
• Reading paid share as ad budget. Share of visits is not share of spend. Use the spend tool separately and say when they disagree.
• Empty or sparse results. Small sites and some countries return thin data. Say so, and do not fill gaps with industry averages you cannot source.

Output

Site: <domain>, <country or global>, <period>, <device if applied>
Scale: <total visits or range> (<source tool>)

Channel mix (share of visits, estimated)
- Direct: <x%>
- Search (organic): <x%>
- Referrals: <x%>, top referrers: <domain> (<share of referrals>)
- Social: <x%>, main platforms: <list>
- Paid: <x%>, estimated spend: <range or "not available">
- Other: <x%>

What it implies
1. <concentration or dependence, with the evidence>
2. <quality of traffic, tied to engagement>
3. <risk or opportunity>

Not established: <gaps, thin data, unattributed direct, missing period or country>

Rules

• Do not treat any figure as measured fact; all of them are estimates from third-party modelling.
• Do not use an industry average or external benchmark that did not come from a tool result.
• Do not merge subdomains or countries into one number without saying so.
• Never call the mix "strong" or "healthy" without naming the evidence that supports it.
• This analysis is read-only. Do not change any saved segment, alert or workspace setting; if a step appears to require one, stop and ask the user for their explicit yes first.

Done when

• Scope (domain, country, period, device) is stated, and channel shares are given with the total volume beside them.
• Each implication names the channel and the evidence behind it.
• Thin data, unattributed traffic and estimate limits are listed under "Not established".